card-tnc-review

name	card-tnc-review
description	Quarterly re-verification of card T&Cs against official issuer sources — diffs the facts the nudges depend on, skips unchanged sources by hash, sweeps for promos. Run with /card-tnc-review.

Card T&C Review

Keeps the numbers the agent asserts honest against what the issuers actually publish: bonus caps, min-spends, pool splits, partner/whitelist merchant lists, exclusions. Those numbers live in tools/card_optimiser.py (nudge thresholds and merchant patterns), pwa/index.html (the card meters), and your Supabase cards / card_strategy rows. Runs in Claude Code (WebFetch/WebSearch available); nothing here touches the Render deployment directly.

Cadence: quarterly, or immediately after any nudge or PWA figure looks wrong. Trigger it by typing /card-tnc-review.

The ledger is the state

docs/CARD-TNC-LEDGER.md holds, per source: the URL, a facts block (the extracted numbers/lists we depend on, in canonical form), a SHA-256 hash of that facts block, and the last-verified date. The hash is over the extracted facts, never the raw HTML/PDF bytes — issuer pages churn markup weekly; the facts are what matter.

That file is not in this repo — it describes whoever's cards are actually in play, so it is created on your first run (see "Baseline mode" below) and is yours from then on. Same for any rendered strategy page you choose to keep alongside it.

Procedure

1. Read docs/CARD-TNC-LEDGER.md. For each source, WebFetch the URL and re-extract ONLY the watched facts listed in its entry (caps, min-spends, pool splits, partner/whitelist lists, exclusions, effective dates). Normalise: one fact per line, sorted, amounts as plain numbers.
2. Skip gate: hash the new facts block. Same hash as stored → update last_verified only, move on. This is what makes a quiet quarter nearly free.
3. Changed → produce a side-by-side diff of the facts block and STOP to present it before editing anything downstream. Changes to caps, min-spends, or pool structures are ask-first (they change nudge thresholds and PWA meters: tools/card_optimiser.py, pwa/index.html, supabase cards/card_strategy rows). After the user confirms, update in the same PR: the ledger entry, any strategy doc or page you keep, and any seed SQL — and list the live-table update statements they must run in the PR body.
4. Promo sweep (secondary, report-only): WebSearch each card name + "promotion"/"nerf", limited to the rewards-blog sources you trust, since the last run date; summarise anything material. Never auto-apply promos — they are time-boxed and the ledger tracks structural T&Cs only.
5. Baseline mode: a source with no stored hash gets its facts block captured and recorded without diffing. On a fresh clone that is every source — create docs/CARD-TNC-LEDGER.md from scratch, one entry per card you actually hold, and say so in the summary.
6. Output: a table — source · verified/changed/baseline · one-line note — plus the diff details for any changes. If nothing changed anywhere, say exactly that in one line.

Hard rules

• Official issuer pages/PDFs are the canonical sources; blogs are news leads only — a blog claim that contradicts the issuer page means "check the issuer page again", never "update the ledger from the blog".
• Never edit docs/CARD-TNC-LEDGER.md facts blocks without a fetched source in hand this session.
• A fetch failure is reported per-source ("could not verify"), never treated as "unchanged" — stale last_verified dates must stay stale.
• If an issuer PDF is image-only or paywalled, note it and ask the user to save a copy into samples/ (gitignored) for manual review.