Stop the grounding gate rejecting drafts for saying "I'd"

[ralyodio]

What was broken

Asking for a draft on https://outreachgraph.com/approvals?channel=email&filter=needs_draft returned:

No draft: the wording kept asserting things nothing supports — I'd.

The draft was fine. extractClaims in packages/ai/src/checks.ts was wrong, in three ways that all trace back to reading an apostrophe as punctuation it is not.

1. A first-person contraction was read as a proper noun. The extractor treats a capitalised mid-sentence word as a name that must appear in stored evidence. I'd is capitalised mid-sentence, and nothing can ever support it — normalize turns the apostrophe into a space, so the claim becomes the token i d. Every draft written in the first person failed, retried (told to "remove these entirely: I'd", which it cannot sensibly do), failed again, and was withheld as failed_checks.
2. Correctly grounded names were rejected. Stripe's normalised to stripe s, which does not match stored evidence reading Stripe. Citing the prospect's own words counted as inventing them.
3. Apostrophes acted as quote delimiters. The quoted-phrase regex listed the straight ' as a quote mark, so two contractions bracketed the prose between them into a quotation: I've seen this. Curious what you'd try yielded the claim ve seen this. Curious what you.

Reproduced against the real checker before changing anything — case 1 returns exactly ["I'd"], which is the string the approvals card renders.

The fix

• Strip contraction and possessive endings ('s 'd 'm 't 'll 've 're, straight and curly) before a capitalised word is read as a name. I'd reduces to I and falls under the length floor; Stripe's reduces to Stripe and matches the evidence.
• Delimit quotations with double quotes only (straight and curly). Single quotes stay out — they are indistinguishable from the apostrophe in I'd.

The gate still does its job

Verified an invented product name, an invented figure, and a genuine double-quoted phrase are all still flagged, including inside first-person prose. Tests cover each.

bun run check is green: format, typecheck, 874 tests across 55 files, 0 failures.

🤖 Generated with Claude Code