Summary

• On a two-way call (the counterparty pitches its own product back), the extractor recorded the reverse pitch as this deal's pricing, competitors, stakeholders, decision process and next steps. It also told the model every non-owner speaker was buyer side, so a co-seller who ran the call was treated as the buyer.
• New shared seller point-of-view block for both the instance and judgment prompts: the whole seller domain is seller side (personal-mail reps fall back to the roster's rep tags), every field is read as "the counterparty buying the seller's product", the reverse pitch and seller-side self-description are excluded, and the counterparty's prior experience with the seller stays in scope. The judgment prompt now also gets the participant roster (preview and persisting runs).
• Precision fixes to 8 default field definitions from a per-field recall/precision audit of one real call: objection.raised/objection.response (capability questions are not objections; default-empty), feedback.moment, competitor.mentioned (the counterparty's existing stack is not a competitor), prospect.call_ran_over (default-empty), stakeholder.engaged (roles/dispositions must be evidenced), qual.gap, tactic.rapport_building. Content-pin hash re-pinned.
• PERSONAL_EMAIL_DOMAINS extended (googlemail, regional yahoo/hotmail/live/outlook) since it now gates the seller-domain rule.

Rollout already done (not in this diff)

The runtime reads definitions from each org's stored AOP, so editing the defaults only affects newly seeded orgs. The same 8 descriptions plus the objection.response.responds_to metadata description were already patched in production on the six orgs that carry taxonomy fields (Pirros, AspireIQ, Warp, Vooma, Admyral, Cedar) with a guarded, hash-pinned jsonb_set (each write only fires if the stored text still equalled the old shipped default; nothing was customized). Full pre-change backups were saved locally; a verification diff against the backups showed only those description paths changed. Warp has no qual.gap field (older instance-only shape), so it received 8 patches. Only Cedar has taxonomy extraction enabled today, so the other five see no change until their flag is on or a backfill runs. Existing extracted rows are unchanged; re-extracting them is a paid backfill and a separate decision.

Test plan

• pnpm run autofix:local chain: brand check, oxlint, deps:check, mail jest, mail tz tests, server vitest. Everything passes except apps/mail chatHistoryList.test.ts (date-sensitive, fails without any mail changes here) and a flaky LinkedColumns.test.tsx that passes in isolation.
• Server: full vitest run src/ green after merging staging (14,470 tests); tsc -b clean.
• New tests: participant-roster-pov.test.ts (domain rule, DEAL FRAME text, personal-mail fallback, roster tagging on a shared provider) and taxonomy-extraction-run.test.ts (judgment prompt includes MEETING PARTICIPANTS on both the dry-run preview and the persisting run).
• Battle-tested against the real prompt on one real two-way call with a GLM dry run (nothing persisted, ~13 runs, plus independent recall/precision audits). A 21-check scorecard went from 5 (baseline) to 20-21 on the final prompts; the remaining misses were single-run flakes that moved between runs.
• Review: three fresh thermo-review rounds; every finding was fixed. Round 3 still had four non-blocking 🟡 items, which were also addressed, but no fourth review was run (loop cap).

Verify in prod

This changes prompt text, not runtime branching, so there is no new log line. Use the existing structured log and the data itself:

• Axiom (cedar-prod):

  ['cedar-prod'] | where _time > ago(1h) and message contains "[taxonomy-stream] complete"
  | project _time, conversationId, eventsProcessed, instanceRowsWritten, judgmentChangesWritten
  

  CloudWatch: /aws/ecs/aws-prod-api/worker-service, same message.
• Data check on a two-way call after