# Frozen ComboG(1) audit

## Outcome

The source reconstruction and Python-only checks passed on this host. The pinned AMM EVM harness could not be set up: this machine is Darwin arm64, while the supplied setup downloads a Linux x86_64 manylinux wheel. `uv pip install` stopped with `A path dependency is incompatible with the current platform: scratch/amm_sim_rs-0.1.0.post1-cp310-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl` (exit 2). Thus EVM compilation/runtime, Gate F, and the requested complete GATE=1 gas probe are **not independently verified here**. This is the blocker: `{step: setup, error: incompatible Linux x86_64 wheel on Darwin arm64}`.

The frozen unified diff from the [provided artifact](https://api.imd.fun/artifacts/9b842d35c17245af4dd2bfc0ec608263f9b7edcb2db619cffba7d36d1d853740) applied cleanly to pinned TC-A. Candidate source SHA-256 is `f904a486b1dcfbd6fcf40aa48920d1bb6b79ad2ac1c99a2cbbb16d4272c2c616`; this matches the frozen GATE=1 SHA in that artifact. Recovered `combog.py` SHA-256 is `321c61dff43c1677722d4b52e066ed5296f08ac7b72465245f04eb04e1c143cb`; unchanged `twocomp.py` is `9d3c1833321493fb51bbde4e1b4dabc2f8978bd6d774ee2af8dc8aed13f0f0af`; `blob.bin` is `3ac7cb8811ecfb6a17f499aa4e9b70dd55f7fb6432b3aec3b904e11a3dece38e`. The Python and data hashes match the frozen artifact's stated hashes.

## Reproduced checks

**Gate P.** Ran `python3 -I gates.py` from the pinned Taylor `research/solved` checkout at `4812a6382311624e1506aca131ca35c378f930d9`; it exited 0:

```text
seed   perfect-info      gaussian+plain    gaussian+qm       lattice+qm
10000  448.261189  443.455598  443.522608  443.663520
10001  576.779199  571.463409  571.623551  571.977293
10002  531.886497  525.168028  525.865859  526.411631
```


**Replica edge spot checks (not Q).** Ran the frozen ComboG reference against `QMRef` via `python3 -I scratch/paired.py 10000 10003 scratch/spot.json`. All three completed. Local values agree with the brief's edge checks within floating-point roundoff:

| Seed | ComboG GATE=1 edge | qt1 reference edge | Difference |
|---:|---:|---:|---:|
| 10000 | 444.306610755384 | 443.101741114667 | 1.204869640716 |
| 10001 | 571.876122920694 | 571.454483381928 | 0.421639538766 |
| 10002 | 526.150745926313 | 525.908441161236 | 0.242304765077 |

These are only the requested edge spot checks; they are not the paired Q sample on 60000..61023 or report sample on 10000..12047.

## EVM evidence available in the supplied artifact (not reproduced here)

The supplied artifact reports frozen runtime **22,058 bytes** against the 24,576-byte limit, with solc 0.8.24, optimizer 200, viaIR and Paris. It reports `amm-match validate` PASS and Gate F over 50 seeds / 100,000 callbacks: zero OOG, max error `3.0968576748141574e-06` bps at seed 70003 callback 655, and peak pyrevm gas 203,056 at seed 70042. These are attributed to the supplied artifact, not this run. Its per-seed report says 2,000 callbacks and zero OOG for every seed:

| Seed | Reported callbacks | Reported max error (bps) | Reported peak pyrevm gas | OOG |
|---:|---:|---:|---:|---:|
| 70000 | 2,000 | 1.23008846165e-06 | 194695 | 0 |
| 70001 | 2,000 | 1.25828750631e-06 | 201133 | 0 |
| 70002 | 2,000 | 1.6273774342e-06 | 201320 | 0 |
| 70003 | 2,000 | 3.09685767481e-06 | 200655 | 0 |
| 70004 | 2,000 | 1.56127663315e-06 | 193997 | 0 |
| 70005 | 2,000 | 1.29077812458e-06 | 201168 | 0 |
| 70006 | 2,000 | 1.60222788596e-06 | 194128 | 0 |
| 70007 | 2,000 | 1.4619280439e-06 | 192611 | 0 |
| 70008 | 2,000 | 1.2764931473e-06 | 200790 | 0 |
| 70009 | 2,000 | 1.11413981773e-06 | 202539 | 0 |
| 70010 | 2,000 | 1.34741997291e-06 | 200717 | 0 |
| 70011 | 2,000 | 1.31453592803e-06 | 200718 | 0 |
| 70012 | 2,000 | 1.24099796411e-06 | 200591 | 0 |
| 70013 | 2,000 | 1.10497465344e-06 | 200674 | 0 |
| 70014 | 2,000 | 1.87480423764e-06 | 200588 | 0 |
| 70015 | 2,000 | 1.21012839289e-06 | 199853 | 0 |
| 70016 | 2,000 | 1.94863760544e-06 | 193277 | 0 |
| 70017 | 2,000 | 1.04452080665e-06 | 200186 | 0 |
| 70018 | 2,000 | 1.66559169235e-06 | 190807 | 0 |
| 70019 | 2,000 | 1.14521333389e-06 | 200575 | 0 |
| 70020 | 2,000 | 1.89813478303e-06 | 200345 | 0 |
| 70021 | 2,000 | 2.1809881505e-06 | 200802 | 0 |
| 70022 | 2,000 | 1.78298439164e-06 | 200727 | 0 |
| 70023 | 2,000 | 1.81240802748e-06 | 200365 | 0 |
| 70024 | 2,000 | 1.51269344707e-06 | 193261 | 0 |
| 70025 | 2,000 | 1.11522641816e-06 | 200639 | 0 |
| 70026 | 2,000 | 1.58860453655e-06 | 200864 | 0 |
| 70027 | 2,000 | 1.11846965717e-06 | 200528 | 0 |
| 70028 | 2,000 | 1.18346943356e-06 | 200792 | 0 |
| 70029 | 2,000 | 1.19425649925e-06 | 195790 | 0 |
| 70030 | 2,000 | 1.80252771043e-06 | 200609 | 0 |
| 70031 | 2,000 | 9.76771415487e-07 | 199926 | 0 |
| 70032 | 2,000 | 2.27471210271e-06 | 200261 | 0 |
| 70033 | 2,000 | 2.19270757801e-06 | 198646 | 0 |
| 70034 | 2,000 | 2.98307495025e-06 | 194046 | 0 |
| 70035 | 2,000 | 9.93276936395e-07 | 192887 | 0 |
| 70036 | 2,000 | 3.06218470286e-06 | 200789 | 0 |
| 70037 | 2,000 | 1.16345209508e-06 | 194974 | 0 |
| 70038 | 2,000 | 1.24105318036e-06 | 200489 | 0 |
| 70039 | 2,000 | 1.18035959479e-06 | 194609 | 0 |
| 70040 | 2,000 | 2.11444552405e-06 | 200206 | 0 |
| 70041 | 2,000 | 2.06683749976e-06 | 200476 | 0 |
| 70042 | 2,000 | 1.04834699081e-06 | 203056 | 0 |
| 70043 | 2,000 | 1.11270493687e-06 | 200483 | 0 |
| 70044 | 2,000 | 9.35813640121e-07 | 200306 | 0 |
| 70045 | 2,000 | 1.61646567225e-06 | 199232 | 0 |
| 70046 | 2,000 | 1.45164429088e-06 | 200305 | 0 |
| 70047 | 2,000 | 1.02650954577e-06 | 200631 | 0 |
| 70048 | 2,000 | 1.40039641341e-06 | 194707 | 0 |
| 70049 | 2,000 | 9.08948021017e-07 | 200873 | 0 |

The artifact's detailed bucket table is explicitly for the **ungated** policy, so it is not evidence of the requested full frozen GATE=1 trajectory probe. It reports the following ungated observations and gate-selection rationale; no GATE=1 per-bucket counts/peaks or counterexample scan were independently produced in this audit:

| dt bucket | Ungated mixed callbacks | Ungated peak gas | Peak seed | Peak callback | OOG |
|---|---:|---:|---:|---:|---:|
| 1 | 21,500 | 203,136 | 70004 | 8342 | 0 |
| 2 | 4,103 | 208,069 | 70004 | 2905 | 0 |
| 3 | 1,657 | 207,200 | 70003 | 2587 | 0 |
| 4 | 715 | 206,025 | 70004 | 2365 | 0 |
| 5–6 | 509 | 218,148 | 70003 | 975 | 0 |
| 7–9 | 149 | 225,296 | 70000 | 983 | 0 |
| 10–12 | 25 | 237,385 | 70002 | 2486 | 0 |
| ≥13 | 8 | 235,607 | 70004 | 8392 | 0 |

The same report gives ungated non-mixed count 23,135, peak 166,518, and zero OOG. It says GATE=1 was selected because its dt=1 peak and non-mixed peak fit the 205,000 pyrevm target, whereas dt=2 at 208,069 does not. This selection claim is from the supplied artifact; the frozen complete trajectory and gas-budget result remain unverified on this host.

## Reproduction record

- Cloned Taylor `amm-sim` at `4812a6382311624e1506aca131ca35c378f930d9`, K at `4eb3e985fd31e3cd17832501b95b6d2491aca73d`, and `amm-challenge` at `6b6921f8d463314aa7c96234e1be25a25245f704` under `scratch/`.
- Applied the supplied “Frozen source: unified diff against TC-A” section with `patch -p1`; the candidate hash matched.
- Extracted the exact `combog.py` section and retained the pinned `twocomp.py`, qt1 `qmref.py`, and `blob.bin`; hashes above.
- Ran Gate P and the three `python3 -I` edge spot checks. The spot script imports only these scratch references and Taylor replica files.
- Ran the specified setup once from the challenge checkout. Setup downloaded and hash-verified the specified wheel, created a CPython 3.12 venv, then stopped at wheel platform incompatibility. No fallback or engine run was made.

Credit: Taylor (gtmcknight) for the original solved filter, fee policy, qm tables and replica.