Capstone: Trust/Verify Log

Capstone step 3 of 3 -- exercises co-17 (trust-vs-verify calibration) and co-24 (human-in-the-loop).

Every change made during this capstone's session, mapped to a trust-or-verify decision and a one-line rationale. A risky change (anything touching the actual conversion logic or its correctness) always maps to a human verification step, not a skim.

#	Change	Decision	Rationale
1	First generation of parse_duration.py (attempt 1)	Verify closely	Novel logic (co-17): the conversion table and token-consumption check are exactly the kind of "looks right, might not be" arithmetic this topic's ex-15 · spot-a-hallucinated-api and ex-47 · reject-a-confidently-wrong-refactor warn about -- ran the full test suite rather than reading and trusting the code.
2	Rejecting attempt 1	Verify closely / human decision	The failure was real (co-16), not a flaky test -- a human read the failure output, traced it to the exact wrong dict value, and wrote the rejection reason before any further generation was requested.
3	One-line fix ("m": 600 -> "m": 60)	Verify closely	Small diff, but it is the exact line that caused the failure -- re-ran the full suite rather than assuming a one-line change is automatically safe.
4	Extract-method refactor (step 3)	Trust lightly, verify the tests	Mechanical, no signature or behavior change claimed -- skimmed the diff for scope (co-15's checklist: does it touch anything beyond the stated refactor?), then let the test suite be the actual proof, matching ex-16 · delegate-boilerplate-safely's "skim + why that was sufficient" pattern.
5	Final acceptance of parse_duration.py + test_parse_duration.py into learning/capstone/code/	Human decision gate	The feature only ships once every one of the seven acceptance-criteria tests (from prompt.md) is green in the SAME run, confirmed by re-running the suite one final time against the exact files being committed -- not by trusting the last individual step's output in isolation.

Verify: every risky row above (1, 2, 3, 5 -- anything touching correctness) maps to an explicit "verify closely" or "human decision gate" entry with a stated reason; only the purely mechanical refactor (row 4) is trusted lightly, and even then the tests -- not a skim alone -- are the actual gate.

Capstone acceptance criteria -- final check

• The feature passes its tests: all seven acceptance-criteria tests pass in the final run (step 3's captured 7 passed in 0.00s).
• Every accepted change was verified before being built on: steps 2 and 3 each re-ran the full suite before the next step began.
• The log above justifies each trust/verify call.
• No unreviewed agent output reached the final diff: attempt 1 was caught, rejected, and never merged -- only the fixed-and-refactored final version lives in learning/capstone/code/.

Done bar: runnable end-to-end -- python3 -m pytest test_parse_duration.py -q against learning/capstone/code/ reproduces the final 7 passed transcript exactly; web-verified -- this capstone names no external protocol, vendor, or version claim requiring re-verification (the syllabus Accuracy notes cover the topic's own protocol/standard citations).

---

← Previous: Session