# Benchmark reports

Generated by `bench/report.ps1` from commit `bb784ce` (2026-09-30).

- Toolchain: `moon 0.1.20260920 (914d7da 2026-09-20) ~\.moon-latest\bin\moon.exe | moonc v0.10.14+7d59c7ec9 (2026-09-18) ~\.moon-latest\bin\moonc.exe | moonrun 0.1.20260920 (914d7da 2026-09-20) ~\.moon-latest\bin\moonrun.exe |  | Feature flags enabled: rr_moon_mod,rr_moon_pkg`
- Instances: MIPLIB 2017 (<https://miplib.zib.de>), fetched by `bench/fetch-instances.ps1`;
  the data itself is not redistributed (see `bench/README.md`).
- Every report below is written by its own script, and every one of those scripts refuses to
  write unless its run finished, covered every manifest entry, and passed its own checks:
  no certificate refused, every reported point an integer point of the model. This file is
  the index of them, and it is the only place that says which of them the current code
  still backs.

## Reports

| report | scope | result | generated at | code behind it |
| --- | --- | --- | --- | --- |
| [parse-report.md](parse-report.md) | model file parsing (MPS/LP reader) | instances attempted: 32, parsed: 32, failed: 0 | `54fec0c` | current |
| [solve-report.md](solve-report.md) | continuous kernel (LP relaxation, presolve, certificates) | instances in manifest: 32, solved to optimality: 20, skipped: 11, other outcomes: 1 | `bb784ce` | current |
| [mip-report.md](mip-report.md) | branch and bound, 300-node budget over the whole manifest | instances in manifest: 32, proven optimal: 3, stopped at the node budget: 12, skipped: 17, relaxations verified by `verify`: 3770, cuts added and verified by `verify_cut`: 325 | `bb784ce` | current |
| [mip-report-small.md](mip-report-small.md) | branch and bound, 30000-node budget over the small-instance list | instances in manifest: 10, proven optimal: 5, stopped at the node budget: 5, relaxations verified by `verify`: 189486, cuts added and verified by `verify_cut`: 205 | `bb784ce` | current |

`current` means no file under the report's own source paths changed between the commit it
was generated from and HEAD, so the numbers still describe the code in the tree. `STALE`
means they describe an older one: the run happened and the report is its record, but
re-running the command would not have to reproduce it. Regenerating a stale report is part
of the round that made it stale, and this table is how that is noticed.

The staleness test is deliberately conservative: it compares the *files* a report's numbers
depend on, so a changed comment counts as a change even though it cannot move a number.
That direction is the safe one - a report is flagged when it might no longer describe the
current code, never cleared when it might not - and the only thing that settles a flagged
report is re-running its command (or measuring that the numbers it would print are
unchanged, which is worth writing down where it was measured).

## What each report is evidence for

- `parse-report.md`: the claim that the reader takes the whole manifest: every instance parses, none fails
- `solve-report.md`: the kernel's answer quality on real models: optima found, size refusals, numerical failures
- `mip-report.md`: that no run ends on a refused certificate, and that every reported point is an integer point of the model
- `mip-report-small.md`: the M5 criterion that small integer instances reach their published optimum; cross-checked by bench/check-mip-objectives.ps1, which compares each proven objective with MIPLIB's own table and requires equality

## How to reproduce

Run these in order in the repository root. The first one downloads the instances and
writes `bench/data/instances/manifest.txt`; the rest read them and take minutes to hours,
because they solve the whole manifest.

```powershell
powershell -NoProfile -ExecutionPolicy Bypass -File bench/fetch-instances.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File bench/report-parse.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File bench/report-solve.ps1 -Relax -MaxRows 1000 -MaxIterations 20000 -Presolve
powershell -NoProfile -ExecutionPolicy Bypass -File bench/check-relaxation-bounds.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File bench/report-mip.ps1 -MaxRows 300 -MaxNodes 300 -MaxIterations 5000
powershell -NoProfile -ExecutionPolicy Bypass -File bench/report-mip.ps1 -Manifest bench/data/instances/small.txt -MaxRows 300 -MaxNodes 30000 -MaxIterations 20000 -Output bench/mip-report-small.md
powershell -NoProfile -ExecutionPolicy Bypass -File bench/check-mip-objectives.ps1
powershell -NoProfile -ExecutionPolicy Bypass -File bench/check-mip-objectives.ps1 -Report bench/mip-report-small.md
powershell -NoProfile -ExecutionPolicy Bypass -File bench/report.ps1
```