[StepenkoAnatoli]

What was asked

"First research Tavily's current terms with the kit." Constraint C-6 put Tavily out of scope on 2026-09-19 with an expiry condition — revisit if a no-training tier ships — so the question of replacing SerpAPI with Tavily is answered by the terms, not by preference.

The answer

The trigger is not met. C-6 stands.

§6.5 and §6.7 are the same sections C-6 cited, unchanged: Tavily and its AI providers may retain inputs and outputs "for purposes of training", and those providers "may not be required to maintain the confidentiality" of either.

Because the finding is an absence, it was counted rather than skimmed:

term	occurrences
opt out	3 — all arbitration
enterprise	0
zero data retention	0
no model training	0

A prior was registered before collecting, and it held. That is worth stating: a protocol that only reports its surprises is advertising.

The bug, found by obeying the tool

preflight warned that one capture prints 176 render-failure notices, and told me to record my review "in a renderReview: front-matter line". I did exactly that, on the live corpus:

• before: PASS, 0 blocking, 1 warning
• after following the instruction: fail provenance/body-unmodified, and the warning still standing

verifyLedger hashes the capture whole, front-matter included. The check's own source comment, ten lines above the message, says the note must not go there — the message contradicted the code. It now names the EVIDENCE row and the [render-reviewed: ...] syntax the parser actually reads.

An existing test pinned that front-matter does not satisfy the review. Nothing pinned that the message does not send you there. Now something does, confirmed red first.

Also

• The topic signal scored 0.92 on a corpus that was 1/8 relevant. Not broken — the topic was seven common words, so the search matched words and the signal measured the same words back. topicMatch now documents that a high score is worth nothing without distinctive terms. It decides nothing, which is why it did no damage.
• A missing newline ran spent and topic together in the run summary.
• ARCHITECTURE said twelve checks; CHECK_NAMES holds thirteen. corroboration landed without it.
• SerpAPI timed out mid-run and the merge seam degraded to Firecrawl instead of collapsing. Unplanned, and the first time that path took a real provider failure in production.

Honest limits

• Seven of eight captures are off topic and cited by nothing. They are marked as such rather than deleted, because the ledger records that the credits were spent.
• Both unknowns rest on tavily.com/terms alone, recorded as accepted single-witness judgements. A vendor's terms have one author, and no third party can testify that a clause is absent.
• The terms carry no effective date and no version string, so nothing in the document can confirm it is unchanged — only that what it says today is what it said three days ago. That is why the trigger has to be a re-read.

Verification

• 874 passed, 0 failed (offline suite)
• Corpus: PASS — 0 blocking, 0 warnings, 14 passing, and under --strict
• Chain verifies; 8 captures, 8 ledger entries

🤖 Generated with Claude Code