[replynodes-ai]

Problem

The live 20-company keyless benchmark (airtable.com) failed the fail-closed pricing
contract with:

pricing.model: no linked evidence excerpt supports the claim text

The Airtable pricing HTML embeds plan JSON. The camelCase key
costPerUserPerMonthInCents matched the per-unit model cue
(?:per|/)\s*(?:user|seat|...) (it matched the PerUser inside the key), was
title-cased into the meaningless claim Peruser, and its linked excerpt was the raw
JSON fragment where the key stays costPerUserPerMonthInCents — so excerpt_supports
could not ground it.

Reproduced deterministically and against the live page
(model_value='Peruser', excerpt = raw JSON, excerpt_supports -> False).

Fix (bounded, harness/extraction only)

tests/run-company-research-keyless-e2e.py:

1. Require a word boundary around the per-unit pricing cue
   ((?<![A-Za-z])...(?! [A-Za-z])) so a camelCase machine key is never read as readable
   pricing prose. Real prose (charged per seat, $20/user/month) still matches.
2. Fail closed in extract_pricing: a cue only becomes the model when its linked
   excerpt actually grounds the derived value (reuses the validator's excerpt_supports).
   If ungrounded, the cue is skipped and extraction continues (or falls back), never
   emitting an unsupportable label.

With the fix, Airtable yields a grounded model Per Seat (excerpt: "plans are charged
per seat") instead of Peruser.

Tests

• tests/test-company-research-v2-quality-fixtures.py: new deterministic regression
  fixture for costPerUserPerMonthInCents — extraction-level (model != Peruser,
  excerpt-grounded) and contract-level (two-page brief validates end to end, pricing
  known + grounded) — plus a fail-closed body where the JSON key is the only per-unit
  text (no pricing) and a control for valid model prose.
• Verified RED (reverting the runner change) → AssertionError: Peruser;
  GREEN after the fix.

Evidence (all on this head)

• Focused: test-company-research-v2-scenarios.py + test-company-research-v2-quality-fixtures.py pass.
• Package: tests/test-package.sh passes (schema fixture, scenarios, quality fixtures, distribution, registry links).
• git diff --check clean.
• Full 20-company keyless E2E DISABLE_TELEMETRY=1 python3 tests/run-company-research-keyless-e2e.py --output ...:
  schema_validation passed for all 20 companies, quality_failures: none,
  unsupported_claim_count: 0, airtable pricing.unknown: false with model Per Seat.

tests/company_brief_validator.py is unchanged; no unsupported claims are accepted.