//ARCHITECTURE

How Verdict works

Eleven stages, run in this order for a reason: identity resolution before spend, extraction before verification, verification before scoring, scoring before a draft is ever written.

1. //00

   Validate + idempotency

   Ordinary code. Required fields, email/domain normalization. Replaying a submissionId returns the existing result — never reprocesses.

2. //01

   Identity resolution

   Runs BEFORE enrichment/scoring, deliberately. Exact email or domain match only counts as "confident" — name similarity, even a perfect score, never auto-merges.

3. //02

   Classify message

   Haiku labels the message into approved categories. Vendor pitches, job applications, and student research hard-disqualify here, before any research spend.

4. //03

   Find research source

   Looks up a domain in the seeded fixture corpus. An ambiguous or missing identity means no safe source to pick — enrichment correctly returns nothing rather than guessing.

5. //04

   Extract facts

   Haiku reads the source page and extracts only explicitly stated facts into a fixed field schema, with a verbatim supporting quote for each. There is no field for a score, a discount, or an instruction — structurally, nothing else can land here.

6. //05

   Verify facts

   Two independent checks per fact: a deterministic quote-in-source substring check (free, catches fabricated quotes), then a batched Haiku pass judging entailment and flagging quotes that read as instructions rather than facts. Only what survives both becomes "verified".

7. //06

   Evidence-sufficiency gate

   Ordinary code. Counts resolved ICP criteria against a floor. Below the floor: no score is ever computed — the lead returns the specific questions that would unblock it.

8. //07

   Score (if the gate is open)

   Deterministic points-based scoring across 7 criteria, each with cited evidence. A veto on B2B-fit or region disqualifies regardless of the numeric total.

9. //08

   Propose CRM change set

   A diff, never a write. Possible identity matches propose nothing — a human resolves the merge. Confident matches propose an update; new leads propose a create.

10. //09

    Draft outreach (gated)

    Only for qualified bands, and only using the lead's own message plus verified facts. Every claim the draft makes is independently re-verified against that same grounding text — never trusted from the drafting call's self-report.

11. //10

    Audit + jobs

    Every stage above writes a started/completed/failed audit event with duration. A Postgres jobs table tracks attempts — the durable queue, not an external service.