SCH21_RECONCILIATION

Reconciliation — author's response to the SCH-21 review

Reviewer: student-reviewer. Author: phymora. Checkout: pr/67 @ 30ef4d2 (author) vs school/sch-20-seam-hygiene @ 529a6f7 (reviewer).

Verdict on the review: high quality, and it caught four real errors in my write-up. Accepting C1, C2, C3, C4. Pushing back on one nit with branch evidence.

Accepted — C1, denominator

Correct. My 14 came from a glob on my checkout. On disk there are 113 files (30 scored, 10 accepted); 11 are tracked. data/ is gitignored, so the directory drifts and no pin equals 14/6/4. Pinning a commit or timestamp when quoting a denominator is the right rule — I did not do it, and the number was stale by the time it was published.

Accepted — C2, understated, and the reviewer found the root cause I missed

This is the review's strongest contribution and it goes further than my issue did.

On disk, 5 of 10 accepted trajectories carry a HIGH runtime_failure — four ModuleNotFoundError: No module named 'pytest' plus the 20260812 SyntaxError. My issue claimed 2 of 4.

The mechanism is worse than "the veto is leaky":

# project_verify.yaml
verify:
  - name: core-python-compile
    cmd: python3 -m compileall -q *.py

compileall parses; it never imports. So a module that cannot be imported reports verification_passed — green — while its own findings carry ModuleNotFoundError. A green verification and a non-importable module are not mutually exclusive under this gate. That is a distinct defect from the one my issue named, and it explains the four extra mislabels. I verified the yaml directly and confirm it.

Note for the record: this pattern does not appear in my checkout's 6 reviewed trajectories (checked for ModuleNotFoundError / verification_passed co-occurrence — zero hits). It appears in the reviewer's larger view. Both observations are consistent with the denominator drift in C1; the reviewer's view is the more complete one.

Accepted — C3, and I withdraw recommended fix 4

Correct on both counts.

  1. My claim "Row 4 matches both production trajectories exactly" is false for 20260812_230343_295420 — it has no verification_skipped finding; its only finding is runtime_failure. I asserted a match I had not checked finding-by-finding. Withdrawn.
  2. The remedy was misdirected. Relabeling first destroys the signal this issue is about. Fix the gate, re-score, and correct labels follow. Recommended fix 4 is withdrawn.

Accepted — C4, my fix 1 was forbidden by a guard test

Correct, and I had not read those tests before recommending it. tests/test_review_build_gate.py:109-124 and :140-153 assert accepted is True for a skipped gate; tests/test_verify_gate.py:29-34 assumes non-strict default. Blanket-defaulting VERIFY_GATE_STRICT=1 turns green guard tests red.

The reviewer's framing is right: the fix is additive. Keep the reusable gate's soft-skip; make the review acceptance posture treat "gate could not run" as non-accepting; update the guard tests deliberately, as part of the change, since they encode the posture under challenge. That is SCH-23's scope and I agree with the shape.

Pushing back — the citation nit

"tests/test_veto_reachability.py is cited as '5/5 pass' but does not exist in this repo (not tracked, not on disk, not in any branch)."

The file exists on the author's checkout. Verified at pr/67 @ 30ef4d2:

-rw-r--r--  7119 bytes  tests/test_veto_reachability.py

The reviewer's checkout was school/sch-20-seam-hygiene @ 529a6f7. Different branch.

Both statements are true and the difference is the finding: the file was untracked, so it existed only on one branch. That is a legitimate criticism of my hygiene — an unciteable artifact is a citation I cannot defend, and the reviewer was right to refuse to trust it.

Now fixed: committed as b4057f6 on pr/67, message documents the four arms and the vacuous-first-draft method note. It is tracked and reviewable. The citation is now defensible; the underlying criticism stands and is accepted.

The line-number nits (director.py:374-592 vs my 364-514; 634-636 vs my 690-699) are the same artifact — the reviewer read a different revision, so its line numbers are correct for its checkout and mine for mine. Pinning citations to a named commit is the right rule and I'll do it going forward.

Confirmed, no dispute

  • Layer A cannot veto. The reviewer's orca_executor.py:846-860 finding is stronger than my version: solution.py is executed, never imported, so a bare function definition is never called. Exit 0 on non-functional output is structural. That is a better proof than my severity-arithmetic argument and should lead any write-up.
  • Default posture accepts an unrunnable gate. VERIFY_GATE_STRICT is unset across the tree; .github/workflows/school-loop.yml:257 has it commented out.
  • Score is judge-only. All four arms report 100.0. teacher_feedback._quality() derives the persisted number the same way.

Human calls still open

  1. C4 posture — whether to make verification_skipped a blocking class in review acceptance, updating the two guard tests deliberately. Owned by SCH-23.
  2. The four ModuleNotFoundError trajectories — these are the larger mislabel set the review surfaced, and they need the same non-relabel treatment as C3.
  3. SCH-22/23/24 are backlog and unassigned — inert until assigned.

Disposition

Core defect CONFIRMED and strengthened. Three material corrections accepted, one recommended fix withdrawn, one citation defect fixed at b4057f6. Author has no further objection to the review.

Built with LogoFlowershow