Recorded-run contract
Planned before the durable run: two seeds, three lift depths, N=2..128 and N=1296, with exact floor/cap checks through 1296. No optimizer, parameter search, random input, or unbounded loop. An implementation check took about 13 seconds, so the durable run has a 60-second wall limit, one requested numerical thread, and a 1 MiB captured-output limit.
The independent algebra review covered the actual RESULTS.md, including the omitted-tail bill, N>=J^2 range, floor-factorial remainder, and the support-three restriction on the N=14 optimum. It found no mathematical failure. This is a proof-reading check, not kernel checking or external verification. The numerical checks use separate Arb and mpmath interval logarithms and exact rational assembly.
The source checker is self-contained. To reproduce the mathematics without the optional manifest runner:
.venv/bin/python hunts/paid_shortfall/construction.py --output /tmp/paid-shortfall-results.json
.venv/bin/python -m pytest -o addopts= -q tests/test_paid_shortfall.pyThe durable run's version-2 manifest also records its exact command, input hash, result hashes, resource limits, timing, and base commit. The worktree is intentionally marked dirty because the new source is checked before commit; its input hash binds the actual source bytes.
check-001: completed
The durable run took 3.136 seconds including process overhead. Its result reported PASS: 2 seeds, 36 seed-period rows, 7,776 floor rows, 3,888 exact base-6 deficit rows, 768 truncation cases, and 1,296 cap integers. All 3,072 inequalities passed on each of the two backends at each requested precision, 35 and 70 decimal digits. All 3,072 cross-backend overlaps passed. The exact N=14 primal/dual identities passed. Manifest validation passed.
id: paid-shortfall-2026-09-13-check-001
hunt: paid_shortfall
started: 2026-09-13T23:43:56.736590+00:00
finished: 2026-09-13T23:43:59.872646+00:00
ran:
- .venv/bin/python hunts/paid_shortfall/construction.py --output hunts/paid_shortfall/computations/check-001/results.json
outcome: All 768 truncation cases and 3072 inequalities passed on both interval backends at both precisions; no asymptotic improvement was claimed
artifacts:
- hunts/paid_shortfall/computations/check-001/manifest.json
- hunts/paid_shortfall/computations/check-001/results.json
- hunts/paid_shortfall/computations/check-001/stdout.txt
- hunts/paid_shortfall/computations/check-001/stderr.txtRegression and preservation checks
The final focused command ran sequentially with one numerical thread:
OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 VECLIB_MAXIMUM_THREADS=1 .venv/bin/python -m pytest -q -m 'not slow' -o addopts= tests/test_paid_shortfall.py tests/test_combined_weight_baseline_review.py tests/test_joint_correction_candidate.py tests/test_discovery_knownness.py tests/test_docs_numbering.py tests/test_repo_hygiene.py tests/test_huntspec.py tests/test_hunt_probe_discipline.py tests/test_doors.pyResult: 203 passed, 4 deselected in 17.05 seconds. This includes 29 cases in the new regression file. The full repository numerical suite was not rerun locally. Contribution structure and manifest hash validation passed. The complete lightweight governance command from checks.yml separately passed all 335 tests in 4.69 seconds. make_context.py --check and git diff --check also passed.
The preservation regression fixed alongside this study was reproduced before editing: the restored baseline checker sets global mpmath precision to 60 digits on import, and the joint checker sets it to 80. Either import made a negative control that assumed 15 ambient digits accept 1/7 to 40 digits. The tests now restore exact incoming bit precision after the whole checker test, including on failure. The negative control now explicitly supplies 15-digit data and is tested under ambient precision 15, 60, and 80. The original checker files, archives, and recorded outputs remain unchanged.