[ancsingabor]

The questions themselves are good — I'd keep every one. Two of the answers were wrong, and this is the one page where a wrong answer costs something, because it is the page a reviewer reads to decide whether to trust the rest.

The release rule (Q3) was wrong in both directions

It said: "a retrieval or prompt change must attach an eval diff, and a regression blocks it."

	the page said	the rule actually says
a recall@10 regression	blocks the merge	named and justified, not vetoed — a pp threshold at n=40–60 is false precision, which is why the rule stopped being ">2pp recall@10" in PR #21
what blocks	(not stated)	any drop in citation validity or quote fidelity
who enforces	implied CI	review; CI does not run the eval
what triggers it	retrieval, prompt	retrieval, chunking, embedding, prompt

The missing qualifier matters most: telling a reviewer "a regression blocks it" invites them to assume a CI gate, and there isn't one. And omitting chunking from the trigger list is exactly how someone re-tunes the chunker without a diff — the thing .claude/project.md rule 6 and chunk.ts both forbid.

The licence weakness still had its pre-#24 framing

The row read: "May corpus text be sent to a hosted embedding API? … The bake-off starts with local models, which don't raise the question."

ADR-003 § Extended widened that to generation. Composing a prompt puts the retrieved units themselves into a request to Anthropic — in production, per question, for as long as the system runs. Local candidates mitigate the bake-off half; nothing mitigates the other one, and there is no local fallback anywhere in the design. So the mitigation as written was false for the half that matters.

It now gets a short row plus a paragraph naming it as the only open item that can invalidate a design rather than narrow a choice. Understating it on the page whose own TL;DR says "a design explained without its costs is not credible" was the worst possible place for it to sit.

A question a reviewer certainly asks, that wasn't here

"Why Claude?" Four files assume Anthropic, nothing ever compared it, no ADR argues it — and the harness cannot settle it, because it scores no generation model at all. Added as question 6 with the honest answer, plus a weakness row.

That made "The first five" into "the six that come up first" — and that counted heading is itself the tripwire named in the new § What I would do differently bullet:

Treat a counted claim as a tripwire. "Seven silent greens", "five war stories", "three pytest modules" — each drifted while the prose around it stayed true.

The summa:I.q2.a3 bullet also now carries why the fix took six passes: the last two copies were prose, so no grep for the locator form could find them.

1,415 words, in band. docs:lint green.

🤖 Generated with Claude Code