Job #765 — check of return #361's verification package: unable on this host
(POSIX resource), with the checker's substance verified in an auxiliary copy
Outcome: unable, blocker kind capability. The unchanged package cannot
execute on this host (Windows, Python 3.14.6): its checker does
import resource and resource.setrlimit(resource.RLIMIT_CPU, ...), and there is
no resource module outside POSIX. A worker on Linux or macOS can run the
unchanged package; nothing else is missing.
That said, the check was not left at the import error. A separate control copy
of the package (package files untouched; one added host-shim module resource.py
with a no-op setrlimit) ran the declared command to completion in 51.1 s,
exit 0, and its receipt reproduces the declared fingerprint exactly once
newline translation is accounted for. So the substantive question — does the
checker consume the submitted target and reach the claimed numbers — has an
answer: yes, and it is recorded below as auxiliary evidence, not as a verdict.
What was reconstructed and how
The package was rebuilt from the manifest alone, by SHA, into a clean directory (no local checkout used):
| manifest path | role | bytes | sha256 (verified) |
|---|---|---|---|
check-chordal-pursuit751.py |
checker | 3266 | c8ce2206…a694b61 |
check-chordal-triage749.py |
dependency | 5656 | 89d328b2… (manifest) |
chordal-pursuit751-inputs.json |
input | 7703 | 5a2a7948…e5e831 |
chordal-pursuit751-results.json |
target | 90156 | d2b003cc…815683 |
All four hashes match the manifest. Declared command (run with python, the
host's 3.14.6, since the python3 name on PATH is the Microsoft Store stub):
python check-chordal-pursuit751.py chordal-pursuit751-inputs.json \
chordal-pursuit751-results.json observed-verification.json
Observed, unchanged run: exit 1 in 0.1 s,
ModuleNotFoundError: No module named 'resource' at line 8 of the checker,
raised before any input is read, so no receipt was produced and every later
assertion is untested by that attempt.
The checker, inspected before execution
validate(inp)re-derives each shape's offsets from the gcd definition and asserts the submittedoffsets,N,first_ordermatch, that the completeda = 46841is excluded, and that the 15 shapes are pairwise distinct in shape-up-to-translation. It does not trust the input file.- It imports the manifest-declared dependency and calls
legacy.recomputeper shape, then buildsexpectedand executesassert target == expected— the submitted target is consumed, in full: minima, worst phases, all four histograms, both loss histograms and the pointwise-improvement counts, for all 15 shapes. It is not a self-check that regenerates an unrelated answer. - Coverage: every phase of
[0,18)×[0,22)×[0,28)×[0,30)= 392863 per shape × 15 shapes = 5,892,945 phase cases, and the receipt it writes asserts exactly that count. Excluded, and declared as excluded in the receipt:completed_shape_repeated: false,global_selector_ranking_reproduced: false— the checker does not reproduce the global ranking over all W starts. - Comparison rule: exact structural equality of decoded JSON objects (integers, no tolerance, no float), plus its internal controls.
Auxiliary run in a control copy (host shim, package otherwise unchanged)
resource.py shim: setrlimit a no-op, getrusage().ru_maxrss = 0. With it:
- exit 0, wall 51.1 s (the checker's own
RLIMIT_CPUis 120 s, so on a POSIX host the unchanged package fits inside its own limit); - stdout:
{"passed": true, "new_shapes": 15, "phase_cases_independently_checked": 5892945, "matching_minima": [...], "strict_robust_gain_shapes": 0, "completed_shape_repeated": false, "global_selector_ranking_reproduced": false, "negative_controls": {"changed_minimum_rejected": true, "changed_histogram_rejected": true, "removed_physical_slot_rejected": true}}; - the written receipt's sha256 is
3ef07bd1d93418efee418714a2de0c63c11de4f05fc40265435820966f5833a0— the declared fingerprint, exactly — after newline normalisation. As written on Windows it hashes451943b1…, becausePath.write_textopens in text mode and translates\nto\r\n(497 CRLF pairs). The declared fingerprint is of the LF form, so it is a POSIX-host hash of the same content.
Controls
- Internal (the package's own):
changed_minimum_rejected,changed_histogram_rejected,removed_physical_slot_rejected— all three reportedtruein the passing run; the checker fails loudly if any of them is not raised (rejected()re-raises "corrupt input accepted"). - External, run here in separate copies:
- corrupted target — shape 0's
minima[2]changed 10 → 11: exit 1,AssertionErroratassert target==expected. Detected. - missing record — the target file removed: exit 1 with an unhandled
FileNotFoundError. Detected, but as a traceback, not as a rejection: the checker aborts before writing any receipt, so a missing target produces no machine-readable verdict. That is a robustness observation about the package, not a correctness defect, and it is the one thing a repair would address.
- corrupted target — shape 0's
Shared components (declared, not hidden)
The checker's use of the dependency is declared in the manifest
(check-chordal-triage749.py, role dependency) and the module is guarded by
if __name__ == '__main__', so importing it is safe. The shared algorithm is the
whole recompute core: integer masks per phase, itertools.combinations(range(6),3)
spanning-tree enumeration (16 trees), pair-atom triangle budget, and the
sandwich first <= tree <= triangle <= exact. So this is one implementation
exercised through two entry points, not two independent derivations — the
"independently checked" in the receipt means recomputed from the inputs by the
checker, not independently implemented. Stated plainly because it bounds what
the receipt can mean.
What a reviewer should take from this
- The failure is host capability only: POSIX
resource(and apython3name that is not the Store stub). One targeted reassignment to a Linux/macOS worker should produce the manifest's declaredpasswith fingerprintdceb4f63…and receipt3ef07bd1…, with no package repair. - The auxiliary evidence makes that prediction sharp rather than hopeful: the recomputation reaches 5,892,945 phase cases, 0 strict robust gains, all three internal controls fire, and the receipt hash already matches.
- Two defects/limits to record against the package, neither fatal: the Windows text-mode receipt is not byte-portable (CRLF), and a missing target yields a traceback instead of a rejected verdict.