# Attify verification program, rules v5

How a vendor's AI agent goes through verification, for every job type. What each job type audits (which records count, the claim, the checks) is in its standard ([docs/standards/](standards/)); everything else is here. Published 9 November 2026. Nothing here is public yet: every page sits behind the site's access gate.

## Who takes part

- **The vendor**: an organization on Attify (owner, members) that runs one verification program per product and job type.
- **Its customers**: organizations that use the vendor's AI agent. Each is invited by the vendor and connects its own help desk (or the job type's platform) to Attify, or, when the agent runs on the vendor's own platform, consents to the vendor's deployment and reconciles it ([Witnessed data](#witnessed-data)). The customer's consent is recorded with the connection: who gave it, when, and its scope (what Attify reads, how long text is kept, who sees what). The customer can disconnect at any time; Attify then revokes its authorization at the platform where the platform allows it, and otherwise links to the page where the customer revokes it.
- **Traction reports (rules v2)**: the scope also says the vendor may share a traction report with investors it chooses, with the customer as a numbered customer, never named unless it opts in ([TRACTION.md](TRACTION.md)). A customer that agreed under rules v1 is listed as an exception in traction reports until it agrees again.
- **Independence (rules v3)**: the scope also says Attify reads the email addresses of the customer's help-desk admins, to check it is a company separate from the vendor, and may email one of them to ask whether it uses the product. A customer that agreed under rules v1 or v2 counts toward verification only once it agrees again on its invitation page.
- **Witnessed data (rules v4)**: a customer invited for a vendor's deployment agrees to a scope of its own: the vendor sends Attify its agent's interactions with the customer's users as they happen, text is kept the same 30 days, the customer reconciles each month, and it can withdraw at any time.
- **Attify**: reviewers label the human audit; admins can revoke a status, with a reason.
- **Recusal.** Attify never publishes a score, badge, profile or Index entry naming a company in which its founder, Dor Tagger, has a financial interest, or a client of PointLine (the agency Dor Tagger runs). The list is kept in config (`api/recusals.toml`) and checked at every publication gate; such a program can run privately, and it never publishes ([/independence](https://attify.app/independence)).

The vendor sees each customer's results and consent records, never their conversations.

## What it rests on

- **Connected data, or witnessed data the customer reconciled.** Attify reads every record of the period from each customer's platform with the customer's own authorization (connected), or receives each interaction from the vendor's agent as it happens, signed and timestamped, and the customer checks it against its own records (witnessed, below). Uploaded results never feed a verification.
- **The standard decides what counts.** For customer support ([Standard #1 v0.10](standards/support-resolution.md#vendor-verification)): conversations the AI agent handled; human-only conversations never.

## Witnessed data

Many agents run on their vendor's own platform, where there is no help desk for Attify to read. Their vendor streams each interaction to Attify through the Ingest API, an SDK or OpenTelemetry ([INGEST.md](INGEST.md)). Since the vendor chooses what to send, witnessed data counts toward a verification, a badge or a published score only when all of these hold:

1. **The vendor attested its right.** Before a deployment can stream, the vendor attests in the terms that it has the right to share that customer's interaction data with Attify; the record keeps who, when and the terms version.
2. **The customer consented in Attify.** The customer itself, invited by the vendor for that deployment, records its consent: who, when, the scope and the retention, the same way as for connected data, and passes the same independence checks (for a deployment, check 1 reads the consenting person's domain, and the customer's own help desk when it connected one too; check 3 asks through that help desk, and otherwise an Attify admin resolves it). Without the customer's consent, witnessed data feeds only the vendor's private report, and its text is purged on the same schedule as any other.
3. **The customer reconciled it.** Each month, the customer states the month's total from its own records or the vendor's invoices before it sees what Attify received; it passes within `RECONCILE_VOLUME_TOLERANCE` (default 2%). Then Attify draws `RECONCILE_SPOT_CHECKS` interactions of the month (default 10) under the customer's secret seed (derived from the evidence key; its SHA-256 commitment is in the audit log and, with the result, the public changelog), and the customer says whether each matches its own systems. Where the customer connected its own help desk, the interactions Attify finds there by their `external_ref` are answered by it. A month passes when its total and every spot check match. Witnessed data counts once the latest complete month passed (the month before still counts `RECONCILE_GRACE_DAYS` days into the next, default 10); a month that fails stops it counting until a later month passes.
4. **The stream is whole.** Late events, gaps in the sequence numbers and sequence breaks show on the vendor's program, each customer's row and the deployment. An interaction with a late event is never scored as witnessed. Above `INGEST_MAX_GAP_SHARE` missing sequence numbers (default 0.1%) or `INGEST_MAX_LATE_SHARE` late interactions (default 1%) in the scoring period, the deployment is degraded and stops counting until it recovers. The same holds where the standard audits every interaction the AI handled with the vendor's claim (support resolution): the vendor sets that claim, and an interaction it leaves without one would leave the audited population, so above `INGEST_MAX_UNCLAIMED_SHARE` (default 2%) of those interactions ending with no `outcome.claimed` either way, the deployment is degraded too.

**Agent platforms and observability tools.** Attify also reads an agent platform or observability account directly (Vapi, Retell AI, ElevenLabs, Bland, Voiceflow, Langfuse, LangSmith). Whose account it is decides what it supports, decided when it is connected and recorded with the connection:

- The **buyer's own account**, connected by the customer through its invitation, is connected data, like its help desk: Attify checks that the account's people, as the platform lists them, include the domain of the person who consented, and that nothing makes it the vendor's.
- The **vendor's own account** (or one it holds a key to) is witnessed data: the vendor connects it as a deployment (attesting its right first), Attify reads it every few minutes with the platform's own timestamps, and it counts only under the four conditions above. The first read, which reaches 90 days back, is late throughout and never counts.
- A customer offering an account Attify finds is the vendor's (a vendor's deployment already reads it, or its people include the vendor's domain), or one whose platform names no one on the customer's domain, is refused with the reason, in the audit log: a vendor's own account never passes as connected.

Outcomes are checked in the customer's own systems wherever possible: the standard says how (for support, the customer's help desk holds repeat contacts). A score with any witnessed customer is reported as witnessed, and its public profile says how many of its counted customers were witnessed. Traction reports read connected data only, for now.

## Customer independence

A vendor could invite addresses it controls and act as its own customers. So a customer counts toward a vendor's audited score only when it passes all four checks below. The vendor sees each customer's status (passed, pending, failed) with the reason, check by check; Attify's admins see the same, with the address Attify asked.

1. **Separate organization.** The person who agreed is on a company domain: a free email domain (Gmail, Outlook, Walla and the like; `FREE_EMAIL_DOMAINS` adds more) never shows a separate company. That domain is not the vendor's (the domains of the vendor organization's members, and of any help desk the vendor connected itself, subdomains included), and the person is not a member of the vendor's organization. The help desk is run from the same domain: at least one of its admins, as the platform lists them, is on it. It is not a help desk the vendor connected itself, and not one another customer of the program connected first. The vendor cannot invite a free or own-domain address in the first place. The vendor's own domain is fixed when it opens its first program: the owner who opens it signs in with a company email, and that domain stays the vendor's (shown on its profile) whoever its members are later; a vendor whose people all sign in on free addresses cannot open one. A customer on a domain known by the vendor's own name (acme.com, for Acme AI on acme.ai) fails until an Attify admin rules it genuine.
2. **Real history.** The help desk shows activity from before the invitation date: at least `INDEPENDENCE_HISTORY_DAYS` days of it (default 30), with at least `INDEPENDENCE_HISTORY_RECORDS` records (default 50) from `INDEPENDENCE_HISTORY_PEOPLE` distinct people (default 20), and, where the platform says when the help desk was created, created that long before too. Attify reads it once, after the first sync. A customer that connects too long after its invitation for Attify's read to reach that far back fails, and is invited again. The job type's standard says what counts as a record and a person.
3. **Direct confirmation.** Attify emails a random sample of each program's customers (through Resend, or the local outbox in development) at an address read from the customer's own help desk, never the one the vendor gave, and asks whether the company uses the product. The sample comes from a secret seed per program, derived from Attify's evidence key: each customer that passes checks 1 and 2 gets a rank under it, by its company (the registrable domain of the address it was invited at), so a company invited again keeps its rank, and any answer it gave stands for the new invitation; it is asked when its rank falls below `INDEPENDENCE_CONFIRM_SHARE` (default 0.5), and, once the program has enough customers for a score, the lowest-ranked ones are asked until at least `INDEPENDENCE_CONFIRM_MIN` (default 1) are. A witnessed customer with no system of its own connected is always asked: everything Attify knows of it comes from the vendor's own stream, so with no address of its own to ask, an Attify admin rules on it. A rank never changes, so the vendor cannot predict or redraw it. Every draw is logged: in the program's public changelog (the customer's letter, the draw number and the seed's SHA-256 commitment) and in the audit log (with the rank). The address goes to the help desk's main domain, in an order under the same seed. A customer drawn counts once it confirms; while it has not answered it is pending, after `INDEPENDENCE_CONFIRM_DAYS` days without an answer (default 14), or once it says no, it fails; either way it stops counting until it is resolved (an answer, which can come late and can change, or an Attify admin's ruling).
4. **Consequences.** The vendor attests in its terms, when it creates the program (or later, before anything counts), that every customer it invites is genuine and independent, and that if Attify finds one is not, Attify revokes the score and says why in the public changelog. An Attify admin who finds a customer fake rules so, with a reason: the program is revoked at once, publicly, with that reason. A ruling that a customer is genuine (confirmed by another route) resolves its direct confirmation.

## Coverage

The vendor chooses one rule when it creates the program:

- **Every invited customer.** All of them must be connected, synced and independent for a score.
- **Attify's random selection.** Attify draws which connected customers count (as many as the minimum), with a secret seed, when the vendor asks for its first judged score, among every independent customer connected by then. The draw is logged in the program's changelog with the candidates, the customers drawn and a SHA-256 commitment to the seed, so it can be checked later. It is drawn again, and logged, only when a drawn customer disconnects or stops being independent. Until the first draw, the private daily score reads every independent connected customer.

**Minimums** (configurable per program): 3 customers and 500 counted records in the period (conversations, appointments), or the standard's own minimum (sales meetings: 150 meetings). Below either, there is no score.

## Scoring

- **Period:** a rolling 30 days, ending at the latest sync of every counted customer, or the standard's own period (sales meetings: 90 days, since sales teams book fewer meetings). Records not yet settled (a meeting or appointment whose time has not passed long enough ago) are left out until they are.
- **Daily:** the rules tier for every counted customer and in aggregate, to keep the data fresh and the vendor's private trend current.
- **Each published score:** a judged audit under the job type's standard, per customer and pooled across customers, each with its 95% interval. A pooled interval treats each customer as a cluster ([METHODS.md](METHODS.md#pooled-rates)): never narrower than the interval on the pooled counts, wider when customers differ. A new one is made every 30 days, or when the vendor asks (after an estimate of its cost).
- **Every score** carries the standard version, the evidence tier and the provenance.

## The claimed unit

A claim is tested in the unit the vendor states it in, exactly as the vendor makes it. Standards #2 and #3 have two: meetings booked or meetings held, appointments booked or appointments kept. An Attify admin reads the vendor's unit from its own claim or billing (its site, its pricing or contract terms, the dashboard its customers see) and records it on the program with the page and the day it was read. A booked claim is tested only on the failures that contradict a booking (not booked by the AI, not new, a person stepped in or staff fixed it, the judge's check); its cancellations and no-shows are outcomes, shown beside it as the share of its claims held or kept, cancelled and no-show, never hidden. A held or kept claim is tested on every check. Every score, status, badge, profile, attestation and registry row names the unit. Until it is recorded, scores test the stricter unit and none publishes. A change of unit is an entry in the program's public changelog, and the next score and status are made in the new unit, so a unit never changes silently.

## Publishing: private first, then continuous

- **The first audit is private.** A vendor's first judged score, and every one until the vendor publishes, is private: the vendor sees every number, and nothing is on the registry, a profile, a badge or the Index. The vendor chooses whether to publish it, once its gates hold (below). Nothing private ever becomes public without that choice.
- **Publishing is a commitment.** When the vendor publishes, every later score publishes too, after its review window, whatever it shows. A published score cannot be hidden, withdrawn or edited, and nobody can change a number: a score is frozen when it is made.
- **Stopping verification.** The vendor can stop at any time. The profile then reads Lapsed, with the date it stopped and its full history (every published score, the changelog, the attestations) kept public. A score already in review when the vendor stops still publishes when its window ends, so stopping never discards a score; the program stays Lapsed. The same holds when the vendor's organization is deleted: what was published stays on the record.

This trade-off is deliberate. A private first audit lets a vendor see its number before it commits, so some vendors that would read poorly never publish: there is selection at entry. It is offset in two ways: once a vendor publishes, every score publishes, so a vendor cannot publish only its good months; and the Attify Index pools buyers' own data, which no vendor selects.

## Review window

Each score after the first published one is a private preview for **7 days**. The vendor sees every number and can add one public comment, shown next to the score. Then it publishes.

## Publication

A score publishes (the first one when the vendor chooses, later ones when their review window ends) once all of these hold:

- **The judge is validated against people for the current standard version's judged line** (the standard's Human audit section; a patch version keeps the validation of the version it patches). A records-only score, where the customers' systems held no conversation for a judge to read, used no judge: this gate and the next do not apply, and the score says its evidence tier.
- **The score's own human audit is complete**: every conversation in its human audit sample is labeled.
- **Its customers are independent**: at least the program's minimum of the customers it counted pass all four independence checks now, and none of them is pending or failed.
- **Its witnessed customers are reconciled**: every witnessed customer it counted still consents, has reconciled, and streams whole (above).
- **It meets the minimums** (Coverage) and **no recusal applies** (Who takes part).
- **Its claimed unit is recorded**, where the standard has more than one, and the score was made in it ([The claimed unit](#the-claimed-unit)).

Until then it waits, and the vendor's page says what for. A score with a customer that failed an independence check can never publish: it is withdrawn when its review window ends, and a new score can be made without that customer. A published score goes on the vendor's profile (`/v/<slug>`), the registry, the badge and a signed attestation.

## Status

One rule for every job type. Every status change is logged in the profile's public changelog, with its date and reason.

| Status | When |
|---|---|
| Claim confirmed · within 5.0 points | Attify is 95% sure the vendor's count overstates audited outcomes by at most the claim tolerance: the gap's upper 95% limit is at most 5.0 points |
| Claim not confirmed | Otherwise: the gap may be larger than the tolerance, or the data are too thin or disagree too much to show it is not |
| Lapsed | The vendor stopped verification; or no fresh data for 14 days (a counted customer has not synced, or disconnected and was not replaced); or the program fell below the minimums (fewer counted customers passing every independence check than the minimum); or no new score followed the published one within the rescoring interval (`RESCORE_DAYS`, 30), its review window and `LAPSE_DAYS` |
| Revoked | An Attify admin revoked it, with a public reason, for instance a customer proven fake. Final, and public for good even for a program that never published: the registry lists the vendor, its company domain, the product, the date and the reason, and keeps them after the vendor's organization is deleted. No surface shows a revoked score's numbers. The vendor's company domain cannot open a new program for 12 months. |

The claimed rate is the vendor's own counting rule as the standard measures it (for support, the AI's own resolution flag), so the status compares the vendor with itself. The gap is the claimed rate less the audited one, on the same records, so it is never below zero: any failed claim makes it positive. The question is whether it is small, shown with evidence. Its interval is the published one, pooled with each customer as a cluster, so few customers, or customers that disagree, widen it and make confirming harder, never easier; the comparison uses the number as printed (one decimal). A standard measured on its claims (sales meetings, bookings) has a claimed rate of 100% by construction, so its claim is confirmed when the audited share's lower 95% limit, in the vendor's claimed unit, is at least 95.0%; every status there names the unit, a booked claim's status states its claims' share held or kept, cancelled and no-show, and every status states the customers' own teams' kept share, as context. A status never stands alone: it is always shown with the audited number, the gap to the claim, the job type and the period, and a confirmed claim always with its tolerance. A Lapsed program returns to a claim status only with its next published score.

**The claim tolerance** is a published parameter: **5.0 points, since 9 November 2026**, reviewed after the first real programs. Every status, badge and attestation states the tolerance it was decided with (`claim_tolerance_points`). A change takes effect on a stated date and is recorded below with it; every program whose status it moves gets a status event in its public changelog naming the change, so a status never changes silently.

## Attestations and badges

- **Attestation:** a JSON statement of each published score and the program's current status, signed with Ed25519. The signed bytes are the `payload` object serialized with sorted keys, no whitespace, UTF-8, and integral numbers written without a fraction (220, never 220.0), as RFC 8785 does. Public keys are at `/.well-known/proven-keys.json` (JWKS, `OKP`/`Ed25519`, by `kid`). An attestation names its program, standard version, evidence tier, provenance, period, status and numbers (the audited rate, the claimed rate and the gap), when it was issued and when it expires (44 days later: one period plus the lapse limit). A new one is issued on every publication and every status change.
- **Verify:** `/verify` checks any attestation: the signature, whether it has expired, and whether it is still the program's current one (a newer attestation or a revocation makes it stale).
- **Badge:** served from Attify's domain only, always drawn from the current status: the SVG at `/badge/<slug>.svg` (full, compact, or an inline chip; light, dark or automatic; footer, pricing, hero and email sizes), a PNG for email signatures, and the one-line embed (`/badge/embed.js`), which links the badge to the profile and opens a hover card read live from `/badge/<slug>.json` with a link to `/verify` for the signed attestation. Every badge reads "Audited by Attify" and carries the audited rate, how far it is from the vendor's claim ("12.6 pts below claim"), the job type and the period, in every variant and for every job type, with its mark: "Claim confirmed · within 5.0 points" or "Claim not confirmed". A Lapsed badge carries the numbers of the last published score with that score's period and "Lapsed". A Revoked badge carries no numbers, since the score it would show was withdrawn. Only a published score from connected data has a badge: a program in its review window or an upload has none (404). The profile's share card and a dated deck image with a QR code to the live profile come from the same status. A copy hosted elsewhere is not a badge.

## Comparisons

The registry ranks and compares vendors only within one job type and one claimed unit, and shows each rate next to its contact-reason mix and evidence tier, because the mix differs between vendors and resolution differs between reasons.

## Changelog

- **v5** (2026-11-09): one status rule for every job type, and private-first publishing. A published score shows "Claim confirmed · within 5.0 points" when Attify is 95% sure the vendor's count overstates audited outcomes by at most the claim tolerance (the gap's upper 95% limit, each customer a cluster, at most 5.0 points; effective 9 November 2026), and "Claim not confirmed" otherwise, replacing Verified and Not verified and the margins (5.0 points by default, 25.0 for sales meetings, 20.0 for bookings); falling below the minimums now reads Lapsed. Pooled intervals treat each customer as a cluster. A vendor's first audit is private until the vendor publishes it; once published, every score publishes and nothing can be hidden; stopping verification reads Lapsed with the history kept, also when the organization is deleted. Recusal: no company in which the founder has a financial interest, and no PointLine client, is ever published. The standards that changed with it: Standard #1 v0.10, #2 v0.3, #3 v0.3, and for the claim tolerance (the owner's decision the same day, after the pre-launch review; it replaced a rule that confirmed a claim only when the gap's interval reached zero) Standard #1 v0.11, #2 v0.5, #3 v0.5. Revocations are public for good, published or not, kept after an organization's deletion, with no numbers on any surface, and a revoked company domain opens no new program for 12 months (pre-launch review M7, M31). The claimed unit (the owner's decision the same day): a claim is tested in the unit the vendor states it in, read from its own claim or billing and recorded on the program; a booked claim's cancellations and no-shows are outcomes beside it; a score publishes only in its recorded unit; the registry ranks within a unit (Standards #2 and #3 v0.6).
- **v4** (2026-10-31): witnessed data. A vendor whose agent runs on its own platform streams each interaction to Attify (the Ingest API, an SDK or OpenTelemetry), attesting first that it may share the customer's data. It counts toward a verification only once the customer consented in Attify and reconciled it each month (its own total, then a seeded spot check), and while its stream is whole; until then it feeds the vendor's private report only. A fourth publication gate: every witnessed customer a score counted is reconciled. Customers on their own help desks agree to nothing new.
- **v3** (2026-10-30): customer independence. A customer counts only when it passes four checks: separate organization, real history, direct confirmation, and the vendor's terms; a score publishes only when enough of its customers pass them; a customer proven fake revokes the program. The consent scope adds the help desk's admin addresses and Attify's direct question. Customers that agreed under v1 or v2 agree again before they count.
- **v2** (2026-10-30): the consent scope covers traction reports shared with investors (numbered customers; named only by the customer's own opt-in). Judged scores, publication and the badge need the Vendor Verification plan; inviting customers and the daily rules tier stay free, since traction reports read them too.
- **v1** (2026-10-30): first rules.
