Skip to content

Governance

Who decides what, what is recorded, how honesty is enforced, and how a release earns its place. With the gaps stated.

Illustration, not product data.

Who decides what.

Five kinds of action. Only the first is automatic, and it changes nothing outside Foundry.

Five kinds of action, who decides each, the gate it passes and what is recorded
ActionWho decidesThe gateWhat is recorded
Read public pagesAutomaticOnly public addresses are fetched. Page text is scanned for orders to the AI first.For each statement: the quote, the page address and the time it was read.
Write a note to your CRMYou, with a previewA permission you granted for CRM writes. A second push of the same company waits for the first.Who, what, which permission, the outcome, and any edit you made before it went.
Send an emailYou, one draft at a timeEight server-side checks, shown below, including a permission you granted for sending.The same record, with the words you changed compared to the draft.
Chat actions that change somethingYou, in your own wordsA domain you typed, or an export you asked for. A plain yes confirms exactly the one refusal just shown.The refusal, stored with the chat turn so a yes can confirm it.
Connect a toolYouYou paste a key or accept Google's consent screen. Permissions to write or send are granted by you, for yourself.Each permission: scope, target, expiry and, if ended early, when it was revoked.

How this is checked. tests/test_permissions.py shows no grant means no action, and that expired, revoked, other-user and other-scope grants are ignored. tests/test_hubspot_push_ledger.py and tests/test_sequence_send.py show a declined push or send is written to the ledger. tests/test_tool_guard.py covers the chat rule. No scheduled job sends email. The watch job re-reads watched pages and makes no model call.

Before one email leaves.

The server runs these eight checks in order and stops at the first reason not to send.

Before one email leavesThe order the server runs them in
  1. 1

    Draft belongs to your team

    A draft from another team is reported as not found

  2. 2

    Entry is live and not paused

    A stopped, replied or paused entry never sends

  3. 3

    Not on do-not-contact

    The company domain and the address are both checked

  4. 4

    No unfilled placeholder

    Bracketed instructions and {{tokens}} block the send

  5. 5

    Recipient looks right

    One real address, at the company's domain unless you confirm

  6. 6

    You granted email:send

    Revocable, and expires after 90 days by default

  7. 7

    This request wins the claim

    A second click or tab is told the draft is already being sent

  8. 8

    Send, then record

    Outcome written to the ledger, with what you changed

How this is checked. The order is written at the top of core/sequence_send.py, and tests/test_sequence_send.py has a test for each refusal: sent twice, two requests racing, do-not-contact by domain and by address, unfilled placeholders, no grant, Gmail not connected, another team's draft, a paused or stopped entry, and a recipient at another company. Nothing is sent by the tests.

The audit trail.

Three records, what each keeps, and the limit of each.

The three records Foundry keeps, what each holds and its limit
RecordWhat it keepsIts limit
Permission ledgerEach permission you grant and each action taken or declined: who, what, to whom, which permission, whether it can be undone, how it ended, what you changed, and the later CRM stage when read back.A failed database write is logged quietly and does not stop the action, so a record can be missing.
Evidence receiptsFor each statement that passed the quote check: the quote, its position in a saved copy of the page, the page's hash, and a hash linking it to the entry before it.A receipt proves our record is unchanged. It does not prove the statement is true. See the two lists below.
Run historyEach research or draft run: what was asked, the company, the sources, the provider that answered, the counts, the time taken, and any error. Also each correction a person makes.Shows that a run happened and how it ended. It is not a copy of the full output.

What a receipt proves, and what it does not.

A receipt proves

  • The page text we fetched, at the time shown, held this quote at this position.
  • Our record of it was not altered since. Changing the text, quote, position or an earlier entry breaks the hashes.

A receipt does not prove

  • That our database was not rewritten as a whole. A chain could be rebuilt consistently by someone with write access.
  • That the page was authentic at that address.
  • The time, to anyone outside our system.
  • That the statement is true. Only that the page said it.

How this is checked. The wording is the receipt page's own. ops/verify_receipt.py is a standalone checker that uses only the language's standard library, so a receipt can be verified without our code. tests/test_evidence_ledger.py pins the hashing rules, and the receipt route is read-only and limited to the account's owner.

How honesty is enforced.

A mechanical quote check on every statement, then a fixed panel of real companies before a release.

The four outcomes of the quote check, what happens to the statement, and the test that pins it
OutcomeWhat happensA test shows
VerbatimThe quote's words appear in order in the cited page. The statement ships as written.An exact quote passes
RepairedThe quote is close. It is replaced with the page's own words and the statement drops from verified to inferred.A near miss is repaired; a short quote is never repaired by similarity
Re-sourcedThe quote is real but sits on another page we fetched. The source is corrected.A quote cited to the wrong page is corrected, not rejected
WithdrawnThe words are on no page we fetched. The statement does not ship. It is listed as removed and counted, never dropped silently.A plausible fabrication is retracted, and so is an empty quote

How this is checked. core/claim_verifier.py is plain string matching, with no model call, so the same input always gives the same answer. tests/test_claim_verifier.py covers each outcome. The brief tells the rep how many statements were removed.

The release panel, with its coverage.

18
companies in the fixed panel
The same companies every run, so a failure means something changed. 6 are also run as campaigns.
10 of 18
briefs scored in the latest full run
16 October 2026. The rest returned a provider-capacity error: unscored, not passed.
0
hard failures in any recorded run
Citation integrity and invented timing, on the runs that produced a result, since 10 October 2026.
What the release panel scores
RuleHard or softWhat it checks
Citation integrityHardEvery cited source and evidence reference must resolve to a real one. One dangling reference fails the whole run.
Invented timingHardIn campaigns, every send date must be a real date before the meeting, in order.
Source floorSoftAt least two pages read from the company's own site, unless the brief itself explains why not.
Unknowns namedSoftA panel where too many briefs list no unknowns is flagged as over-claiming.
Confidence labelsSoftEvery evidence label must be one of the four real values.

How this is checked. ops/eval_panel.py holds the panel and the rubric, and ops/eval.sh runs it against the live site. Run results are recorded with their dates in the team log. The adversarial run is described on the security page.

How changes ship.

Six gates. Four block a deploy. One reports. One is run by hand.

The release gates, what each checks, when it runs and whether it blocks
GateWhat it checksWhen it runsDoes it block?
TestsThe backend test suite, and the frontend types, tests and build.The deploy script runs the backend suite. A CI file runs both on every push to main and every pull request.Blocks the deploy script. CI result is as configured in the repository host.
Craft gateText contrast of at least 4.5 to 1, no colours hard-coded in components, a fixed text-size scale, focus rings, no em dashes.Deploy script.Yes, nothing is deployed on a finding.
Python gateNo undefined names, store filters that match the schema, and the tenancy gate for missing owner checks.Deploy script.Yes, for undefined names, schema and tenancy.
TypeScript gateTypes and lint, which the production build does not check by itself.Deploy script.Yes.
Live verificationThat web and API serve one shared route table, plus live checks of the deployed site.After each deploy, and on a schedule every six hours.No. It reports a failure and the deploy script prints how to roll back.
Honesty panel18 real companies run through the live pipeline, 6 of them also as campaigns, scored against a fixed rubric.By hand, before a release that touches research, briefs, drafts, campaigns or the verifier. It needs real credentials, so the deploy script only prints a reminder.A hard failure means do not deploy. Nothing enforces that automatically.

How this is checked. The sequence is in ops/deploy.sh. The tests run on push and pull request through .github/workflows/ci.yml, and the six-hourly check is .github/workflows/post-deploy-verify.yml. We have not shown here that these workflows are running on a hosted repository today.

What the models see.

Public page text and your selling context, and in the chat, your messages.

What each feature shows a model provider and what comes back
FeatureThe model is givenWhat comes back
Research briefCleaned public page text of the company, and your selling context.The brief. Each quote is then checked against the page, not by another model.
First-touch draftThe same research text, your selling context and the company name.A draft you edit. Sending is blocked while it has unfilled placeholders.
Campaign planResearch text, the page title after scanning, and your context.A touch plan. Send dates are computed in code, not written by the model.
ChatYour messages and a short, scanned digest of each tool result.An answer, or a request for an action that the rule above may refuse.

How this is checked. Prompts are assembled in core/brief.py, core/draft.py, core/campaign.py and core/orchestrator.py. The providers are tried in the order Groq, OpenRouter, Google Gemini. We make no statement about what a provider does with the text; see Not yet.

Leaving and deleting.

What you can remove yourself today, and what you cannot.

You can delete today

  • A chat conversation, with its messages
  • A saved connection key
  • Your Gmail and Google Sheets connections (our copy of the token)
  • A permission you granted
  • A list, and the companies and people taken out of it
  • A person saved on an account, one at a time
  • Entries on your do-not-contact list

Not available as a button

  • Accounts, briefs and saved statements
  • Evidence receipts and run history
  • The team, or your sign-in
  • A full export of your data
  • A written deletion process with a response time

How this is checked. The delete and disconnect routes are the ones in the request handler: conversation, connector key, segment, grant revoke, do-not-contact entry, and the Gmail and Sheets disconnects. There is no route for the items on the right.

Who is responsible for what.

A short split, so nothing is assumed.

What Foundry does

  • Keeps teams apart and shows the checks that do it
  • Keeps the quote and page behind each statement, and withdraws what it cannot match
  • Asks for your permission before a write or a send, and records the outcome
  • Encrypts connection secrets before storing them
  • States what is not done, on this page

What you do

  • Read each draft and the quote behind its claims before you send
  • Keep your do-not-contact list current
  • Follow the laws and your email provider's rules for your own outreach. We make no compliance claim
  • Revoke access on Google or in HubSpot when you stop using Foundry, because disconnecting here deletes only our copy
  • Keep your sign-in private

Not yet.

What is not done, in plain words. This list is part of the record.

  • Independent certificationNone held. We do not use certification names.
  • A second approverA permission is yours alone and covers any recipient for 90 days by default. A second person approving a send is not built.
  • A guaranteed ledgerA failed database write is logged quietly and does not stop the action, so a record can be missing.
  • Receipts anchored outside our systemThe chain's latest entry is not published anywhere else, and the time on a receipt is our own clock.
  • A fully scored honesty panelIn the latest full run, on 16 October 2026, 10 of 18 briefs and 3 of 6 campaigns were scored. The rest hit provider capacity. 0 hard failures on what ran, but that is not full clearance.
  • A check for invented personal namesThe panel does not test whether a brief invents a person's name. That gap is known.
  • An automatic block on the honesty panelIt is run by hand. Skipping it is possible.
  • Self-serve deletion and exportSee the list above. There is no retention schedule either.
  • A promise about model providersWe make no claim about how Groq, OpenRouter or Google handle the text we send. Their own terms apply.
  • Automatic reply detectionReplies are marked by hand, with Replied on the entry. The Gmail access we ask for can send mail and cannot read your inbox, so Foundry cannot see a reply.
  • A daily limit on emails sentEach email needs your approval, but no cap on how many you approve in a day is built. Your email provider's own limits apply.
  • Named customersNone listed. Foundry is pre-revenue.

The security page has its own list of what is not done yet.

Ask us anything on this page.

Rather talk it through with a person?

Book a call