0007: Catch-up records only what it observed
Context: Adopting the rules in an existing repo tempts the agent to fill templates with plausible guesses: expectations it never saw, architecture it inferred.
Decision: The catch-up prompt runs every QA tier it can and writes expect only from observed runs;
writes architecture.md only from verifiable code; logs failing tiers as papercuts without fixing them;
proposes oracles as proposed. Separate commit from code changes.
Consequences: Baseline is honest, possibly incomplete. Fixes come later as normal work. → reflected in architecture: adoption.
Alternatives: | option | why not | |---|---| | fill templates from intent/README | unverified claims become "docs" | | fix failures during adoption | mixes knowledge capture with code change; unreviewable |