What changed

Buying on each contract: purchase requests, approval, orders and subcontracts, receiving, and supplier invoices that become the contract's costs. Start from the bill of materials it was priced with.

A Material card on each contract: budget, actuals, open commitments, requested and left to buy, with the forecast at completion and what is late.

Open orders in Finance, with the cash going out by month, and each supplier's on-time and rejected record in Relationships.

A purchase checklist: the FAR and DFARS rules that may apply to a purchase by its value, the contract type and the contract's clauses, each with its citation.

A bill of materials on the price volume: each part with its quantity, source, supplier, unit price and scrap and loss, priced through your rates. Import a parts list from a spreadsheet.

Quotes on the price volume: ask a supplier for a quote, record the answer, use it on its lines, and see how much of the material is backed by a quote. A request for quote file to send.

Units per period, development and production, the price per unit, and a learning curve that works out each lot's hours.

Pricing stages kept as they were, from rough order of magnitude to negotiated, and any two compared by cost element, line and rate.

Each role says how its hours were estimated and what they rest on, and each line item says what its contract type means for the price.

Your contacts on the Contacts screen, with Add a person, and the contracting people named on notices one press away.

Import contacts from Outlook, Google, a phone or a spreadsheet as CSV, Excel or vCard, and take them out as Excel, CSV or vCard.

Pursuits that are not on SAM.gov, such as a prime's request for a quote or a commercial job, with a capture plan, a price volume and a contract once won.

Suppliers in Relationships, each checked against SAM.gov for its registration, any exclusion and its federal work, with a notice when either changes.

The sole source list goes back as far as SAM.gov publishes justifications, about two years, and new ones keep arriving every day.

Negotiating Federal Contracts is in Courses: five lessons, from the rules of the table to the negotiation memorandum, with a quiz.

Canceled, the American way, everywhere. Agreements, proposals and settled pursuits say canceled, and old links and saved filters still work.

Clients, for advisors and consultants who hold a seat on several accounts: every client side by side, with its registration, readiness gaps, pursuits, next deadline and CMMC score.

GSA schedule holders by SIN, and the budgets of the civilian agencies year by year.

A CDRL register on a contract's deliverables, and revenue spread over the period of performance in Finance.

Email into Relationships: copy one address on an email to a buyer, a prime or a partner, and it is filed on the right person and organization.

Notifications on the installed app: a moved deadline, a new notice or a task due today, on your phone or computer.

Joint venture and mentor-protégé agreement templates, two short courses on them and on export control, and ITAR registration on the readiness checklist.

On a phone, lists read as cards with the deciding line first, and the summary folds away. Pursuits are compact rows, with Plans and Decisions as views of the pipeline.

A contract's tabs are in five groups. The notice page opens with its bid facts.

The plan cards say who each plan fits, and the code picker can start from your company's registration. Each screen goes by one name everywhere.

A CMMC and NIST SP 800-171 self-assessment in Readiness and compliance: all 110 requirements, your score as DoD counts it, what stands between you and Conditional status, the system security plan as a Word file, the plan of action as a spreadsheet, and reminders before each date.

A limitations on subcontracting check on the bid sheet, in the bid or no-bid decision, on the pursuit brief and in Teaming.

Answer a sources sought notice or a request for information: a response filled in from your company record and past performance, with the notice's own questions, as a Word file.

Proposal drafts keep their versions. When two people save the same part, nobody's text is lost: the second sees both and chooses. A history compares and restores.

Write the past performance volume from the records you choose for a proposal.

Your first week on Today, one short thing a day for seven days. Every screen has a question mark with what to do there and the lesson that teaches it.

Book a call with us from Help.

The course certificate no longer shows the exam score.

A certificate track for the sixteen-module course. Each module ends with a short quiz that opens the next, then a final exam. Pass at 80 and you get a certificate with a verification code anyone can check, a QR code, and a button to add it to your LinkedIn profile.

On One Code, a button that leads to a paid feature shows a lock and says what the paid plans add, instead of an error.

Back and the filter chips keep your place on the page, and Today appears sooner.

Find your company by its UEI or CAGE code, including a SAM.gov registration with no awards yet.

Follow-ups in Relationships that are not tied to a pursuit now reach your calendar, your week and Today, and a task is emailed on the day it is due.

The weekly email goes out even when nothing on your watchlist changed that week.

A new invoice takes the next number, and an invoice is never saved without one.

A buying office page no longer waits on a slow read of its newest awards.

A download of everything your account holds now finishes. Before, it could stay at "being made".

Agreement end and renewal dates land on the right day at the end of a month. January 31 plus one month is February 28, not March 2.

Older activity records are removed on the schedule the privacy page gives.

Sign-in pauses a network that keeps guessing passwords across accounts, and tells you by email when someone keeps guessing yours.

The app opens faster. It loads the styles of the screen you open first, and the rest as each part is opened.

The dollar total shown beside the number of awards on record adds up every award again.

The NAICS explorer and the NAICS export count the awards of every code, not only the busiest ones.

Hourly alerts go out each hour as new notices are posted. Daily and weekly digests are as before.

The search box at the top finds companies by name.

When a screen cannot load, it says so plainly, offers a way to tell us, and a way back to Today.

On phones and tablets, buttons, chips and the tick boxes on lists are easier to tap.

When someone leaves a team or is removed from it, their open approvals and queue items pass to the account owner, and their training record is no longer kept on the account.

An API key reads what the API reference describes, and nothing more.

After an update, the app downloads only the files that changed.

The NAICS explorer says when the award figures by code are still being counted, instead of showing none yet on every code.

Signed in, the app no longer shows the live demo's bar, and saving works again. In 1.138.0 the app took every signed-in person for the demo and held their saves back.

Lobbying filings and defense budget requests: when the source is down or turns a request away, the reader says so plainly and asks again by itself.

The live demo is the real app: one code, read only, with no account. Every screen a One Code account opens, on real records.

A mailbox joins the outreach campaigns only once its warm-up is healthy, and one that falls behind is taken off.

One market reads its sources in turns: what one request could not read in time is asked for again on its own, and each source fills in as it arrives. A source not read in time says so, and is never shown as having nothing on record.

Subawards, Budget and every vehicle page open in a fraction of the time on the second visit: what they work out is kept for a few minutes.

Vehicles sorted by last order no longer open on a date the source typed wrong, such as 9019.

Download everything this account holds: one file, made in the background and sent as a link that works once, for seven days. A seat downloads its own data.

Every screen has a picture of its own, drawn for what the screen does, in your theme's colours.

Rows in every list are links, so a notice, an award or a company opens in a new tab the usual way.

Saving a search asks for its name on one small sheet and can email you when something new matches, from the moment you save it.

One Code: you pick your code and then confirm it, and you can change it once in the first day.

Watch buttons say what you watch, and every score says what it scores.

On a phone, lists no longer start with an empty column, the search panel closes after a search, and a notice opens with its two main actions first.

Vehicles and the other long lists sort by any column in a few seconds, and a file of a sorted list is read in one order.

Your password and email address can only be changed on the Account screen, and moving two-step sign-in to a new phone asks for your password or a code first.

When somebody leaves an account, its agreement notes leave their inbox too. The interest counts are deleted after about thirteen months, and no figure under 25 is shown.

What you and your team typed is kept out of the screen translator.

New accounts confirm their email address before the app opens. We send one link, good for seven days, and a first payment or a password reset confirms the address too. Until then nothing else is mailed to it.

Team seats say what each seat entails: add a teammate with a title and a role, see what that role may and may not do before you send the invitation, compare every role, and read what the account shares and how to share a view.

An invitation applies the title and role when the person joins, and its email names the role and who invited them.

Sign out in the app goes straight out, without a second page asking whether you meant it.

Refer a company: the month is yours once the company's first payment is past its refund window. It is one month of the cheaper of your plan and theirs, up to 12 a year, while you are on a paid plan.

Vehicles and Data status open in a few seconds even when the servers are busy: their figures are kept from the last update instead of counted again on every visit.

A list sorted by a column that would take too long to sort now comes back at once in its usual order, with a note saying so.

The pages of Inbox, Notes, Proposals and other lists match the rows on them, and Next on page 2 goes to page 3.

Deadlines count your own calendar day, so a date that is today says today and is not marked late in the morning.

Negative amounts read -$1.50M, and times read "5 minutes ago" wherever you are.

Every file you download treats a cell that starts like a formula as plain text.

A role that may not change records can read the people the account found through Apollo, and cannot spend its Apollo credits or change them.

Your password and email address are changed only from the Account screen, which asks for your current password.

The API reference describes 28 more routes: defense contract announcements, SBIR and STTR, wage determinations, occupational wages, FedRAMP, the notice archive, subgrants, registered suppliers, exclusions, labor rates, the defense budget and states.

Refer a company: every paying account has its own link on Account. When a company you send starts a paid plan, your next month is on us.

A search that finds nothing now says what that list holds and where else to look, on Awards, Open notices, Grants and fourteen more lists.

Vehicles opens much faster: its first rows are read from an index instead of sorting the whole list of more than 700,000.

A screen's title no longer shows a blue outline when the screen opens.

The 15-minute walk through your company page is offered once your first payment goes through: by email, and on Today for 21 days.

More improvements to the tools we use to run LastAwarded. Nothing you use changed.

Improvements to the tools we use to run LastAwarded. Nothing you use changed.

Another fix to the tools we use to run LastAwarded. Nothing you use changed.

A fix to the tools we use to run LastAwarded. Nothing you use changed.

A company's page lists the NAICS codes it registered under in SAM.gov, where it showed "[object Object]" for each one. SAM registrants shows each company's other codes the same way.

The founding member rate is given with a code we send to each company we approve for it.

The owner of an account can choose the AI model that writes its readings, briefs and questions: Claude Haiku 5.5, now the default, or Claude Haiku 4.5. It is on Account, and each counts as one reading either way.

A new account is offered a 15-minute walk through its own company page, on Today and in the welcome email.

Cancelling a plan from Account asks for the main reason first, then goes on to cancel.

One market, Competitors and SAM registrants each say what they search, and where to look for a company instead.

The app's answers, the checkout and the receipt pages tell every cache between you and us not to keep them.

The registration directory finds a company by its UEI, as it already did by its CAGE code.

A UEI or CAGE code typed in the search box at the top opens the company search. When Competitors has no award winner with that code, it offers the registration directory with the search carried over.

A request to stop emails from LastAwarded, or an address that bounces, is now honored everywhere we send from, at once.

An account opened from the link in an email from LastAwarded starts with that company filled in and, when its awards are on the record, as Your company.

An account opened from a company's own vendor page starts with that company as Your company, watched, on its primary NAICS code, and opens on Your company.

Occupational wages are brought down as BLS sends them and read whole again. A read that had stopped starts again by itself, with nobody pressing anything.

Connecting QuickBooks Online, Xero, DocuSign, Adobe Acrobat Sign or Apollo is finished only in the browser that started it.

A signature packet goes out once, even when Send is pressed twice, and its signed copy is brought into the library once.

A month sent to QuickBooks or Xero that gets no clear answer is finished later and is not booked a second time.

A connection to the books or to Apollo is no longer marked as disconnected when the trouble was on our side.

An account that connects its own Apollo account does it with Apollo's own sign-in, and allows LastAwarded to search people, reveal the ones it chooses and read company facts. No key is pasted in, and the access is renewed before it runs out.

A signature packet saves again. Saving one had said there was no such agreement.

Changing the state of a past performance questionnaire saves it on its record again.

Messages has a status beside every name: Available and Away on their own, or Busy, In a meeting, Do not disturb, Out of office or Away when you set it, with a line of your own and a time it clears. Do not disturb also holds back the emails about messages.

Up to three screens open beside the one you are on, on a desk: Shift and a click on any link, or the button beside Go to anything (Ctrl and backslash). Drag the line between them for more room; each can be made the main screen or opened in a tab of its own, and they are back the next time you open the app.

Messages has New group beside New message, the people on your account across the top of your conversations, and a group's faces at the top of it. A group no longer needs a name: it is called by its people until somebody names it.

The day's new awards come in every day again. They had fallen nearly two days behind.

Assistance listings are asked for a hundred to a page, the most SAM.gov takes, and the read that had stopped on it starts again by itself.

Wage determinations ask SAM.gov for what it sends, so the count of them comes in.

Occupational wages read the national total across every kind of employer, and the reads that had stopped on its name were cleared.

Other transactions are asked for one kind at a time, agreements first.

Court of Federal Claims: bid protests are found by the court's own case codes, and vaccine cases are left out.

Listed companies: one that files no annual or quarterly report is said so plainly at once, rather than asked for its figures six times.

A brief on a company no longer trips on a contract whose buying office is named only by its agency.

Rules, dockets, bills, labor rates, the regulation, the defense budget and withdrawn subcontract reports are read every day again while the award history loads. They had waited behind it.

A day whose plan was cut short is planned again a quarter of an hour later, rather than the next day.

Data status says when each source last brought in a good page, including a read that found nothing new.

Everything related to a record, and the intake forms, name agreement types as your company does.

The alerts and messages panel uses the app's own text sizes.

On a phone, the one-time tip about the shortcut bar no longer replaces a message you are reading, such as the note that QuickBooks Online is connected.

The guided tour goes screen by screen: from Today and the search box through every screen in the menu, part by part, pointing at each one and saying what it is for.

The bell and the messages icon open a panel along the right edge with your newest alerts or conversations, over the screen you are on. The full inbox and messages are one press away at the top of it.

The part of the menu you open is shown in the accent colour.

Data status says when each source was last updated, in plain words.

The plan cards and the comparison show how many API calls each plan allows a minute and a day.

At a text size other than 100%, the tour's spotlight sits exactly on what it points at.

A guided tour. It walks through the app from start to finish, opening each screen and pointing at what to use there, with Back and Next. Each part has a short tour of its own. Start it from Getting set up, Help or a part's overview, and pick up where you stopped.

The menu's extra rows stand out: More of the market, Companies and records and every All of row are in the accent colour, and a closed fold says how many screens it holds.

Data status lists every source by group, each with what it holds, when it was last read and a link to its screen.

A list now counts for up to a few seconds instead of stopping at your plan's rows. A count that finishes is exact. One that stops says "At least" and how long it counted, and a total says when it covers only the rows counted.

The plan cards show every way the plans differ, in groups, with what every paid plan includes under them and a comparison of all of them, on Users and billing and wherever a plan is chosen.

Defense contract announcements, SBIR and STTR, the FedRAMP Marketplace, wage determinations, occupational wages, other transactions, Court of Federal Claims decisions, bills and SEC filings each get a regular turn to be read, so none waits behind the larger reads.

Assistance listings are read again: SAM.gov counts their pages from 1.

Reminders, calendar items and scheduled saved views use the agreement type names your company chose, and a filter on one of your own types holds in them.

SEC filings are tied to every registration of a listed company under its name, and to the companies that name it as their parent, so the largest contractors are found. The ticker links to its quote on the exchange and its filings at the SEC.

The search box at the top finds screens by name, and the views inside them, such as the Agreements board, Counterparties or Incurred cost. Enter opens the one marked; anything else searches the record.

Defense contract announcements, SBIR and STTR, wage determinations, occupational wages and the FedRAMP Marketplace now start on schedule however much else is loading. Data status says "starting" until their first rows arrive.

Agreements keeps seven views in its row: Overview, Agreements, Board, Calendar, Counterparties, Obligations and Signing. The others are under More views, which names the view when one of them is open.

The page of a very large company or buying office says when its breakdowns are being counted, and how old they are.

A role given the Finance overview alone, or Your company alone, opens that screen and not the whole part. A person can be closed out of either one alone, and can hide either one from their own menu.

Finance, the eighth part, between Contract and Company. It opens on its overview: backlog funded and unfunded, billed, unbilled and cash expected from invoices, costs against budget by contract, indirect rates against provisional, direct and indirect labor hours, audit readiness, the next incurred cost due date and the accounting connection. Timesheets, People and pay, Indirect rates, Cost ledger and Audit readiness are under it now.

The Agreements home: the figures with a year behind each, where every agreement is, charts of value, risk, days in each stage, what ends or renews and what is owed, what needs attention and what the portfolio says. Each person arranges it. A Board moves an agreement along the stages its rules allow, and a Calendar shows every date an agreement carries.

Each agreement opens as a workspace: a lifecycle stepper with the time in each step, its key facts, the document with each clause coloured by how it reads and a clause inspector, two versions side by side, the approval as a chain, and the history as a timeline.

Counterparties, a page for each company you contract with, and Families, a master agreement with its orders and amendments as a tree. Agreement types each company renames, hides and adds. Four ready-made federal setups: subcontract management for a prime, teaming and nondisclosure for capture, supplier and purchase agreements, and consulting and professional services, each a starting point for counsel to review.

QuickBooks Online and Xero, connected live from To the books on Audit readiness. Each account connects its own company, picks each account from its chart of accounts, and sends a month as balanced journal entries. A correction goes as a reversing entry, and nothing there is edited or deleted.

The incurred cost schedules on Audit readiness: each schedule of a final indirect cost rate proposal for a fiscal year, worked out from the records, with what it was drawn from and what is missing, as one Excel workbook, and the date each proposal is due on the calendar.

Everything related to a record: a Related button at the head of every record page shows what is linked to it, across your work and the federal record, one step out or two.

Other transactions, read from SAM.gov's contract award records: tagged on the award, a filter on Awards, Recompetes, a company's and an office's awards, One market, exports and the API, and counted on a company's page.

Plain-English readings of FAR and DFARS clauses, paragraphs for NAICS code, agency and buying office pages, and tags of what the work is on closed notices and recent awards, each written once with AI and checked before it is kept.

The breakdowns of a large company or buying office, the capability statement and a rule's market card read figures kept for them, and the other pages that read the awards read a bounded slice, so they open quickly however large the record grows.

An approval asked of everyone in a role is decided by the first of them to decide.

Signature packets still out with DocuSign or Adobe Acrobat Sign are voided when an agreement is removed.

Key personnel flags a certification or clearance that lapses before the work starts.

Proposal approvals can be required in order.

Proposals, under Proposal: every proposal from the solicitation to award, with its volumes, team and schedule, submission, clarifications and discussions, the award, the debriefing and the protest dates. Start one from a notice.

Compliance matrix, under Proposal: the solicitation and each amendment uploaded to a proposal and read into requirements, each with an owner, a status, evidence and a due date, what an amendment changed, and every proposal's requirements on one screen.

Colour reviews and approvals on each proposal: who reviews which volume, findings as strengths, weaknesses and deficiencies, who closes each, and approvals with stand-ins that must be given before a proposal is marked submitted.

Key personnel and Past performance, under Proposal: your bench with degrees, certifications and clearances checked against each position's minimums, letters of commitment, and every contract you can cite with references, CPARS ratings, questionnaires and relevance to the work.

Agreements: playbooks with positions and fallbacks, concessions approved before signature, the other side's paper read against them, Word files with tracked changes out and back, comments and mentions.

Agreements: intake forms, requests to legal with time targets, approval rules in stages, obligations with evidence, entitlements, a repository with entities, currencies, your own fields, search, saved views, holds and retention, and analytics.

E-signature through your own DocuSign or Adobe Acrobat Sign account, once the version is approved, with the signed copy kept on the agreement, and an agreement shared with the other side through a link of its own. AI in Agreements (summaries, questions and redlines) is off until an account turns it on.

Contract administration: each contract opens on its control center, with modifications as SF 30s, clause currency, the flow-down to every subcontract, health on seven dimensions, vehicles with their orders, and the PIEE invoice packet prepared for the biller.

Audit readiness, under Company: the week certified by the person and approved by somebody else, late entries marked, unallowable costs, the floor check, provisional rates against the records, journal files for the books, and the pre-award survey answered from the records.

Messages, at the top of the menu: the people on an account write to each other one to one, in groups of up to fifty, or on a pursuit, a contract, an agreement or an organization, which everybody who may open the record may read. Up to ten files go with a message, from your computer or your library. The count of what you have not read sits beside the bell, Today shows a block for it, and somebody who is away is sent an email, at most one an hour for each conversation.

Roles and access, under Administration: roles the account defines, each with the systems it opens and whether it may change records, take files out, see pay or add people, with four ready-made roles to copy. One person can be closed beyond their role, everything is enforced on our servers, and every change is kept. Each person can also hide rows they do not use from their own menu.

Security, under Administration and on every plan: two-step sign-in with an authenticator app and ten recovery codes, trusted browsers for thirty days, a lock screen after the minutes you choose, every session with a way to end it, and your last fifty sign-ins. Failed sign-ins pause a name from an address. The owner can require two-step for everybody and set the longest a session may stay unlocked.

Capture is the nine rows of the owner's list, in his order. Capture plans is every pursuit and its plan in one register. Bid / no-bid scores every pursuit out of 100 on eight factors the account weights, with what decided it, the day the decision is due and who makes it, and recording a decision writes the pursuit's Bid gate review. Competitive intelligence lists the companies you meet most in your market and reads each one as how to beat it.

Facilities and capabilities, under Company: your sites with their clearance and certifications, equipment, capabilities with their codes, where you perform and how many people are cleared, and every pursuit you track checked against them as eligible, eligible with a gap, or blocked. The share you are eligible for is on the Readiness and compliance overview.

The handbook can be made your company's own: rename, reorder, hide and add sections and pages, rewrite any page with every version kept, compare versions, mark pages as policy, procedure, template or lessons learned, and print it whole. The Glossary takes your own terms beside ours. Courses add a catalog of short courses in five tracks, each with a quiz and a certificate, given to people with a due date and kept in a training record.

Discuss is at the top of every pursuit, contract, lifecycle record, agreement and organization in the CRM, and a company page links to how to beat it.

The menu is the seven parts of the product, in this order: Intelligence to find the market, Capture to decide what to pursue and win it, Proposal to build the response, Contract to run the award, Company to know whether you are ready, Knowledge to teach the organization, and Administration to run the company. Rows are named for what they are: Awards, Opportunities, Recompetes, Competitors, Regulations, CRM, Contacts, Teaming, Customer intelligence, Proposals, Price volume, Document library and Contract administration. Each part ends with a page that names everything in it and where each is done, and the site has a page for all seven.

Five more official sources, each a screen of its own: the contracts of $7.5 million and more the Department of War announces each business day, SBIR and STTR awards and open topics from SBIR.gov, wage determinations as SAM.gov publishes them, occupational wages from the Bureau of Labor Statistics, and cloud products from the FedRAMP Marketplace. SBIR and STTR and FedRAMP each add a card to a company page, and a labor rate role links to what the same job is paid where the work is.

Tasks and calendar, under Capture: every task and date the account tracks as an agenda, a week, a month, a task list or by person. Tasks are given to people, dated and closed in bulk, and the next step on every record is listed with what will show it is done. Today can be arranged: choose the blocks and their order.

Each stage of the contract lifecycle keeps its own documents and its team by role, and offers the forms the stage calls for, filled in from the record, as a Word file.

Getting started and help, under Knowledge: the setup steps at any time, the tour of the seven parts, where everything is, and the answers to common questions.

The first screen loads a fraction of the script it used to, and each part of the app arrives the first time one of its screens is opened.

Agreements, under Contract: contract lifecycle management for every agreement the company makes. Requests, a repository searched through the words of every agreement, drafting from your own clause library and templates, every version kept with a word by word redline, and each version as a Word file.

Each agreement read clause by clause against your standard and fallback language, with a risk score that gives every reason and the words it was found in, the key terms read from the text to confirm, and your language put in place of theirs in one press.

Approval rules by kind, value and risk, asked in turn and on the approver's calendar; signatures with the signed copy; obligations read from the text, with owners, dates and recurrence, on the calendar; renewals and notice dates as tasks; and agreements signed before, brought in from the library.

Readiness and compliance, under Company: what the company is eligible for and what blocks it, with papers, OCI reviews at Gate 1, supplier declarations, DPAS rated orders and their answer dates, DODAACs, the vehicles you hold with the GSA sales report date, and offsets.

The contract lifecycle checklist adds lines for each kind of contract, such as billing rates and the final indirect rate proposal for cost type work. A company in Relationships shows its UEI, CAGE code and when its registration runs out. An other transaction agreement is a contract type in the ledger.

Relationships, under Capture: a GovCon CRM kept beside the federal record. Organizations, from agencies and buying offices to primes, partners and competitors; the people at each; who on the team knows whom; and every call, meeting and email, against a person, an organization and a pursuit.

Who to know: the contracting people named on the notices of the offices you pursue, watch or keep, ranked by how much of your kind of work they handle, less the ones already in your book. And a forecast of the open pipeline by federal fiscal quarter.

Add to Relationships on every contracting person, buying office and company page. Each pursuit and its lifecycle record show who is on it and every touchpoint, and a next step with a date becomes a task on the calendar.

Contract lifecycle, under Contract: every pursuit and every contract as one record, from the first look at a market to closeout and the recompete, in fifteen stages under Win, Perform and Renew. Where each record stands is worked out from what is on file, and can be set by hand. Each stage has a checklist and can require a sign-off, each record has a health score with its reasons and a history of every change, and the obligations register lists what is owed under every contract.

Your stages: the owner or a manager renames, hides and adds stages, says which need a sign-off, and writes the checklist for each, with lines that apply to one kind of contract only.

The menu is grouped the way a government contractor is organized: Intelligence, Capture, Proposal and Contract, then Company, Knowledge and Administration. Nothing was removed. Pinned views sit under Today, One market and Ask, and the guided tour follows the new groups.

The contracts handbook is rebuilt in the order a large contracts department keeps its own: 139 pages in eleven sections, with the forms at the back. New pages cover revenue recognition step by step, each of the twenty-two contract provisions on its own page, the Uniform Commercial Code, the international sale of goods, commercial offers and the battle of the forms, foreign tenders, letters of credit and guarantees, and foreign military sales, financing and direct sales. A page is set in the same faces as the course, a size down.

SAM registrants can be narrowed to the companies that have won a federal award, or to those that have not, and sorted by it. The export says how many awards each has won and their total, its first and last award, the subcontracts it has been paid for, and when it was last active.

Every list can show 25, 50, 100, 250 or 500 rows a page, up to what the plan allows, and remembers the choice.

NAICS codes, product codes, states, set-asides and notice types can be chosen several at a time.

A ZIP code filter on notices, awards, vendors, subawards and grants.

The library holds far more: 25,000 documents and 50 GB on Enterprise, 5,000 and 20 GB on Capture, 1,000 and 5 GB on Pro, 250 and 1 GB on Shop. Its storage is shown in GB, and one file may be up to 50 MB.

Data status lists every source the record is built from and when each was last read, not only the six read every day.

The founding member rate is for the first 250 accounts.

On a phone, pills, chips and buttons are smaller, a row of chips scrolls sideways instead of wrapping, and long buttons no longer run off the screen.

The contracts handbook is in the app, under Learn, on every plan. It has 107 pages: agreements made before a bid, the terms of a contract one at a time, proposals and negotiations, administering a contract through to closeout, and work for commercial, university, state and foreign customers. Its definitions and acronyms are in the Glossary.

The directory of SAM registrants can be narrowed by county and by ZIP code, and its export carries the county.

A maintenance release. Nothing a member or a visitor sees has changed.

A maintenance release. Nothing a member or a visitor sees has changed.

The earlier years of contract history keep loading while the newest awards are being read, where they waited for those to finish.

The load of the earlier years of contract history no longer waits on one month. A month the source is slow to provide is asked for in smaller parts.

Awards under $2,500 are read again: the source changed how an amount has to be asked for. The page of one code, one state or one buying office shows its figures only when all of them could be worked out, and says they are not loaded otherwise.

The list of set-asides and the list of product and service codes are counted the way the states now are, so they no longer say the record is still loading. An account on the highest plan is no longer told to upgrade when a list reaches its ceiling.

Assistance listings from SAM.gov are asked for in the form the source accepts, and a read that had stopped carries on by itself.

The new screens are in the menu: Describe the work, Per diem rates, Proposal drafts, Listed companies, Assistance listings, Court of Federal Claims and Lobbying filings. Proposal drafts lists every draft your account has started. Each has a page on this site, and the new lists are in the data reference.

AI writing in a proposal draft is optional. It is off until the owner of an account or a manager turns it on, and the pages that speak of your library now say exactly what is sent when it is on. Matching your documents to requirements still uses no AI.

A review of the screens added in the last release. Searches by a description of the work answer faster. A worksheet with a figure that could not be kept is no longer saved empty. The name of a product code no longer waits on the whole award record.

A first draft of a proposal is laid out from the compliance matrix, with your own passages under each requirement, the gaps listed, and a Word file to take away. Notices can be searched by describing the work in a sentence, and each result says why it matched. A pursuit can be sent to your own system as it changes. Travel can be added to a bid from the per diem rates.

A grant program has a page of its own, and every product and service code shows its official name and what it covers. A company that is listed shows what it reported in its annual filings. Decisions of the Court of Federal Claims and lobbying filings are being added to the record and appear on a company once they are read.

Some earlier defense budget requests were on file without all three of their exhibits. The missing exhibits are being added to the record.

Federal Register notices and presidential documents are read every day beside the rules, and the years before are being added to the record. The bills of earlier Congresses and earlier defense budget requests are being added too. A grant shows who the recipient passed the money on to, and those subawards have a search of their own under Grant awards. The Rules tab opens newest first and can be narrowed by date.

Open for comment lists the soonest deadline first, as it says. A subcontract report is kept once for each prime that filed it, so two primes that number their subawards the same way no longer share a row, and each report carries the identifier of the company that was paid.

Closed notices can be searched as they arrive: by words in the title, a solicitation or contract number, an office, a code and a year. Each one opens with what it said, what it led to, and everything else posted under its number. The way in is under the filters on Open notices.

Closed notices from earlier years are being added to the record. As they arrive, an award shows the notice it came from even when that was years ago, with what the notice said and everything else posted under the same solicitation number. The time an office takes from notice to award is measured from those years as well.

Tables in the course and the guides inside the app showed stray marks around some words and ran off the side of a phone. They now read cleanly, and on a phone each row is a card of its own. Figures on the site also stay on screen while they are counted again.

Recompete scores, the buying rhythm behind them and the list of likely bidders on a notice could go a long time without being renewed. They are now worked out in small steps and renewed on schedule.

Housekeeping behind the scenes. Nothing changes on your screens.

The total dollars obligated could stay blank while the award history was loading. It now shows, and the count of awards arrives sooner.

While the award history was loading, the count of awards could show as zero. It now shows the count, and while the count is being taken it shows what has been loaded so far and says so.

The capture read. Pursuits now opens with a read of the whole scope you set: how many contracts, their value, how many are high priority, who holds them, who keeps winning there, which incumbents are in a weak position and why, and the largest one coming with an estimated solicitation window. You can set the scope by sub-agency and by words, and copy the read as text.

Who is positioned. On an ending contract or an open notice: the incumbent and any weakness the record shows, the companies that keep winning that work at that office, the incumbent's subcontractors, the vehicle and its other holders, and your own record there.

This week. Today opens with a ranked list of what to do in the next seven days, from your pipeline, your tasks and the contracts you watch. The weekly mail opens with the same list.

A company page opens with a read of the company: what it holds that is ending, how much it depends on its largest offices, its direction, how it wins, and where the two of you meet. Customers are shown by year and partners are grouped by company.

A vehicle page shows every company that has received orders with its share, orders by office and by year, first orders by year, and what is left. Who to team with is ordered by how many reasons apply to you, and the account keeps a shortlist.

The app opens faster. Its files are about a third smaller, the first screen needs one request fewer, and a list no longer recounts everything each time you turn a page.

An invitation to an account is accepted only from the link in its mail. A seat that cannot see pay no longer sees labor dollars in a short date window.

Tracking something that is already on your pipeline no longer moves it back to Tracking. A filter set by a link shows in the row of applied filters and clears there. Forget it in Settings forgets every remembered filter.

Forecast, The way in and the price bands keep the figures they had for a code when working them out again is cut short.

Every screen in the app opens with the answer: a sentence and the figures that decide something. Quick filters are remembered, an empty list offers the next step, and you can act from the row.

With your own AI key you now choose from named models for eight providers, including Claude Fable 5.1, and set how long an answer may be and how much thinking a model is asked for.

Sole source reads three years of justifications, and a justification that carries no NAICS code is filed under the code of the solicitation it belongs to.

Routes, What it goes for and the forecast are worked out for every code. Before, they reached only the first few.

Your market this week shows its three tables again, and figure tiles on five screens open the list behind them.

Subcontract reports read from SAM.gov carried the wrong state for the subcontractor. The state is now read correctly, and the reports on file are being read again to put it right.

The scenes in the app play by themselves on every device. There is no button to start them. Settings still has Follow my device and Still.

The scenes in the app play from the start, whatever the device asks for. Settings has Follow my device and Still for anybody who wants either.

Every screen in the app now opens with a moving scene of its own, larger than before. If your device asks for less motion the scenes stay still until you press Play animations, and Settings has a Motion choice for it.

Your own AI key now works with nine providers: Anthropic, OpenAI, Google Gemini, xAI, Mistral, Perplexity, DeepSeek, Groq and OpenRouter. You choose the model from the list the provider gives, test the key from the Account screen, and decide whether every reading uses your key or only the readings past your plan.

Moving pictures in the app. Today opens on a live scene, every empty screen has one of its own, and the guided setup, the Ask screen and the waiting states are drawn and animated in your own theme and accent. They stop when you turn motion off.

Ask about this page opens with a shorter introduction.

An alert inbox in the app. Every alert on what you watch is now kept behind the bell at the top of the screen, as well as sent by email.

Private notes on any record. Write down what you learned about a company, an agency, a buying office, a notice, an award, a vehicle or a person. Notes belong to your account, and every note is also on one screen of its own.

Compare two companies side by side: awards by year, buying offices, codes, set-asides, vehicles, size, certifications and exclusion status, with a short list of where they differ.

Switch between accounts. If you hold a seat on more than one account you can move between them, and back to your own, from the side menu. Your own work stays in your own account.

Roles for seats. The owner can make a seat a viewer, who reads everything and changes nothing, or a manager, who can also invite and remove seats. An Activity card on the Account screen shows who joined, left or changed role.

Pay rates now carry a start date, with a history for each person. When payroll arrives after hours were invoiced, the ledger shows the difference and adds one true-up line when you ask for it.

Eleven more lists export to Excel and CSV, including who to call, who is chasing what, about to graduate, who to team with and the directories. On One Code your readiness checklist, company record and pursuit brief now download as files.

A deadline in the evening no longer shows as the next day. Dates follow your own time zone, on screen and in the pipeline workbook.

Small controls are easier to press on a phone, one press of Escape closes a sheet, and two labels on the site that were hard to read are fixed.

AI readings now have a daily limit per account, which keeps the service quick for everyone. An account using its own AI key is not limited.

The books. Timesheets with a change log, people and pay, pay runs you can bring in from a file, a cost ledger, and an indirect rate calculator that works out fringe, overhead and G&A from your own figures. A saved rate set feeds the bid worksheet and the contract ledger, so a bid and the contract it becomes are priced the same way. The account owner decides which seats can see pay.

The bid worksheet now prices a whole bid: a labor rate table, pricing by contract type, hours by period, other direct costs, fee rules, scenarios side by side, and price bands from what the work has gone for.

The contract ledger now runs a contract after award: modifications, funding and ceiling as of any date, billing by contract type, invoices from draft to paid with aging and days to pay, estimate at completion, line items, deliverables and dates, and documents kept on the contract.

The pipeline is now a capture system. Every pursuit has an owner, a next action, a value and a chance of winning. There is a weighted forecast, a board view, a page per pursuit, tasks, and links from a pursuit to its bid worksheet and its contract.

A calendar of everything dated: deadlines on what you watch, pursuit dates, contract periods, option decisions, deliverables, invoices due and tasks. It also comes as a feed for your own calendar program.

The way in is rebuilt. It ranks the ways into a market for your company and writes a plan for each target: who holds the work, how they got it, which vehicles and partners lead in, and what to do next. Routes now lives inside it. Every company a plan names is checked against the exclusions list, and where the record for a code is thin the plan says so instead of guessing.

Ask a question on any screen. You can add your own AI key so readings are not counted against your month. The key is kept encrypted and only the account owner can set or remove it.

Excel, CSV and print almost everywhere. Every work system exports, nineteen lists of the record export with your filters, a saved view can arrive by email weekly or monthly, and each document has its own print layout. A file of your own work is never counted against your monthly rows.

Change your email and password inside the app. New accounts get a short guided setup. Send us a message from any screen, for a problem, feedback or an idea. The language picker is now in the side menu. The app shows its version in the menu and the site shows its version in the footer.

Award history is being extended back to fiscal 2008, the larger awards first.

Inviting someone who already has an account now waits for them to accept, and a seat can leave an account. Choosing a new password also removes the API key and any email change that was waiting.

A contract started from a pursuit now prices labor exactly as the bid did. A cost already on an invoice cannot be billed twice. A fixed fee stops at the fee. A payment short of the invoice stays owed.

Charts by fiscal year no longer draw a year that has barely begun, and they print the true total. Record pages read properly on a phone. Long values no longer shrink the figures beside them.

Every reading the AI writes now carries its working. Each figure in a reading is numbered, and pressing the number shows you the fact it rests on, how many rows that fact was counted over, and a link to those rows. Two claims this product was already making are now true rather than aspirational: nothing said where a sentence came from, and nothing checked whether a figure in a reading appeared anywhere in the facts it was given. A figure that appears in the writing and in none of the facts is now named on the page.

The network. Every other screen here answers "what is open" or "how do I reach that office". This one answers "what is around me": every company, buying office, code, vehicle and contracting person within three hops of where you are standing, drawn rather than listed, with how strongly each one connects to you counted over every route at once rather than over the single best one. It can tell you that three offices are near you through the same company, which makes that company one conversation instead of three, and no table could ever have said it.

A question with several parts in it is now broken into those parts before anything is looked up, and you see the plan. Each part says what it settles, what came back, and opens the rows behind it. Where the record does not hold part of an answer it says so instead of writing around it. The old behaviour asked for what it needed one step at a time inside a loop, which meant there was never a plan to show you, and nothing it produced was kept, so the same question cost the same again.

What to give it. One list of everything this product can use about your company, what each one turns on by name, where to put it and roughly how long it takes. Several screens now say in place when a thin answer is thin because of something missing rather than because the record is empty, which is a distinction they could never draw before.

A firm registered in SAM.gov under a set-aside programme who had not also ticked the box on their account was told, on every notice they were plainly eligible for, that it was set aside for a certification their profile does not claim, and lost thirty points for it. The product held the proof from their own registration and argued against it. Two more of the same shape: the pursuit reading never saw the win theme and the weakness you had typed into the four boxes on the same screen, and the compliance reading was told an item was a gap without being told the score you had entered against it.

The code, customer and end date you can put on a document now decide which document is suggested for which requirement. They were collected by the form, read back into the same form, and used by nothing at all, so the product asked you for work and gave you nothing for it. A write-up from the same code or the same customer now sorts above one that merely shares words, and work finished within three years sorts above work finished a decade ago.

Motion, everywhere, rebuilt as one system. The two halves of this product shared nine timing values out of forty-six and not one easing curve, so the same idea moved at one speed here and another there. Six durations and five curves now, declared once. And the things that had no motion at all: a panel covering half the screen appeared and vanished between two frames, the command palette had none, every confirmation ended by disappearing rather than leaving, and on the public site six hover states changed with nothing in between while the reduced-motion setting reached almost none of it.

Every confirmation this app has ever shown you was painted a full screen below the fold. "Saved." "Copied." "Added to your pipeline." Sixty-five messages, including every failure, every result of a bulk action and the one telling you your session had expired: the app was doing the work and saying nothing, and an error message had no background colour at all. One line put it back where it belongs. If you have ever pressed something here and wondered whether it took, that is why.

A list that failed to load kept shimmering as though it were still loading. Nine screens did this, and the request that fails is the slow one, so what you got was a table that claimed to still be working on something already abandoned, with the error a screen below the fold. They say what happened and offer to try again. Two more from the same family: the screen you land on after a search no longer jumps from one layout to another while you are reading it, and a table that failed now looks like a table that failed.

Two things the tables promised and did not do. The first column was supposed to hold still while the rest scrolls sideways, so the dollar figure you are reading still has a name attached: it was measured sliding 400 pixels off the left edge, because one line of styling made the browser resolve it against the wrong box. And a long list was supposed to keep its column headings: they left the screen after about fifteen rows. Both work now, and on a phone the Clear button on a filter could not be pressed at all, because the bottom bar was sitting on top of it.

The Contrast theme, which exists for people who need maximum contrast, was shipping a primary button at 1.7 to 1 on ten of the twelve accent colours, and the button changed colour under the cursor. It measures 11 to 1 now. Also: the system setting for reduced motion did not reach the one animation that runs longest, a loading bar drew a near-white band across every dark theme, a bar chart drew its own tallest bar six per cent short so the biggest year and the second biggest looked the same, and twelve of the figures on record pages were links that took you back to the start screen when you clicked them.

Sorting a table had no keyboard path at all, filter boxes were not joined to the words above them, moving between screens left the caret nowhere and the browser tab said the same thing on all fifty-eight screens. There is a way past the menu now, the screen you land on announces itself, and Ctrl-K behaves like a real dialog rather than a panel you can tab out of behind your own back. And the bid worksheet and the contract record, which are hours of work each, asked nothing at all before deleting themselves.

Analysis was fixed once already and half of it was missed. The four headline figures were being counted by the engine rather than inside your request, which was the timeout; the six tabs underneath them were not, so the screen still ran five counts across the whole record before it drew anything. Those now come from the engine too, and each tab is remembered for fifteen minutes, so moving between them is instant rather than paying for the same counts six times. Measured before the fix, the tab this screen opens on cost three quarters of a second on every single visit, first and fiftieth alike.

The AI now reads the question you actually have. Every reading in this product answered "what does this record say", which the table above it already answered, for nothing. Two new ones answer "should I bid this, and what would stop me": the first reads your own pursuit brief, the incumbent, what the work goes for, how often one supplier answers, and your own checklist, and comes back with pursue, watch or walk and what to do this week. The second reads the clauses the solicitation actually cites against what you say you hold, and tells you which gaps cannot be closed before it shuts. Both are built from figures already on your screen and nothing else.

The notice brief was reading the cover page. On a real solicitation the description field is a two-paragraph synopsis and the statement of work, the instructions and the evaluation factors are in the attachments, which this product downloads and reads for the compliance matrix and was not giving the brief. It reads them now, says when it has only read part of a long package, and may no longer name a clause number that does not appear in the document or turn a period of performance into a deadline. The Refresh brief button did nothing at all for thirty days at a time; it refreshes. And the fit score could be written as 1 out of 5, the worst possible, by a model repeating the question it was asked.

Your market this week. Every other screen here answers a question you arrived with; this one answers the question nobody has an hour for on a Monday, which is what actually happened last week in the codes you work in. What was awarded and for how much, who took it, what opened, what closes in the next three weeks, and what comes up for recompete within a year. Every figure carries the same figure for the thirteen weeks before it, because four awards means nothing and four against a weekly average of eleven means the market slowed down. The line worth reading first is whether anybody won there for the first time: a market that admitted somebody new last week is a market worth a first bid, and one that has admitted nobody in a year is a market to enter as a subcontractor instead. It is on its own screen and attached to the weekly email, and a week where nothing happened says so in one line instead of padding itself out.

Four things about your own company that the record already knew and nothing ever said. Your SAM registration expiry, which is the one date that decides whether anything else matters: an expired registration means nothing can be awarded to you and buyers doing market research cannot see you, and it was sitting in the record unmentioned while the checklist tracked only dates you had typed in yourself. What gets reserved in your own market, priced in the codes and offices you watch rather than nationally, against what you already hold, with no recommendation attached because eligibility turns on facts about owners and premises that no federal database holds. A capability statement built from the record rather than retyped, which is the document every sources sought and every prime introduction asks for and the one in your folder is a year old. And whether answering a sources sought at a given office actually works: how often one became a real solicitation, how long that took, and how often it came back set aside, which is the whole reason to answer.

The date that is missed most often, and it is not the response deadline. A solicitation usually shuts the door on questions about a third of the way into the window, and after that there is no way to ask, no amendment to clear an ambiguity, and a choice between bidding on a guess and not bidding. That cut-off is now pulled out of the document with the date it names, along with site visits, which are sometimes mandatory, and how long your price has to stay firm. All three sit at the top of the compliance matrix. A related fault went with it: a solicitation that stated its question cut-off in the same words it states its response deadline had that earlier date reported as the deadline itself, which is two dates a fortnight apart with the wrong one reading as safe.

A pursuit can belong to somebody now, and carry the one thing that has to happen next. Seats are sold from the second plan upward, and until now the pipeline recorded who started a pursuit and nothing about who is carrying it, so a shared list was something everybody read and nobody worked. An owner, a next step and a date for it, filterable to just yours, and the next step goes into the calendar feed as its own entry rather than as a note on the deadline, because it is nearly always the sooner of the two and it is the reason a deadline gets missed. Also new: a saved search can now run its export on a clock and arrive as a file weekly or monthly, and a member pointing their own spreadsheet or script at the product has a per-code data endpoint to read, on the API key they already hold.

Analysis and maps, which was the slowest screen in the product and sometimes gave up altogether. The four numbers at the top were counted inside the request that drew them, and on a filter reaching the whole record that means reading all of it before anything appears. A clock cannot help, because a clock is read between one piece of work and the next and nothing can stop one already running. So on a wide filter that counting now happens in the background, the same way every other expensive figure in the product already worked, and the screen opens straight away and says the numbers are still being counted. Three things it used to get wrong are gone with it. A count that was cut short was remembered as an answer, so for the next quarter of an hour the screen told you the record held nothing at all. A measure of how concentrated a market is printed a verdict of "not concentrated" from a number it never worked out, which is the wrong answer in exactly the market it exists to warn you about. And two of the six views had no time limit on them at all.

And a sweep across the rest, by readers who had not written any of it. Two screens were describing how this company is run rather than what is on record, which is nobody's business but ours, and a check in the build now refuses that wording anywhere a customer can read it. A drawing on the home page contradicted the number printed underneath it. A table on the sources page had a stray extra cell that pushed its heading out of line. Public pages counted contract awards the government has since retracted, while every page inside the product excluded them, so the two halves published different totals for the same office. A debarred company could slip past the check that keeps it out of a recommended route. A buying office that admitted a new supplier today was scored as though it had admitted nobody, and sorted below one that had not admitted anybody in a year. A figure on every buying office page could never be worked out at all, so it said "not measured yet" on every site for ever. And the number of feeds said twelve in four places on a product with thirteen.

Three fields taken back out of the award record. They were collected because the government publishes them and it seemed wasteful not to, and then nothing was ever built that showed them. Keeping a field that no screen displays is the wrong trade in both directions. They will come back in the release that gives them somewhere to appear. The five fields added at the same time that the contract page does draw, including the option money still unspent on a contract, are untouched. There is now a check in the build that refuses a new field on that table unless something reads it.

More of the same repair, found by reading the repair rather than the fault. A change to how the record is kept that did not take is no longer recorded as one that did, which is the quiet version of the same failure and the worse one, because nothing would have said so. Two of them starting at the same moment can no longer both begin.

The last release could take the site off the air, and this one is the apology. A change to how the award record is kept was being started by whoever happened to visit the page. A page that takes too long is stopped, so the change never got to finish, never got to record that it had happened, and started again from the beginning for the next visitor, until nothing on the site could answer at all. Three things are different now. A visitor never sets one off, ever. Only one of them runs at a time, so a second one cannot pile on top of the first. And it saves its place as it goes, so a change that is too big for one pass finishes over several instead of starting over for ever. The screens keep working while it runs.

A thirteenth feed, and it is the one that answers a question nothing here could answer before: who is subcontracting on a contract vehicle right now. Subcontract reports are read from the system the primes file them into, rather than from the system that republishes them later, so a teaming lead arrives weeks earlier and carries the vehicle the prime contract sits under. If you cannot get onto a vehicle, the way in is somebody already on it who is buying help to deliver under it, and that list is now on every vehicle page. Reports a prime withdraws are carried too, which the republished copy does not publish at all, so a withdrawn subcontract no longer sits on file as a live lead for ever.

Two numbers on every buying office page that decide whether its work is worth chasing, and neither is about the work. Whether the office exercises its options or lets contracts lapse and re-compete, because a recompete date at an office that quietly extends everything twice is a year of planning spent on a date that will not arrive. And how often a solicitation from that office turns into an award at all, which is the answer to "should I spend eighty hours on this", counted from the record and published nowhere else. Both appear only where there is enough history to mean something.

And on a defense budget line, what successive requests said about the same year. A request projects several years ahead, so the same year of money appears in several requests and rarely at the same number. A line falling by half between one request and the next is a programme losing its argument inside the Department, two years before a solicitation exists and a year before anything else in this market says so. The figures were already on file and only the newest was being kept.

Analysis and maps took far too long to load. It was six aggregates over the whole award history with no ceiling on any of them, and it was the only major screen in the product without a cache or a limit. It now works to a clock: the figures at the top are cached across the six tabs, the two counts that no index can serve stop at a sensible ceiling, and anything that cannot be counted in time is marked as still counting rather than leaving the screen to spin. Also on that screen: the span read "FY1950 to FY9019", because a contract start date typed wrong in the source record parses perfectly well and became a fiscal year. Dates outside the range a federal contract can plausibly fall in are refused now, and the ones already on file are cleared on update.

A cancelled solicitation used to stay open here for ever. The search this product runs returns live notices by default, so a notice the buying office withdrew simply stopped appearing and the copy already on file was never touched again: every field on it still correct except the one that matters. Those are asked for by name now and closed. Six fields the government publishes on every contract action and this product was dropping are also read: the unexercised option value, which is the money still on the table and the reason to be talking to an office before a recompete rather than after; whether the prime is obliged to subcontract; and the authority actually cited for a sole source, which this product had a whole screen about and was inferring.

A full sweep of everything, by three readers who had not written any of it, and the most serious thing they found had been quietly doing nothing for months. The newest award feed, the one that gets a contract on record days before the main system republishes it, was being refused on every single request, and nothing said so on any screen. That feed has therefore never returned a row, and now does.

Four places where a plan boundary held on the way in and not on the way round. An award refused by code could still be opened through the pass that fetches its detail. A notice refused by code could still have its readiness panel read, which is built out of the attached documents that plan cannot open at all. A vendor page printed the vehicles, the published rates and the sole-source justifications, which are three whole screens the narrow plan closes. And the screen for your own company was marked open in three places and closed in the fourth, so it drew, called, and came back as a locked panel headed "That" for anybody whose subscription had lapsed. Also: a plan's watch limits were checked when a watch was added and never again, so an account that moved down a plan kept every watch it had bought on the one before.

And the reading itself, including two faults in the previous week's own work. A reliability measure was being recorded in a way that could overwrite itself, so a perfectly healthy day reported most of its reading as lost and sent an alarm saying so. Asking for a backfill on a second day re-read a hundred and five weeks of notices already on file, because the span was measured from the moment of asking rather than from the calendar. One company with a malformed registration number permanently blocked two whole passes, because it sorted to the front and the failure stopped the batch without marking it. A retry could take longer than the attempt it was retrying. And three internal lookups were reading whole tables to pick four rows.

A day of reading that was running at a small fraction of what it should have been, and four separate reasons nothing said so. Reading happens in timed stretches, and a stretch that is cut short loses everything after the cut, including the record that it ever ended, so the next one stands down and waits. Losing most of every minute that way looked identical to simply being slow. That is measured properly now, and when a day comes up short the product says so first and says what caused it, rather than explaining a setting that was never the limit.

And the thing that reads as broken and never was: work that has not started yet shows as waiting rather than as running, so the state is readable rather than alarming.

A dozen things found by one last read of everything written this week, before any of it had been used in anger. Three could have cost money or trust. The newest award feed had a rule that skips an ordering vehicle rather than counting it as a contract, and that rule had been left reading the old field names while everything round it moved, so a ceiling would have been added to spending as though it had been spent: a total roughly twice what it should be. The clause library decided whether something was a clause or a solicitation provision by reading the heading, and a heading never says; the document says it in its own first line, so every provision in the FAR would have been filed as a clause. And the fault list keyed a failed request on the exact record it failed for, so one broken screen on a fifty-row list would have written fifty separate faults, pushed every other fault out of the list, and mailed twenty of them as brand new.

And nine smaller ones, most of them in things added in the last three days. The feedback sheet opened over a page that could still scroll behind it. Tabbing onto either settings dropdown showed no focus outline at all, in every theme including the two built for people who most need one, because the new wrapper clipped it. Keyboard focus in the menu used a colour mixed for the content area rather than the menu, so it was close to invisible whenever the menu and the theme disagree. The error reporter filed errors from other software on the page under this product. The menu slid during first paint if your opening screen was in a collapsed group. "Something looks wrong" said "What is wrong on buyer?" on a buying office instead of naming the screen. A single browser error turned the whole engine health red and started a daily "something has stopped" email about an engine that had not stopped. Two optional sources warned for ever after being switched off on purpose. And pressing "forget the faults I have dealt with" also erased the record of a refused email.

Housekeeping, and the only reason it is worth a line here is that the last three releases each fixed a defect caused by the same thing. The app was one script of twelve and a half thousand lines and one stylesheet of two and a half, and at that size a rule written for one part of a screen quietly reaches another part three hundred lines away and nothing about either piece looks wrong when it is read. Both are now written as smaller files by subject and joined back into exactly the same two files your browser loads: same single request each, nothing added, no change to anything on screen. The build refuses to ship if the joined file and the pieces have drifted apart.

Three things that were never checked, and two of them were wrong. The menu did not fit on a laptop: eleven headings, the brand block and the footer add up to about 1,190 pixels with the working groups open, which is 290 more than a 1440 by 900 screen and 420 more than a 1366 by 768 one, so the footer carrying your plan and Sign out sat below the fold on every machine in use. The rule meant to prevent that counted rows, and none of what overflows is a row. The menu list scrolls on its own now and everything round it stays put, which also frees the structure to be arranged by what belongs together: Your company has moved to Capture beside what you can perform, where it always belonged and was not, and the two reference drawers are called Companies and people and Records and reference rather than Look up somebody and Look up a record, which was a distinction you had to guess at every time.

Two screens would have got slower and slower as the record filled up, and the narrower your market the worse it would have been. Open notices on your codes and contracts ending on your codes each filter on one column and sort by another, there was an index for each and none for the pair, so the database walked the date index from the newest record backwards until it had collected a page. On a busy code it finds one immediately. On a rare one there is nothing to find and it reads everything to prove it, which is the member whose market is narrow waiting while the member whose market is crowded does not. Both now have an index that does the filtering and the sorting in one pass. Found by running the real queries against a quarter of a million rows and reading back what the database intended to do, which is a thing the build does on every run now.

The newest award feed was reading almost nothing. SAM.gov replaced the old contract-action feed this year and the reader for it was written in the old vocabulary, so against the published schema a contract arrived with its number and its dates and nothing else: no vendor, no UEI, no NAICS code, no set-aside, no money and no buying office. Nothing would have failed. The rows would have arrived, the screens would have drawn, and a whole agency of recent awards would have been filed under no company and no code. Six other sources that had never been handed a payload of any kind are now checked the same way, including the clause library, which was reading the section symbol in front of every clause number as part of the number.

The product can now tell you when it is broken. Until this release every defect in it was found by a person looking at a screen and saying so, which works while the only people looking are the two who built it and stops working the moment anybody else is using it. Four things are watched now. A warning or a crash in our own server code, which was previously invisible because a warning is not a crash and nothing fails visibly. An error in your browser, which is the real blind spot: the request succeeded, the figures were right, and the screen is blank, and nothing on the server can see that happen. A request that answers with a server error rather than an honest no. And mail the transport refused, which matters because sending a digest counts what was handed over and not what arrived, so a site with no working mail could report perfect health for ever. Each distinct fault is one line with a count beside it rather than ten thousand log entries, a new one is emailed once and never again, and one that comes back after being marked dealt with returns to the top of the list by itself.

One thing you can now press. "Something looks wrong" sits in the menu on every screen: it sends what you type, the name of the screen and your email address so a reply reaches you, and it sends nothing off the screen itself, not your filters and not your search.

The two dropdowns on the settings screen were the largest controls on a screen made of small ones. A picker showing the word Today was drawn a hundred and ninety-five pixels wide, because a dropdown takes the width of its longest choice and that one holds thirty of them, and the digest picker was sized by a choice a member on the wrong plan can never pick. Both were also set a size larger and stood three pixels taller than every button and segment beside them, so the two least important controls on the screen read as the most important. They now stand the same height, at the same size, with the same corner as the button under them, and each is as wide as the words it is showing rather than as wide as the longest thing in its list. The Webhook row above them was reserving two hundred and forty pixels of empty space under its own sentence, from a rule written for a row that sits side by side being applied to one that stacks, so that card is a third shorter. All of it is measured by the build now. Nothing in the test rig had ever measured a control, which is how a screen went out with its proportions inverted and everything green.

Twelve palettes and twelve accents, and a menu that can sit on either side of the theme. The five light palettes differ by warmth and by how much contrast they carry rather than by name, the five dark ones by which way they lean off neutral, and Daylight is the accessibility theme that was missing: Contrast is maximum contrast on black, which is the wrong answer for a large share of the people who need maximum contrast, and this is the same setting the other way up. The accents are spread round the wheel so no two are told apart only by their label, and each one now carries the colour of the text that sits ON it, which nothing did before. The sidebar setting offers a light menu as well as a dark one and every palette declares its own pair, so a dark menu on the warm theme is warm and a dark menu on the navy theme is navy rather than one fixed slab dropped on top of whatever is running.

Ten things that were wrong in the appearance settings and in the newest screens. The Sidebar control did nothing at all on five of the eight themes: it moved, it saved, it survived a reload, and the menu did not change, because the rule doing the work sat above the palettes and lost to every one of them. The primary button hardcoded a text colour picked for orange, so a member on Blue had a button measuring 2.7 to 1 where the floor is 4.5, and its hover turned every accent orange. Dialogs were zoomed twice at any size but Normal, so Largest came out half again bigger than intended and taller than the window. The Ctrl K palette followed the operating system rather than the app, so a dark laptop running the app on Paper opened a dark panel over a light page. The theme was not applied until two round trips after the page loaded, and never at all on the plan gate or the code picker. Preferences saved in the browser were shared between accounts on one machine. Reset left the theme buttons memory behind and said there were ten settings when there were eleven. Every accent is now measured by the build against the surface it lands on, in both themes, and a ratio below the floor fails it.

And six in the newest code. The estimated value on an imported agency forecast took the smallest number anywhere in the cell, so a line reading five million dollars over five years was stored as five dollars, and a range took the high end instead of the low one. The date reader matched a month name inside an ordinary word, so a row saying a decision was pending in 2026 came back as a confident three-week window in December, while a column that said Fiscal Year 2026 in words was dropped as unreadable. A column headed Agency Ref was taken by the agency field, which put a line number where a department belongs. Every imported row hashed the office name as typed instead of resolving it, so it joined to no buyer page. The cost audit trail was the one table left behind when work moved to the account, so every entry explaining a colleagues change to a total became unreachable on a team. And the two company figures behind every small business verdict stayed on the person who typed them while the rest of the checklist moved, so two people at one firm got two different answers.

The digest learns about your own company. Every block in the alert mail was about the market: what is open on your codes, what a buyer you watch has bought, a rule closing for comment. Two facts worth more than any of them were computed nightly and shown only on a screen somebody had to remember to open. A contract of your own inside its recompete window, which is the most expensive thing that happens to a small supplier and the most predictable, and which only ever reached you before if you had thought to put your own firm on your watchlist. And a readiness answer running out, where the lead time is the whole point: something that takes six months to renew has to be started six months out, not on the day it lapses. Both sit above the market blocks and below an amendment, because that is the order in which they can cost you something.

Work belongs to the company that pays for the seats, not to whoever typed it. The pipeline, the bid worksheets, the contract ledger, the pursuit briefs and the readiness checklist were all filed against the individual person, on a product that sells seats: two people at the same eight-person shop could not see each other pursuits, and the checklist, which is a list of facts about the COMPANY, held a separate answer per person, so the readiness result on a notice said something different depending on who opened it. All of it now belongs to the account. Everything already written moves across on upgrade, with a collision merged rather than skipped, so nobody opens the app after the update and finds a colleague work gone. Every row says who put it there, once, on the day they did, and an edit by somebody else never quietly takes their name off it. The deadline and amendment mail now reads the shared list, so a seat holder still gets told.

Your own company, which is the one this product could say least about. Every competitor page here can tell you what a firm holds, what runs out when, what it is published at and whether it is barred from federal work. All of that was on file about the member too and no screen had ever shown it to them. Three things on it can end a decision by themselves and are ordered that way: whether the company is on the federal exclusions list, whether something on the readiness checklist has lapsed or lapses inside the lead time it takes to put back, and whether one of your own contracts is inside its recompete window. Losing a recompete on work you already hold is the most expensive thing that happens to a small supplier and the most predictable, because the end date has been on the record since the day of award. The short version appears on Today when there is something to say and is silent when there is not. The company match is a name lookup, is presented as the guess it is, and is correctable in one click.

Agency procurement forecasts, where an agency has published one. Agencies put out what they intend to buy as a spreadsheet on their own site, in their own format, with no API and no two of them agreeing on a column name, which is why this product derives a forecast from the award record instead and always will. What was missing is the other half: a published forecast says things the record cannot, a requirement never bought before, a named person to talk to, an estimated value on work with no history. Those lines can now be put in and they sit in the same list, never blended with the projected ones. A published line is labelled as an intention, carries the requirement in the agency's own words, turns a quarter or a fiscal year into a real window rather than a bare date, and carries no confidence score at all, because a plan has no rhythm in it to measure and a number there would be manufacturing evidence for somebody else's intention.

The compliance matrix now reads the solicitation instead of the paragraph about it. On a real request for proposal the notice body is a two-paragraph synopsis and the instructions, the evaluation factors, the page limit and the address to send it to are in the attached documents, which this product had never opened. The matrix came back with nine rows where there should have been ninety and looked exactly as finished as a complete one, which is the worst thing a compliance tool can do, because a missed instruction is a proposal returned unread rather than points off. The documents are read now, and the screen states in one line how many of them the matrix was built from and refuses to hide it when some are still unopened. The readiness check and the clause list read the same package, which is where clause numbers have always actually been. Every row can now be ticked off, given an owner and pointed at the volume that answers it, shared with everybody on the account, and downloaded as a spreadsheet, because a matrix that cannot leave the screen is a reading exercise rather than a day of work saved.

A moved response deadline is how a shop goes non-responsive on work it was going to win, and nothing here could see one happen: the record was overwritten with the new date and the old one was gone, so there was no before to compare against. The previous deadline is kept now, an amendment is noticed the moment it lands, and it goes to the top of the mail above everything else with the direction stated, because a date that moved earlier is an emergency and a date that moved later is not. Anything read off the superseded documents is dropped at the same moment, since a matrix built from the old package is more dangerous than none. Deadline reminders also now cover what you are actually chasing: pressing track in the pipeline is the one gesture that means you are bidding, and it used to buy a calendar entry and not one email.

A loss was recorded as a word and nothing else, so a year of bidding came back as a tally. Settling one now asks why, from a short fixed list, with what you bid and what it went for, and the winning figure is filled in from the record where the award has been published rather than asked for. Two readings come out of it. Why you lose, counted, because most of what a shop loses it loses on one thing and that is worth knowing as a single fact rather than as an impression. And whether you are actually expensive: the two numbers have been sitting in this database under identical keys the whole time and nothing compared them. It says nothing at all until four settled bids carry both, because a pricing verdict from two is how a shop talks itself into cutting a margin it did not need to cut.

The product counts how many accounts have opened a federal record, so it can eventually say how crowded a pursuit is. That is the only thing it records about what a member does rather than what they typed, and the privacy policy now says so in full: only federal records, the count published and never the list, nothing at all below twenty-five accounts, one signal per account per record per day, deleted after about thirteen months. The code always said a member could turn it off and nothing could: the flag was read and never written, so there was no control anywhere. There is one now, under Settings, and turning it off deletes what was already recorded rather than only stopping the recording.

The last of the work the audit opened. Three screens ran a query per row where one would do: the bidders panel and the likely-bidders list asked how many contacts each company had, once per company, up to four hundred times on a busy solicitation, and the nightly forecast builder read its buying history one group at a time. The same piece of work could be handed out twice at once, which meant the same page of a source read twice. And three parts of the product that nothing had ever executed are now driven by tests: the API, which is a hundred and ninety handlers and the permission checks every one of them passes through; the reading path and the federal debarment list write; and the sample data the browser rig feeds the app, which was supplying a bill stage the product does not have, so a third of the policy progress bars drew blank in every screenshot.

Every source class in this product has a half that builds a request and a half that reads the answer, and only the first half had ever been tested. The rig now hands SAM.gov, USASpending, the Federal Register and Congress a recorded response and checks the row that comes out: the fields that arrive in two different shapes, the ones that are genuinely absent, the row that cannot be identified and is dropped rather than half-stored, and the links that are checked before they reach a page. Writing it found a job that could lose its date window and spend a request to discover it. Separately, ten of the illustrations had labels that ran into each other at phone type sizes, and the build now measures every label in every drawing at two widths and fails if any two touch.

The three text boxes on Award History were reading the whole table. Vendor, Agency and Buying office matched anywhere inside a name, which no index can answer, so each search read every award, and the count, the total and the facet lists evaluated the same thing again. The match is unchanged and now happens against the directory of offices and companies beside it, which is thousands of rows rather than millions, and the award table is asked for the exact names that came back. Two parts of the product that had never been exercised by a test are now: the pipeline, which is member-owned writes, the win-rate maths and a calendar feed with no login behind it, and billing, which decides what plan somebody is on and whose monthly allowance a request spends.

An audit of the whole product, and the faults it found. One Code held on every list and not on the home screen: that screen is built from your watches rather than from filters, the plan allows a watch on a state or a buying office, and the clause those watches build is an OR, so one state watch opened that whole state. The calendar feed carried pipeline dates to a plan that cannot open the pipeline. An office page built its price band from the OFFICE's main code rather than yours. A notice refused by code could still have its documents pulled by asking for them directly. A seat was counted on the day it was added and never again, so an account that moved down a plan kept every seat it had bought on the one before. The AI allowance was charged for requests that failed and never refunded, so twenty minutes of provider trouble emptied the day. A company merge moved six tables and there are fourteen, and one of the eight it left behind was the federal debarment list. Links carried on a notice went into the page without their scheme being checked.

Recompetes are Pursuits now, and the change is not the name. The screen was eight columns and a score, which is a search result, and it put the least useful number on its own in the second column while the facts a capture manager decides on were spread across the other seven with nothing joining them up. Each one is a card now: a sentence saying what it adds up to, then the eight facts, then the capture plan. Nobody buys a sorting aid. What a team buys is being told which contracts are worth their capture budget and what to do about each one, which this screen has been able to answer for two releases without ever saying so.

Two facts that were never counted, and they are the ones that change the decision. How often the agency has bought this kind of work: an end date says when a contract stops, not whether anything replaces it, and a requirement re-bought in five separate years since 2004 is worth two years of positioning while a one-off with the identical end date is worth none. Until somebody counts it the two look the same, and nothing did. And how many different suppliers have ever won that work at that agency: one is a displacement rather than a competition, which is worth knowing before a month goes into it rather than after. Both now move the score, both are stated in words on every pursuit, and the list can be sorted by how often the work comes back rather than only by when it ends. Counted from distinct years rather than from award actions, because a requirement bought on four task orders in one year is one buy and counting the orders would report it as recurring quarterly.

The recompete score now takes your own record into account. It used to be one number built nightly from the federal record, the same for every member looking at the same contract, which is the right answer to what the market thinks and the wrong answer to what you should do. Two suppliers are not in the same position: one has won at that office three times, the other has bid there twice and lost both to the same incumbent. You already wrote that down when you settled the pursuit, and nothing was reading it back. Now it is, with three guards. Nothing moves until you have settled enough pursuits for your own history to mean anything, because a rate on two bids is not a rate. No single office or incumbent counts until you have met it more than once. And the market score stays on screen beside the adjusted one with the reason it moved, because a number that silently changed is a number nobody can argue with. None of it is written back: your record adjusts your screen and nobody else's, ever.

Who to ask first. Every guide tells a new supplier to talk to the small business specialist, and none of them tells you who that is for the office that actually buys your work; the department-wide small business office is one mailbox covering two hundred buying offices. The answer was already here and going unread. A solicitation carries its contact block, and a good share of those blocks name somebody whose title says exactly what they do. The directory filters to that role now, and a buying office page leads with them, because they are the one thing on that page you can act on today rather than a record of what the office has already done. Where no notice from an office has named one, the page says so rather than showing an empty list: an office that has not published a name and an office with nobody are different things, and only the first is what the record can show. Roles come from the title the government typed and from nothing else. A contracting officer holds a warrant and a contracting officer's representative does not, so the filter tells them apart.

One Code held on the lists and leaked everywhere else. A notice links to its award, the award to its vendor, the vendor to its other work, and three clicks from a narrowed list you were reading a contract in a code you do not have, with the header still saying you were on one. Records now check the code they carry. So do the pursuit briefs, before the meter rather than after, so a refusal costs nothing. The directories of every NAICS code, every product and service code and every set-aside are on a plan now, because a screen whose whole purpose is to show the other codes cannot sit on the plan that is one code; the market page for your own code still opens. A buying office and a vendor both publish a top-codes list, and those were naming other codes on pages that had otherwise been narrowed. They are cut to your code too.

An empty list said nothing matched, which is a statement about the federal record and was usually a statement about the account: the market is full and you are looking at one code of it. Read the first way it says the data is thin, which is the opposite of true. Every narrowed answer now says which code it was narrowed to, and the empty state says it in place of the old wording rather than under it. The screens stopped claiming breadth they were not delivering, and the two toolbar buttons a plan buys are marked rather than offered and then refused a click later.

The plan comparison on the account screen was printing a dash for three of its six rows and drawing all four cards as lozenges with an ellipse round the current plan, on the screen where somebody decides what to pay. Two separate faults that met there. The rows were looked up in the price table by three names that table has never carried, and the list of names existed in two copies so fixing one left the other wrong; there is one copy now, checked against the table by the build. And a tag added two releases ago for the evidence ladder was called the same thing as the plan card, six hundred lines apart, with enough specificity to win: every card became a pill. The build now refuses a class name that means two different things.

One Code. An account now opens on one NAICS code and everything this product knows about it: every federal award in that code back to FY2001, every open notice the day it posts, what the work goes for office by office, who holds the contracts and when they end, the recompetes, the forecast, and three full pursuit briefs a month. No card, nothing to cancel, and no day it stops. What a paid plan adds is breadth and the tools: every other code, the bid worksheet, the contract ledger, win and loss, saved searches, email alerts, CSV export, the API and the AI readings. The line is drawn that way on purpose. The usual version of this is the whole product with every number turned down, which teaches a visitor that the product is thin: they see every screen, none of them answers a question properly, and they leave having concluded the paid version is the same screens with a bigger number on them. The real thing on a narrow slice says what this is. A code can be changed after thirty days, and a paid plan that lapses drops back to one code rather than going dark.

The screens a plan buys stay in the menu rather than disappearing, marked, and opening one draws a panel naming the thing it is, what the account already has, and what a plan adds. Two reasons. Hiding them would make the product look smaller than it is to the one reader who most needs to see its size, and would change the shape of the menu on the day somebody upgrades. And a locked screen that says upgrade is an advertisement, while a locked screen that says the contract ledger tracks what you have won, what you have invoiced and what is left is a decision somebody can actually make. The panel leads with what the account has, because a member who only ever reads what they are missing concludes the product is a funnel they are standing inside.

What the narrow plan may reach is a named list of routes rather than a set of numbers turned down. Relying on the dozen existing limits would have made every one of them load-bearing for revenue, and a bug in any single one is a hole nobody notices for months. Anything not on the list is refused, so a screen added to the product next year is not given away on the day it ships; a lock inside an open area wins, so a notice opens and its AI brief does not; and a list asked for before a code has been picked matches nothing rather than matching everything. That last one is the important direction: failing open here would hand over the entire federal record, and it is asserted in the build.

The pursuit brief, which is now the first thing you open on a notice rather than the last. Everything the product knew about one pursuit was already computed and was spread across six screens, so a member was expected to hold a notice, a price band, an incumbent, a clause list, a competitor list and a readiness checklist in their head and reach a decision. It is one page now: the recommendation with its reasons, who holds it and what they paid, what the work goes for here, who else is chasing it and at what strength of evidence, whether you are small for the code, what you would have to be able to prove, and your own plan. None of it is new data and none of it is a new request to anybody. It is the same facts, assembled where the decision is made. Added to it: a capture timeline, counted back from a date the government published rather than predicted, saying what to do twelve, nine, six, three and one month out, with the stages already behind you marked as passed rather than hidden, because knowing you are late to a pursuit is more useful than a tidy list that pretends you are not.

The menu is five verbs in the order the work happens: find, understand, target, capture, win. It used to be named after what the screens were rather than what you are doing with them, which is an inventory, and with forty-six destinations underneath it an inventory is what a new member had to read before they could do anything. Nothing was removed. Every screen is now where somebody would look for it while doing that step. The home page leads with one pursuit worked end to end instead of with a count of feeds, because nobody pays for a feed count, and the page says plainly that the example is illustrative rather than inventing a supplier to sell you something. The pricing headline says what the product is for: know which contracts are worth your capture budget before you spend it.

The product has started recording how many accounts open, watch or pursue a federal record. Nothing is shown to anybody yet and nothing will be until a record has been seen by at least twenty-five separate accounts, because a smaller number identifies who is in it rather than describing a market. The recording starts now because that history cannot be backfilled later. It covers federal records only, never a bid, a cost, a contact, a note or anything else typed into the product, those remain private exactly as promised. It is one row per account per record per day, no query reads an account back out, only counts of them, and any member can switch themselves out of it entirely.

The question that comes before the price. Every screen in this product answered something about the work and none of them answered whether you are allowed to do it, which is the question with a lead time attached. A cost-reimbursement contract requires an accounting system a government auditor has examined and called adequate, which takes months. A defence contract carrying the safeguarding clause requires a security plan against a 110-control standard and a score posted before award. A build to print assumes a quality system a registrar certified. A facility clearance cannot be applied for at all; it has to be sponsored. None of that is written on a notice in words. It is carried by the clause numbers in section I, which is forty references with no text against any of them, so the requirement is published and unreadable, and the usual way a small supplier learns about it is six weeks into a bid. Thirteen requirements are now tracked, each one read from a clause the notice itself cites and printed with that clause beside it. Fill the checklist in once and every notice says what it needs, what you have, and where a gap is too long to close before the deadline. Nothing is verified, nothing certifies anything, and a requirement you have not answered for is never treated as a pass or as a failure.

The SBA size standards are on file, from 13 CFR 121.201 rather than from a copy of it. Almost every decision a small supplier makes turns on whether they are small for a particular code, and the product was making set-aside recommendations on every screen while being unable to answer it. The table is now read from the same regulation reader that holds the FAR and DFARS text, so it needs no new key, and it is refreshed only when title 13 has actually been amended. Give it your average receipts and your headcount once and the answer appears on every code page and every notice, with the threshold, the amendment date and a link to the regulation. Where you have not said, it says not known rather than guessing, because the two are one click apart from a false certification on an offer. Separately, the labour rates page now states out loud that schedule rates are the ceiling and that the Department of Labor wage determination attached to a solicitation is the floor, that there is no public interface to read those in bulk, and that this product therefore does not hold them.

Attachments on a notice were sending members to SAM.gov instead of downloading. The route signed the link and, on any failure at all, redirected to the SAM.gov notice page, so a site whose key had expired quietly turned every download into the one trip this page exists to save you. The public site had the same files working the whole time from the raw link. It now tries the signed link, then the raw one, and only sends you to SAM.gov when there is no usable link at all. Also: the place of performance on a notice is a link to that state, the way it already was on the public site.

The named contacts on a company registration were being read on the back of the pass that reads its integrity records, because both ask the same endpoint for the same company and it seemed like no extra work. It was no extra work and it was wrong: that pass stamps a company and will not look at it again for a month, so where it had already run the contacts were never fetched at all and the screen sat empty with nothing on it to explain why. It is a pass of its own now, with a stamp of its own, and there is a button that runs both of the new internal tools rather than leaving them behind several million award rows. The company list also counts what it is listing rather than saying how many fit on a page.

The second half of the internal work, and the harder half: for one solicitation, which companies are chasing it. There is no public answer to that question. SAM.gov keeps an interested vendors list, but reading it needs a federal government system account, meaning a government email address and a contracting role, so a supplier cannot have it. Who submitted an offer is protected during the procurement and is never published, and after award the government names the winner and says how many offers it received rather than whose. So the screen refuses to produce a bidders list and produces evidence at six strengths instead, drawn so that a fact and a guess can never be mistaken for each other: who won it, who holds the contract it replaces, whose name appears in the text the government published on it, what somebody looked up and wrote down, who has won this code at this office, and who is merely registered to do the work. The strengths are never added into a score, because a score is how the difference gets hidden. The middle rung is the interesting one: it reads the published text across every notice sharing a solicitation number, which is where the questions and answers live, and shows the sentence it found a company name in so the match can be judged rather than trusted. It also says, on the screen, that most agencies strip the asking company out of a questions document, so finding a name there means something and not finding one means nothing.

Work has started on the supplier side of the contacts question. The product has always held the government side, because the government publishes it: the contracting officer on a notice arrives with a title, an email and a telephone number, so suppliers can ring them. The company side is almost entirely unpublished. Nobody releases who bid on a solicitation, at any access level, and a SAM registration gives up the name and the title of its named contacts and nothing else, because the telephone numbers and email addresses in it are controlled information that needs a federal system account to read. So the new screen does what the record actually supports: it holds the company, the named roles the registration publishes, and the open work that company is a plausible bidder on, worked out from what it has won and at which offices, with every notice saying why it is on the list. What it never does is present a guess as a bidders list. It is internal while it is being used in anger and is on no plan, so nothing here is sold yet.

The bid worksheet prices section B. A solicitation does not ask for a price, it asks for a list of priced lines: program management as one, engineering as another, each with its own number, quantity and unit, and the cost volume has to follow that list. A total with nothing under it is a volume an evaluator cannot check against the schedule they handed out, and the first thing that comes back is a question. Each line now carries its own type, its own quantity and unit and its own materials, subcontracts and travel; every role is assigned to the line it is priced into; and each line is burdened on its own through the same chain the whole sheet uses. The line prices add to the period price exactly rather than nearly, because the period total is taken as the sum of the lines rather than worked out separately and reconciled afterwards. That is exact because every step of the chain is a percentage of what is below it, so burdening each line on its own is the same arithmetic as burdening the sheet, and there is no allocation rule for anybody to argue about. The numbering follows the convention as well: 0001 in the base is 1001 in the first option and 2001 in the second. A sheet with no lines on it prices as one line, which is what every existing sheet does, to exactly the number it always did.

Clear all did not clear. Two screens put a default filter on an empty arrival, which is right, but they put it on every arrival including the one immediately after Clear all, so the pill the member had just taken off came back and the product said in writing that it was still filtering by something it had been asked to drop. The registrant directory also defaulted to a state, which on a site still loading its directory state by state opens an empty table and reads as a product with nothing in it; it now opens on the code the account already told us about, which is read across the whole country. The on-or-off filter that sits beside Apply was inheriting a text field's height, padding and full width, so it drew as a large empty square with its label wrapped beside it. And on Incumbents about to graduate the NAICS code ran straight onto the end of the money with no heading and no space, so a line read $6.09M541330. Registrants and exclusions can now also be loaded on their own rather than sitting behind four million award rows, which is why both were empty on a new site for longer than they should have been.

The Ask screen now says it uses AI before you use it rather than only underneath the answer, and so does every reading offered on a record. Both say what the model does and does not do: it reads the figures on the page and the records it searched, every number in the answer is counted from those records rather than recalled, every claim links back to what it came from, and it can still be wrong about what you meant. In dark mode the warm tinted blocks were reading as brown and tan against a blue-grey page; the tint is now a raise in the same family and the accent lives in the text and the strokes, where it belongs.

The bid worksheet prices the term rather than the year. Almost no federal contract is one year: it is a base period and its options, and the option prices are agreed at award, which is the commonest way a small supplier wins a five-year contract and loses money on it from year three. Each period now carries its own length, its own escalation on labour, its own escalation on everything else, a level of effort for the option year that is a different size, and its own indirect rates where the provisional rates move. Escalation compounds, which is the part that is got wrong by hand: three per cent a year is 1.0609 by the third period, not 1.06. The screen states what the same term would have cost with no escalation at all, so a bid quoted flat reads as the cost you are absorbing rather than as a keener price. The whole term is what gets placed on the band of what this office has paid, because an award record is a contract rather than its first year; the wrap rate and the role check stay on the base period, because the published rates they are read against are current. Every sheet written before this prices to exactly the number it always did.

A twelfth source, and the one that makes a solicitation readable: the regulation itself. Every FAR and DFARS clause is now held here as text, read from the eCFR and re-read only when title 48 is amended. A solicitation incorporates its clauses by reference, which means section I is forty numbers with no words and perhaps ten of them decide what the work costs. Open any notice and the clauses it names are listed, looked up, and sorted by what they cost, with the families that take months to be ready for at the top. Every clause number on a rule is now a link to the text of the clause it rewrites, which closes the loop the rules screen opened: this is the change, this is what it currently says, and these are the open solicitations citing it.

In settings, the two Figures options were the same number drawn in the same face, which reads as a bug rather than a choice; each is now drawn in the face it is offering and says what it is. On a phone every settings row reserved 240 pixels of nothing between its own sentence and its own control, because a flex basis written for a row becomes a height in a column, so the Figures and Motion rows were each about four hundred pixels tall with one line of text in them.

The loop closes. Five things, all of them the part of federal contracting that happens after somebody says yes. Every change to a recorded cost now needs a reason and leaves a line in an append only trail that nothing in the product can edit or delete, which is the entire substance of what recorded hours are worth. Funding actions are tracked over time, so the limitation of funds notice is measured against what is obligated today rather than against whichever figure was typed first. The actual indirect rates, once they are worked out, are set beside the provisional ones and the settlement is reported in dollars, in March, rather than being discovered in January. A period can be prepared as an invoice in the cost element order a voucher is read in, using the same arithmetic as the bid and the ledger so the three can never disagree. And your real cost of doing the work is finally set against what that buying office has actually paid other people for it, which is the one reading no combination of the products in this market can produce, because each of them owns a third of the lifecycle and cannot see the other two. Also new: win and loss, which joins what you bid to what the record says happened and gives you a real win rate by buying office and by code, counted only over the pursuits that were actually settled.

Merging two records for one company no longer flattens what is inside it. A merge is a claim about ownership rather than about sameness: Lockheed Martin in Fort Worth and Lockheed Martin in Sunnyvale are one corporation and two completely different competitors, and rolling twenty registrations into one total would throw away the only line that matters to somebody deciding which of them to chase. So a merged vendor page now lists every registration inside it, each with its own identifier, its own city, its own award count and its own money, and each one opens on just its own awards. The split is read from the identifier published on every award rather than from the merge, so it stays right whatever is merged, and the parts always add up to the whole, including money paid under an identifier that never got a record of its own.

Merging duplicate vendor records looked like it was not saving. It was saving. A suggestion is a group, and eight registrations of one company in eight cities is one finding with eight rows in it; the panel drew all eight, merged the first two, and then removed the whole card from the screen. So it reported eight merges, performed one, and on the next page load the other six were still there. It now ticks the rows to merge and the row to keep, does the whole group in one call, prints the count the server actually returned, and reloads the list from the server instead of hiding anything. Merging the same group twice is also a finished merge now rather than an error. Separately, a lot more motion: eight new pictures that draw what the product does over time, including a contract burning down against its own clock, a salary turning into a billable rate one pool at a time, the federal year with its September spike, and every award behind a price band settling into it. Connectors that carry data now show it moving. In the app, screens and cards arrive rather than appearing, charts draw themselves, and big figures count up, all of it switched off completely for anybody whose settings ask for stillness.

The contract ledger, which closes the loop. This product already held what was awarded and what a bid should cost; it had nothing for the part after you win, which is where the money is actually made or lost. Now a contract can be filled straight from the award record, cost recorded against it and against each line, and burdened by the exact arithmetic the bid worksheet uses, so what you planned and what it is actually costing are finally the same numbers. It watches the four things that cost real money and that nothing tells a small supplier until it is too late: the limitation of funds notice on a cost type contract, spending faster than the clock, a line gone past its own funding while the total still looks fine, and a period ending with cost still unbilled. Also in this release: the search box at the top of the app now searches the record rather than submitting to the AI screen, which means looking something up no longer spends a reading; the guides, the course and the longer reads are drawn as a reading surface with a section rail and a progress bar rather than a stack of identical cards; and the primary button lost its gradient.

The bid worksheet. Two screens in the product already answered what the work goes for and what a named role costs an hour; neither answered the one that decides whether to bid, which is what your own number has to be. Roles, hours and salaries in, with your indirect rates applied in the order they actually apply, and out comes the price, the wrap rate, and every step between. It is then read twice against the record: the total is placed on the band of what this buying office has actually paid, and every role on the staffing plan is checked against what that role is published at, because a price can be right in total and lose on one line. Nothing is stored until you press Save, and a worksheet is never mixed into any figure the product publishes. Also fixed: the course was published out of order, so every index of it opened on modules ten, eleven and twelve before starting again at one, and the primary button carried a moving highlight that read as an orange gradient.

The size and density settings did not reach the menu, so the sidebar stayed at one size while everything beside it moved. They reach it now. Dropdowns and text boxes were drawing as flat grey slabs in the dark theme, here and in the app, because the stylesheets named the forms they were written for and any control outside those fell through to the browser's own widget; there is one rule for every control now. The guides are read inside the app rather than opening a browser tab that replaces it: they are grouped by the job somebody is doing instead of alphabetically, the one written for a complete newcomer leads on its own, and everything written is searchable in one box.

Ten things you can change about how the app looks and reads, saved to your account so it follows you between a laptop and a phone: light, dark or match your system, an accent colour, a size that scales the reading area without the menu eating the screen, density, a sidebar that can stay dark while the records stay light, your own system font, and stillness for anybody who wants the fades off. How figures read is a preference too: money short or in full, dates month first, day first or year first, and which screen the app opens on. Also in this release: a new guide, Federal contracting 101, for anybody who has never sold to the government, and forty acronyms added to the glossary the app reads, so they are searchable beside the record that used them. And a fix worth naming: reading a rule puts it on file, and tagging it is what connects it to a NAICS code. On a site still loading its history the tagging never happened, so rules arrived daily and almost none of them reached the code pages they belong on. They reach them now.

Notices posted today showed yesterday's date. A notice posted on the 20th read "Sep 19" in the app while this site, reading the same row of the same table, read "Sep 20", so it looked as though the two halves of the product had fallen out of step. They had not. The notices were there, on the right day and in the right order, and being printed one day early: a date with no time on it and no timezone means midnight in Greenwich by definition, and midnight in Greenwich is the evening before almost everywhere in America. Every deadline stored the same way was a day early too, which is a date somebody plans a bid around. Dates are now printed from their own digits, and the rig checks them from six timezones rather than only the one the server happens to be in. Also in this release: a source that briefly could not be reached now comes back on its own within fifteen minutes. And the band naming what is different here is fifteen rows rather than eight, rebuilt so it reads as a list of cards on a phone instead of a wall.

What the people cost, on the notice. A services bid is mostly payroll, and the hourly rate a company is allowed to charge a federal buyer for a named job title has been on file here since the rate screen shipped. Nothing could reach it. GSA records those rates by job title and records no NAICS code, no product class and no buying office against them, so no page in the product could point at a rate from the thing being bought, and the one question a services bid turns on had a screen of its own that nothing linked to. The notice page and the price band now both carry the roles that match the work: what each one typically bills an hour, the middle half of the range, how many awarded rates and how many companies stand behind it, the experience and education usually asked, and how often it needs a clearance. The match is made on the words describing the work, taken from the code title, the product class title and, where the notice has been read, the processes named in it. Words are cut back to the part their relatives share before they are compared, so that a code called Engineering Services reaches a role called Systems Engineer, which comparing them whole never would. A product buy gets no rate card at all: a rate for a systems engineer under a notice for bearings would be worse than an empty space.

What the work goes for is now on the notice itself, in full: the median, the usual range, how many companies have ever won this work at this office and what share the largest of them holds, how often it drew a single offer, the competed and sole-source split, the fixed-price share, how much of the money went to a small business, and which way it is moving. All of it was already computed and the page was printing two figures of it. The buying office named on a notice is a link now, to the page that says what that office buys and what it orders rather than competes, which was one click away with no click to make. That office page answers the half it was missing: whether it competes the work, whether anybody new has won there lately, and how long it takes from a notice going up to a signature. And a notice carries the defense budget lines behind its code, which is the only thing on that page looking past its own deadline.

The guides, the longer reads, the notes on how each figure is made and the glossary are reachable from inside the app. They were all out here on the public site, which is to say behind the member's back: looking up what a CAGE code is, or how the recompete score is worked out, meant leaving the product to come and find them. They still live here, because this is also where they are indexed and where one can be sent to somebody who is not a member; the app now lists them and opens them, and reads the glossary inside itself, searchable, because that is the one piece somebody wants beside the record it explains. Also in this release: a record page whose left-hand story is short, a vehicle nobody has ordered against, a vendor with one award, no longer leaves two thirds of a column empty.

An action button that fails now says why, rather than putting its label back and leaving somebody looking at a button that appears to do nothing.

The app would not load at all on the three previous releases, reporting that no route matched the URL to a member who was signed in. The endpoint list added two releases ago gave the route registrar an optional flag so it could list the routes without registering them, and every callback on that hook is handed a first argument of its own, so the flag arrived as an object and every route in the product returned before registering. Also in this release: a signed-in member can reach their account, their invoices and their plan from this site, which they could not before, and the pricing cards are a little over half as tall because a plan row is now one line rather than two.

Notices were not arriving for a day at a time. They arrive again, and the same fault was found and fixed in every other source at once. Also here: the sitemap is aligned, one card per group, flowing down columns rather than across rows.

Waiting for a list now looks like work rather than like nothing: it used to draw the same box a table draws when nothing matched, so the most repeated moment in the product said no results while it was still fetching them. The first column of a table stays put while the rest scrolls, which is what makes an eight-column award row readable on a laptop. The sort arrow no longer resizes the header under the cursor. API access lists every endpoint the key opens, read off the route table so it cannot go stale, rather than showing a credential and nothing else. And the engine illustration on two pages here still drew the inside of the engine and counted them in its caption, three releases after the prose on those pages stopped doing it.

An audit of the Ask rewrite the day after it shipped, and it had four faults worth naming. The AI answers stopped far sooner than they should have. Every grant total was wrong, because a lookup was reading the wrong value as its row limit. A question about open notices searched closed ones as well. And a model that started circling could spend every remaining search on the same query. Also in this release: sorting by vendor on the longest lists, which was far slower than it should have been; two more facet menus that still collapsed to the value you had just ticked; sorting on the Agencies table, which had none; and four screens that explained missing data by naming how the work is rationed, which is not a customer's business.

Ask now reaches the whole product. It used to turn a question into one set of filters, run one list and write four sentences from the first twenty rows, which is a good search box and is why the AI read as narrow. It now works a question across all nineteen lists here, as many searches as the question needs, changing what it looks for as it learns: awards, notices, subawards, assistance awards, buying offices, vendors, vehicles, rules, bills, exclusions, contracting people, justifications, budget exhibits, labor rates and the rest. Every figure in an answer is counted over the whole matching set rather than over the sample, so it can tell you what three offices hold of two billion and mean it, and the answer lists the searches behind it with what each one found so any figure opens as a screen you can check. The rule it works under has not moved: the searches run through the same queries the screens use and under your own plan limits, it may use only what they returned, and where the record does not settle a point it has to say so.

An audit of the app and the site, and four kinds of fault came out of it. Multi-select was still single-select on thirteen screens: ticking one value left that menu holding only the value already ticked, because a menu was counted over the list as it had already been filtered. Notices and Awards were fixed two rounds ago and nothing else was. A multi-select of set-asides returned nothing at all on three screens, where the comma list was compared as one string against a column that holds one code. Every date and time in the app was drawn in the reader own timezone while the server sends UTC, which on the east coast is four hours of drift on every last-refreshed, saved, checked and closing time in the product. And sorting by vendor had no index to walk on either table that offers it, so one click sorted the whole matching set from scratch. On this site: the pages that described how the work is scheduled no longer do, the Award History page understated the span of the record by more than two and a half times, and the recompete horizon named two plans out of the four that have one.

The status screen stopped reporting how the business runs. Several figures that describe the running of this company rather than the state of the record were on a screen members read. None of them answers the question that screen exists to answer, which is whether what you are looking at is current, and they read as a fault to anybody who does not run it. They are gone from it now. The screen says which sources are current and nothing else.

Notices were stopping for a day at a time, and the cause was not the one the last three rounds fixed. Today's notices now come first, ahead of everything else, and nothing can crowd them out. A source that is behind also recovers on its own rather than waiting to be restarted by hand. "Stopped" became "pending", because a source that is a little late is usually just late. And the Data status screen keeps the running of this company for the owner: a member sees whether the record is current, which is what that screen is for.

Five faults the owner found in the running product. A facet was counted over the list as already filtered, so ticking one state left the State menu holding only that state and there was no second one to tick, and the same narrowing cut every other menu at the same time. Notices were asked for only once a day, so the part of the reading reserved for them had nothing to do. The history was crowding out work that has to be current, which is why notices reported stopped on a quiet site; current work is now protected from it. And one screen reported a figure it believed to be impossible, because one of the two numbers behind it was hard-coded.

An audit of the whole product at production size, which found things no test at fixture scale can see. The tie-breaker on every sorted list was pinned to descending while the column followed the reader, and an index is stored in one direction, so the second click of any column header stopped being a walk of the index and became a sort of everything that matched: twelve milliseconds against a hundred and twenty seconds. The facet counts decided whether to sample by comparing a total that had already been capped at the plan's row limit, so on the two cheapest plans they never sampled and the award list did not finish loading at all; the flag that marks a withdrawn award was in the WHERE of every query over that table and in none of its indexes, which is why. Two figures on the home screen were joining the award table for no reason and took twenty-seven seconds between them. The screen that reads every source at once was computing twenty-two breakdowns and discarding all of them. Every list sum is bounded and priced now, and a new suite checks the SHAPE of these queries, because the shape is wrong at every size and only expensive at one of them.

Eight screens had been empty since the previous upgrade and said "Still counting" for ever: Agencies, Set-asides, States, Buying offices, Companies, PSC codes, Subawards and Grant awards. Nothing was building those figures at all. They build now, and a notice posted this morning can never wait behind a fiscal year from 2019. Narrowing a list became one row of buttons with every value behind each, searchable, instead of five rows showing six of two hundred; several values can be chosen at once, which the filters have always accepted and nothing ever sent.

A design pass driven by one screenshot. Every AI capability now carries a line saying what it gives you, and they draw as a grid rather than as eleven rows of two different heights. One shape per job across the app: an allowance is a labelled bar wherever it appears, a long secret is a field with a copy button, an identifier you paste elsewhere carries one too. The account screen is five bands sized to their contents instead of two fixed columns running twelve hundred pixels past each other.

Twenty defects in one sweep, and five more introduced while fixing them and caught before they shipped. The ones that mattered: a duplicate route registration meant pinning a saved search and un-sharing one answered with success and did nothing, and a double-escaped query silently killed every saved-search alert whose filter text began with one of five letters.

Seats and AI packs are sold beside a plan rather than built into it, priced by tier for seats and flat for briefs, with the checkout and the grant fields to match.

Five things the product did not have. Integrity records: the criminal, civil and administrative proceedings connected to federal work that a contracting officer has to read before deciding whether a company is responsible, on every company page. An empty one is not a clean one and no screen here pretends otherwise, because that system is reported only by companies above a threshold of roughly ten million dollars in active awards, records drop off after five years, and this product checks companies a few a day, so four different situations produce an empty record and only one of them is reassuring. Second, 8(a) graduation dates and the incumbents they make vulnerable: the programme runs nine years and cannot be extended, and inside it a company can be handed contracts nobody else may compete for, so an 8(a) incumbent holding work that runs past its own certification is the clearest opening this record contains. There is a screen listing every one of them, sorted by the size of the gap. Third, who to team with, which asks three questions of three different records and ranks the answers into one list: the primes who actually report subawards in your codes, the firms who win at the same offices in codes you do not work in, and the registered firms holding a set-aside you do not. Every row says why it is on the list. The first of those is read off the subaward record rather than the award record on purpose, because a company winning large awards in your code is your competition and a company reporting subawards in it is a customer, and from the award table alone the two look identical. Fourth, bulk actions: tick rows and track or watch all of them in one request, with an answer that separates what was done from what was already there from what a plan limit refused. Fifth, saved searches became saved views: they keep their sort as well as their filters, pin to the menu, and can be shared with your account, where anybody can open one and only the person who made it can change it.

A full sweep for defects, and it found two that mattered. The document library stored whatever file extension it was handed. There is an allow-list now, checked before a single byte is written. And a team seat could invite: the seat cap was counted against whoever asked rather than against the account, so somebody holding one seat on a ten-seat plan had an empty member list of his own, counted one seat used, and could add nine people who could each add nine more, every branch getting its own monthly AI allowance on one subscription. Invitations belong to the owner now. Also fixed: a figure the server declined to count read as $0 rather than saying it was not counted, which made a market worth billions and a market worth nothing look identical exactly when the server was too busy to tell them apart; the capture plan crashed on any notice published without a NAICS code, which is most sources sought; Clear the filters did not clear, because six screens read the last filters back out of storage the moment the address bar was empty; sorting a list kept you on page four of the old order; and a request that finished after you had moved on replaced the screen you moved to with an error, despite having succeeded. New in the same round: a settings screen. Theme, how tightly the lists are packed, the email digest, the phone shortcuts and the keyboard reference, in one place instead of five. Every fix went into the test rig as a regression, including two structural checks that catch the whole class rather than the instance.

The AI works on every record screen instead of one. A brief of a notice was the only thing a member ever saw it do, so it was on one screen out of fifty-eight and every other screen handed you a table and left the reading to you. There are now readings of a rule, a bill, a company, a buying office, a defense budget line, a NAICS market, a vehicle and an award, each shaped for the decision that screen exists to support rather than a summary of the table above it. Ask answers the question too: it used to turn a question into filters and hand back rows, and it now writes a short answer from the rows that came back, so every name and date in it is in the table underneath. The rule behind all of it is one sentence and the whole design rests on it. The server rebuilds the same figures the page was drawn from and sends the model those and nothing else, so it has no query access, no search tool and no memory of you, and it cannot assert a figure you are unable to check because it was handed the screen. Where the record does not settle a point it is required to say so in that line rather than fill it. Every block is labelled with the model that wrote it and what it was written from. Also in this release: Ctrl-K opens a palette that reaches every screen, every NAICS code by number or by name, your saved searches and the last places you were, with whatever you typed offered to Ask as a fallback. The letters j and k move down and up a table and Enter opens the row. Twenty-one lists that used to say "no rows" now explain which of the three reasons they are empty for and offer the links that answer each. And the site finally says what the AI does: it had been describing the version from 1.9 while the app grew past it, which is a capability nobody was told about and therefore one nobody was paying for.

The app looks like a desktop application now rather than a dashboard. One radius scale, one shadow scale, one focus ring, a translucent toolbar with a hairline under it, and colour saved for the one thing on screen that is actually urgent. The sidebar is light in light mode: it had been the single near-black slab on an otherwise white application, which fights the content for attention and forces the selected row to shout louder still to be seen inside it, and it is now the quietest surface on screen. Dark mode keeps the deep panel, where quiet is what it already was. A selected row is a filled pill rather than a line with a bar on its edge. The stat tiles lost the orange stripe down the left of every one of them, which marked nothing because all four had it, and their numbers are larger and set in the text face rather than the monospace. The card grid stops stranding a fourth card alone on a row of its own. Lists and tables gained row highlighting and roomier cells, tags became pills, and the scrollbars stopped drawing a grey trench down the side of every panel. A new test reads the stylesheet as a design system: a custom property that was never defined does not throw and does not warn, it is simply dropped, so every token used must exist, every token defined must be read, and every colour must be restated for dark.

The menu down the side of the app. Thirty-four links, all on screen at once, in three groups named Work, Explore and You, and seven of them drawing the same little house because those screens had no icon of their own and the code quietly falls back to one. Nothing told you which group a screen would be in, so finding anything meant reading every label. The groups now answer a question rather than naming a category, they collapse, and the four a member uses daily are open: eighteen rows on screen instead of thirty-four with no screen removed from the product. What is open is remembered per member rather than per browser, and the group holding the screen you are on always opens itself. Four names were wrong rather than unclear and are fixed: Opportunities is Open notices, Incumbents is Companies, Contacting people is Contracting people, and Grants is Grant awards, which had been sitting four rows under Grant opportunities meaning the opposite thing. Every heading is a real button, so the keyboard and a screen reader get what the mouse gets, and a new test checks the menu as a structure so a row without an icon or a screen without a home cannot ship again.

Rules and funding. An award is a record of something that already happened. A rule says what the buying is about to require, and it says so first: a proposed rule runs a year or more ahead of the solicitation it produces, and the comment window is the only point in the whole process where a supplier has any leverage. The Federal Register, Regulations.gov and Congress.gov publish it in three different places, none of them knows what a NAICS code is, and nobody else in this market joins them to the awards. Every rule and bill here is tagged to the codes it touches and every tag says which of three layers made it: quoted in the document, matched on a phrase, or suggested by AI, because a screen that renders those three identically is lying about the two weaker ones. Every rule shows what actually went out under those codes in the last twelve months, in money, and where a rule rewrites a FAR or DFARS clause it lists the open solicitations citing that clause right now. Also in this release: key personnel, which reads the positions a solicitation names, the minimums stated in its own paragraph, and which of your people states each one and which they are silent on. And One market, where one filter asks every source at once.

Your library: a member's own past performance, proposals, resumes, rate sheets and certificates, kept where the requirements are, so a shredded solicitation shows which of your own documents answers each requirement and how many of them your library already covers. Nothing in it is sent to a language model; the matching runs on this server and prints the words it matched on. Files sit outside anything the web server will serve and are scoped to the account rather than the person, checked on every read and every write including the search. A document that cannot be read for text is stored and marked rather than stored silently as a document with no words, and pasting the text in fixes it in a minute. Separately, the award-reporting website the government retired in February 2026 is gone from every page, guide and setting; the guide that named it moved and the old address answers a permanent redirect.

The shredder: a solicitation turned into a compliance matrix, where every row is the document's own sentence stamped with the section and paragraph it came from rather than a model's summary of it, plus the Section M evaluation factors that have no Section L instruction behind them, and the page limit, deadline and address pulled out first because those are conditions for being read. Five defects found by an audit and fixed, including four per-code pages that could run a long count while a visitor waited. The award feed this product read every day had been retired by the government and nothing here had noticed; contract awards now come from the SAM.gov Contract Awards API, with the identifier kept byte for byte so four million existing awards update rather than duplicate, and every page, guide, source list and link that still described the old award-reporting site as live has been corrected. The defense budget: every procurement, research and operations line the Department is asking Congress to fund, from the Comptroller's own exhibits, read by column name rather than by position so a moved column cannot shift a figure, and refused loudly rather than imported as nulls when a workbook cannot be placed. Sole source: the published justification behind every award made without competition, treated as an object rather than a notice, sorted by the date the lock comes off rather than the date it was posted, with the authority taken from an explicit citation and statutory citations translated, and any field the document did not state left empty rather than estimated. The compare page now names the two companies in this market that publish their prices and quotes their published figures.

The errors after an update are fixed, and the cause was ours. A figure that had not been counted yet answered with an error, and just after an update that was every figure at once. It answers normally now and the screen says the number is still counting while showing everything else. On top of that a busy moment now makes a page slightly thinner rather than failing it: the rows you asked for are never optional, the extras around them are. The release itself is Routes, which asks a different question from everything else here. Every other screen answers what is open. This one answers how you get to it: each path from your company to the office that spends money on what you make, ranked by how hard it is to walk rather than by how much is at the end of it. Compete directly where an office really does take on newcomers, subcontract to a prime who already pays subs in your code and already wins there, get onto the multiple-award vehicle the orders run through, or find out that the only opening is a recompete and when it is. It works because one company is one company across five separate federal files, which took months to build and is the reason nobody else has this. Plans also compare row for row now.

A notice and the award it became, joined at last. Both federal records carry the solicitation number and nothing had ever put them together, so the question of who won a notice was answered by looking for other awards at the same office in the same code, which is an inference. It is now a fact: this notice, that contract, that company, and the number of days between the posting and the signature. Where an office has enough matched pairs the page states its measured lead time, which is the one number nobody publishes and the difference between planning for eight weeks and planning for eight months. Maps are now on every screen whose data has a place in it rather than on one tab of one screen, and clicking a state filters the list under it. States and regions is a screen of its own; the app had linked to it since the analysis workbench shipped and the route never existed, so every one of those clicks landed silently on Today. Country is captured, so an award performed in Germany stops being indistinguishable from an award with no place recorded, and the overseas work is drawn as equal squares by region. On the engine: two cache guards were inverted, so a notice page ran a grouped count of every award and discarded it. The filter-and-sort pairs the busy screens actually use are faster. And about forty columns that were being stored and shown nowhere are now on the page, including what a GSA rate does in the option years, which is what a services firm needs before quoting a three-year bid.

The notice page, rebuilt. A member put ours next to SAM.gov's and ours was thinner: no attachments where SAM.gov showed a link, a bare product code where SAM.gov printed its name, a city where SAM.gov printed a full address, and no opportunity type or inactive date at all. Some of that the product never read, some it read and threw away on the way into the database, and some it stored and never printed. All three are fixed, and a test now proves structurally that a field read on the way in is a field written and a column that exists, so the next one cannot be lost quietly. The page is organised the way SAM.gov organises theirs so the two can be read side by side, and then carries the four things SAM.gov cannot tell you: what the work goes for here, whether this buyer ever takes on anybody new, when they will buy again, and what they order through a vehicle rather than competing. Every long record page gained a rail of its own sections, which on a phone is a strip rather than a column. And the home page now says what is different here as six facts you can check rather than as adjectives.

Three more federal sources and a screen for asking your own question. GSA CALC brings about 450,000 awarded hourly rates: what a named role actually costs, which is the only place the government publishes price by role rather than by contract and the number a services firm needs before quoting. Grants.gov brings grant opportunities, including the forecasted ones an agency has not posted yet, each one opening onto what has already been paid under the same assistance listing. The SAM.gov exclusion list brings the companies barred from federal work, marked at the top of their vendor page and checked against every teaming partner the product suggests before it suggests them. And Analysis and maps takes any filter that works on the award history and describes that slice six ways: trends, shares with a concentration index, categories, two maps, vehicles and awardees. The map is squares rather than an outline, because federal money is concentrated in DC, Maryland and Virginia and on an area map DC is one pixel. Three things a competitor has are still missing and the compare page says why: protest history, national stock numbers and state and local opportunities have no machine-readable source between them.

The contracts you have to be on before you are allowed to bid. Most federal money never passes through a competition anybody can watch: it leaves as an order written against an IDIQ, a GWAC, a schedule or a blanket agreement somebody won years ago, and an order is never posted on SAM.gov. Every one of those vehicles is now its own page in a directory you can search: whether it is still writing orders, when it ends, how many companies are already on it, how much of the ceiling is left, and what has actually been bought through it. Sorted by when each ends rather than how large it is, because the end date is the only part a company on the outside can act on. How many companies hold a vehicle is counted from the ones sharing its solicitation number, after each is resolved to a single identity, because a single-award vehicle held by one firm under two spellings would otherwise look like a door with room on it. The capture recommendation now says outright when most of an office's money in a code leaves as an order rather than a competition. And the home page leads with what is open today rather than only with what was bought before.

The way in. Every tool in this category tells you how big a market is. This one tells you whether you can get into it, which for a supplier without federal past performance is the only question that matters. For every code at every buying office: how often somebody wins there for the first time, how large that first win typically is, which door it came through, and whether first-time winners ever win again. Markets are ranked by how open they are rather than by how much money is in them. Open one and you see the suppliers who actually got in, with what their first award looked like. It can only be measured because a company is counted as one company however the federal record spells its name, which is the same work that makes everything else here trustworthy.

Capture plans. Open any notice or recompete you are tracking and there is a workspace for it in three parts: a brief assembled from the federal record, a pursue or walk-away recommendation with every signal written out, and a short plan you write on top. The recommendation weighs competition, incumbency, set-aside fit against your own certifications, whether the award size suits a company your size, how much runway is left, and whether the incumbent has options that move the real recompete date. A set-aside you cannot bid is blocking rather than merely a low score. Teaming candidates come from the subaward record. The brief is recomputed every time you open it, so a plan cannot drift out of date with the facts behind it.

Award history now reaches back to FY2001, which is as far as the federal award record goes, on every paid plan. Depth is the product here, not an upgrade. Fixes: an award page could sit on "Loading" forever, because the query behind one panel had no index and read the whole table every time you opened an award; a screen that had failed looked exactly like one still loading, on all twenty-seven of them; on a phone every filter group reserved seventy-two pixels of empty space under its label and the pipeline stage buttons stacked one per row; the pipeline is a card per item now rather than a seven-column table on a phone; the buttons on the shortcuts sheet were invisible in dark mode; and the membership and checkout pages were never actually loading the site stylesheet, which is why they stayed white.

What the work goes for. For every code and class of thing with five awards behind it, the last five years of what was actually paid, as a range rather than a number: the median, the middle half, and the full spread. Once across the country and once for every buying office, so the two can be read against each other. An average would be no use, because one ceiling award among a hundred deliveries gives a figure nobody has ever been paid. Beside every band, the four things that decide whether the price is reachable: how often it is competed, how often only one supplier answered, whether it is fixed price, and how much of the money one incumbent already holds. Two filters for finding a way in: markets where half or more drew a single offer, and markets no supplier dominates. Every band opens onto the awards behind it, because a price nobody can check is a price nobody will trust.

Three things the record could always support and nobody was showing. A forecast of what each buying office is likely to order next, worked out from what it has actually bought, with a window and a stated confidence, and no prediction at all where the history is too thin or too erratic to justify one. A directory of the contracting officers and specialists named on the notices, searchable by office and by the codes they put out. And a filter for work that drew only one offer, which is usually the shortest route to a first award. The plans, the compare table and the feature pages carry all three.

An audit of the whole engine, with every fix pinned by a test so it cannot come back. The ones that mattered: a detail update that quietly blanked fields it had just written, recompetes failing outright on the list, the dashboard and the calendar feed, unsubscribe links that could be forged, a dropped connection permanently writing off rows it should have read again, and the daily feed of new awards stopping for as long as the ten-year history was loading. Every screen in the app now shows an error when a request fails instead of a spinner that never stops.

Subaward records, contract modification history and the awards under a vehicle, all from USASpending. The public pages now show a code, vendor, office or agency and what the record holds for it; the organized detail behind that opens with a plan. A plan and the RSS feeds are retired. Pages for the three jobs people arrive with: small business contractors, capture and BD teams, and subcontractors and teaming partners. Headline numbers in a row share one size instead of each being shrunk on its own length. Header dropdowns no longer close while you move the mouse into them. The obligated-by-fiscal-year chart is readable again. It covers the years the engine loads rather than every start year on record, so a base contract from the nineties and options running into the thirties no longer add a bar each; the amount is printed over the tallest and the newest bar with the rest on hover, and the year labels thin out as the row gets crowded. The fiscal-year range under the headline numbers follows the same window.

Notice text, notice history and the registry stopped competing with each other, so all three keep arriving. A source that had to wait now comes back on its own.

The engine now tells a wait from a retry from a real error, and never gives up on a source that drops connections. The award detail pass can no longer starve the history load. Tidier facet chips on the lists. Tile numbers never wrap.

Alert webhooks to Slack, Teams, Zapier or Make on Capture and Enterprise. RSS feeds of open notices by code, state, PSC and set-aside, with an RSS button on every open-notices block. A share image for pasted links. On phones: filters fold behind a button with the active count, tables scroll sideways with readable columns, live tiles sit two across, and the home page cards are full width again.

The ten-year history load never ran: USASpending accepts NAICS filters of 2, 4 or 6 digits and the engine asked by 3-digit subsector. Every subsector now goes to the API as its 4-digit industry groups, retired groups included, and the history that had stalled was started again.

SAM.gov sources that had to stop now pause and resume on their own.

The engine reads several sources at once without ever doubling up on the same work. The daily catch-up waits for the newest fiscal year to finish loading, so history is read before the last four days are read again. Readiness checks for reading, notices, mail delivery and the payment gateway. On the site, the bento sections use a fixed grid that never leaves a lone card, and a round of visual fixes on the home page.

Coverage widened to the whole listing: every 2022 NAICS code with retired codes, every awarding agency, every fiscal year back to FY2001, every SAM.gov notice, registrant directory by state. One page per NAICS code on the site, all 1,012, plus the NAICS index by sector. Accounts and profiles with access only on an active plan; open tier and no-card trial removed. Enterprise plan. Audience pages for large suppliers and services firms.

First release. Award history from USASpending with detail pass, SAM.gov notices with descriptions, SAM.gov contract awards, SAM entity directory, entity resolution with reviewed merges, recompete table and score, buyer roll-ups, NAICS explorer, watchlists and email digests, AI briefs, classification and plain-English search, CSV exports, team seats, API keys, four plans with trial, and this website with guides, NAICS pages, buyer profiles, region pages, insights, glossary and FAQ.