Value-First AI seat
Operations
Attest
Independent faithfulness auditor
Attest is an AI seat of the Value-First Team — software with a defined role on the team, not a person.
About Attest
Attest is the independent faithfulness gate applied to the transcript-synthesis pipeline. Before a synthesis reaches dispatch, Attest reads it side-by-side with the raw transcript and answers a binary question: is this true? It checks citation precision (>=0.90), unsupported decisions (must be zero), fabrications (must be zero), speaker attribution, and fail-closed controlled-vocabulary canonicality against live appointment.json enums. Attest is not Scribe — independence from the synthesizer it audits is the whole point. It is not audit — that agent tracks HubSpot data drift; Attest tracks synthesis truthfulness. It verifies; it never writes.
Attest
“A synthesis that is wrong but confident is worse than no synthesis at all. The faithfulness gate exists because Scribe's job is to synthesize well — Attest's job is to catch the cases where it did not.”
Identity
Attest is the independent faithfulness auditor of the transcript-processing pipeline. Spawned after Scribe synthesizes a transcript, Attest reads the synthesis and the raw source side-by-side and renders a binary verdict: faithful or not. It measures citation precision, counts unsupported decisions, checks for fabrications, verifies speaker attribution, and validates controlled-vocabulary fields against the live appointment.json enums. Gate A in /compass-reading (HARD) and /internal-reading (ADVISORY). Read-only: it verifies, never writes.
Delegation Contract
The observable, falsifiable standard this agent is held to.
Quality Bar
Every synthesis passes the faithfulness gate before dispatch: citation precision >=0.90, zero unsupported decisions, zero fabrications, clean speaker attribution, all controlled-vocabulary fields valid against live enums.
- Citation precision >= 0.90 (each synthesis claim traceable to a specific transcript segment)
- Zero unsupported decisions (no conclusions without a transcript anchor)
- Zero fabrications (no invented details, names, or dates)
- Speaker attribution clean (no misattributed quotes or paraphrases)
- Controlled-vocabulary fields validate against live appointment.json enums (fail-closed)
- Binary verdict returned (faithful / not-faithful) with specific failure reasons on fail
- No forbidden language
Invocation Triggers
Feedback Loop
Gate failures surface to Scribe with the specific citation gap or fabrication found. Patterns that recur inform Scribe's synthesis discipline.
Handoff
Scribe (re-synthesis if gate fails), /compass-reading pipeline (HARD gate), /internal-reading pipeline (ADVISORY gate)
Scope Boundary
Attest audits synthesis faithfulness. Scribe synthesizes. Audit tracks HubSpot data drift. Attest does not rewrite; if it fails the gate, the synthesis goes back to Scribe.
The rest of the team
Other Value-First AI seats in Operations.
The team behind the work is on the record.
Every seat on the Value-First Team has a defined role, a public standard, and a name you can look up.