A statement that may not answer its question is never kept as an example

[sandeep-agami]

Spec: ACE-120 (feature reconcile-evidence, brief F10)

Stacked on #298 (which stacks on #291, #290, #289, #288). Base is ACE-119-reconcile-prose-evidence. Merge the five below it first, then retarget this to main. Do not merge without manual review.

Summary

Phase 3e keeps a match row as a worked example, which teaches the AI a question-to-SQL pairing. A sound statement paired with the wrong question was the most harmful thing that step could keep, and nothing stood in its way: the ledger grades a statement against the semantic model and the data, never against the question it came with. Now the skill reads the two side by side and the ledger withholds a doubtful row.

What changed

• Phase 1.5g in the skill, for every row with both a question and a statement: doubtful when the grain differs, the measure differs, a filter is present the question never asked for or absent when it did, or the time window differs. Writes question_fit.json as {"fit": "plausible" | "doubtful" | "no_question", "reason"}. A judgment made by reading, named as the one part of the ledger that is; it can withhold a row and never proves anything about the semantic model.
• The ledger's question_fit part: plausible → confirmed, the note saying it was read, not measured; doubtful → unresolved with the reason, so a matching number grades match_unverified and never reaches Phase 3e (whose pinned text is untouched); no_question → no part; file absent after a successful run → unresolved, by the same missing-input rule every expected file follows.
• Phase 3b.5 shows a doubtful fit under "What couldn't be checked" and says how the person settles it: reword the question or the statement and re-run the row.
• part-ledger.md, statement-check.md (step 10) and evidence-row.md name the step, the file and the part.

Verification

Run tests with uvx --with pytest-cov --with-editable "packages/agami-core[model,server]" pytest … or uv run dev.py check.

• New: four ledger tests (plausible, doubtful with row_status reaching match_unverified, no question, missing or unknown after a successful run and nothing after a failed one); an end-to-end fixture whose statement counts items for a question about orders and is withheld with matching numbers; pins on 1.5g's four doubt signals and file, on 3b.5, and on the three reference docs. The two clean-row expectations gain the question_fit part deliberately.
• The reconcile files together with the verb tests: 271 passed.
• uv run dev.py check at the top of the stack: ruff and gitleaks clean, 5766 passed, 12 skipped, 1 pre-existing failure: tests/test_golden_run.py::test_a_client_that_cannot_be_found_still_fails_as_a_generation, which fails identically on untouched origin/main on this machine because a client binary exists at /opt/homebrew/bin/claude.

Review round 4

A panel pass found the one way past the gate this spec builds: a no_question fit removes the part, and nothing checked that the row really carried no question, so a malformed file could send a sound statement paired with the wrong question to the keep-offer. Fixed in the second commit, with tests:

• The findings verb cross-checks the row record against the ledger: a row with a question and a statement whose ledger has no question_fit part is never treated as a statement that held.
• 1.5g and statement-check step 10 say the file is written for every statement row and no_question is only for a statement that came alone. part-ledger.md names question_fit among the parts a missing file opens, and the 3b.5 example quotes the code's own note.
• Fixture 14 asserts the fit is the only open part, and a twin row read as plausible reaches match.

Rebase note (2026-09-13). Replayed onto main at 99e44a8 with every branch below it; content unchanged, force-pushed with lease.

🤖 Generated with Claude Code