Methodology and evidence

publication revision 20schema groupmixer-benchmark-publication-v1

Execution contract

Native benchmark resource classes
Class	Deadline	CPU	Attempts
heavy	300s	8 pinned logical CPUs	native randomized: seeds 101, 202, 303, 404, 505 · seedless hosted: five fresh attempts · deterministic: one
standard	60s	8 pinned logical CPUs	native randomized: seeds 101, 202, 303, 404, 505 · seedless hosted: five fresh attempts · deterministic: one

Input preparation, solving, and output conversion count toward the deadline.
Refreshed native and browser attempts use fresh processes, an 8 GiB physical memory cap, and no swap. Earlier native rows used an 8 GiB address-space ceiling.
Returned schedules are independently validated without repair.
Elapsed times are not compared across hosted, browser, or different-machine native runs.

Status vocabulary

No candidate, timeout, error, and infeasibility are separate outcomes.

Benchmark publication status definitions
Optimal	A zero lower bound or declared target was attained.
Optimal, proven	The solver returned a matching bound and verified proof evidence.
Feasible	At least one canonically valid schedule was returned.
Infeasible, proven	The exact hard-feasibility problem was proven infeasible.
No candidate timeout	The declared search envelope ended without a candidate.
Error	The process or evidence pipeline failed; no timeout or infeasibility is inferred.
Unsupported	The tool cannot represent the case or exceeds an observed product limit.
Unavailable	A runnable licensed environment was not available.
Invalid output	Output was returned but failed canonical validation.
Pending verification	The workflow or capability has not entered publication evidence.
Not run	No execution was admitted for this row.

Configuration rules

Allowed

Case-specific formulations and symmetry.
Disclosed runtime construction and staged search.
The strongest frozen end-user setting admitted within the case resource envelope.

Forbidden

Embedded schedules, answers, or cross-solver hints.
Proxy cases, weakened semantics, or repaired output.
Selection by a metric outside the declared objective.

Authorship disclosure

GroupMixer authors this benchmark and the OR-Tools formulations. Measured 11 October 2026: all OR-Tools cases and 12 GroupMixer browser cases ran on an Intel Core Ultra 9 386H. Other GroupMixer rows retain the earlier native Auto Extreme runs. Machines, runtimes and memory policies differ; elapsed times are not compared across those runs. Returned schedules are independently evaluated against the same canonical rules. This is not an independent third-party benchmark.

Charter Manifest OR-Tools formulations

Publication files

SHA256SUMS

Publication evidence files with sizes and SHA-256 checksums
File	Size	SHA-256
publication.json	393.7 KiB	sha256:3783b760ba6a1bfb3d5771d1d705f053585d1941acdb8bf2816a72696b46b2fc
attempt-aggregate.json	552.5 KiB	sha256:da8f744048f800400cfecae3f503d6df6a5388005b665fb859787e579c96ea83
publication.schema.json	9.1 KiB	sha256:0b13828818f97dad7050a8dad70ce195fadd7bd25e453f61fc2a525a70754ad0
evidence.tar.gz	9.5 MiB	sha256:73c7262e123906bc595d3560db5534bb17bd6df6c8be2be09ba52f7c320be757
publication-provenance.json	151.6 KiB	sha256:595eea5a3f640c9efb0ce60777ebf5cc694d714c000ea8170cdfb11597ed9592

Reproduce and verify

cd webapp && npm run benchmark:data:check
jq '.summaries[]' public/benchmarks/public-group-generator-v1/revision-20/attempt-aggregate.json
cd public/benchmarks/public-group-generator-v1/revision-20 && sha256sum -c SHA256SUMS

Comparator evidence revisions: r1 (151 rows) · r2 (11 rows) · r19 (20 rows) · r20 (78 rows). Current summaries exclude dry and superseded attempts; the archive retains historical evidence for provenance.