# e6-freeze-combo / freeze — result

Credit: Taylor (gtmcknight) for the replica, Gaussian filter, fee rule and qm tables that ComboG builds on.

Input D = https://api.imdnotdead.fun/artifacts/9b842d35c17245af4dd2bfc0ec608263f9b7edcb2db619cffba7d36d1d853740 (gd-ab-1 report). Base: trunk 4eb3e98, `contracts/Candidate.sol` (TC-A).

## Files changed (all inside the assigned paths)

| File | Action | Bytes | SHA-256 |
| --- | --- | ---: | --- |
| `contracts/Combo.sol` | new: TC-A + D's diff | 47,171 | `f904a486b1dcfbd6fcf40aa48920d1bb6b79ad2ac1c99a2cbbb16d4272c2c616` (matches the brief) |
| `contracts/combo.diff` | new: D's "Frozen source: unified diff against TC-A" fence, verbatim (lines between the ```diff fences) | 10,251 | `ee1e263770b2beecda27dbf2b85365c6536dd1e6f4c69f3ed5ad6451d39290b2` |
| `reference/combo/combog.py` | new: D's "Exact combog.py" fence, verbatim, `GATE = 1` | 6,011 | `321c61dff43c1677722d4b52e066ed5296f08ac7b72465245f04eb04e1c143cb` (matches the brief) |
| `reference/combo/GATES.md` | new: P/Q/F/gas-probe evidence, every figure linked to and attributed to D | 5,175 | `7dce3475…5065` |
| `README.md` | edited: TC-A marked as confirmed fallback; new section "Combo (ComboG(1), frozen, NOT yet confirmed)" with hashes, prerequisites, rebuild/replay/compile/F/engine commands and startup imports; seed table gains Combo A 9803572..9804595, B 9804596..9805619, TC-A gap 9805620..9805635, next free 9805636 (9803188..9803571 listed as reserved outside this item); confirmation pending, requester submits, no board score | 8,952 | `80980dcb…3b07` |

Unchanged and re-verified: `contracts/Candidate.sol` `b87234c8…a2` (TC-A); `reference/combo/twocomp.py` `9d3c1833…0af` (byte-identical to `reference/twocomp/twocomp.py`); `reference/qt1/qmref.py` `1dfc9b7a…b92`; `reference/qt1/blob.bin` `3ac7cb88…e38e`; `reference/qt1/Strategy.sol` `c68887cf…ec08aca`. No config, lib, tools or other files were touched. The exact header (`// SPDX-License-Identifier: MIT` … credit line) and the final newline (`}\n`) of Combo.sol come from the patch unchanged.

## Checks run (all locally, this machine)

1. **Patch reconstruction, two ways.** `git apply -p1 --verbose combo.diff` on a copy of TC-A in a temp dir: "Applied patch Candidate.sol cleanly", which produced `f904a486…c2c616`, 47,171 bytes. I also wrote a strict Python unified-diff applier that requires every context and removed line to match exactly (`test/scratch/apply.py`). It gave the same SHA-256. (`patch` is not installed here, so `git apply` was used.) I re-ran the README rebuild block (with `git apply`) and got the same hash.
2. **Hashes and byte counts** of all deliverables and the unchanged inputs (`sha256sum`, `wc -c`): the table above.
3. **Python AST:** `ast.parse` of `combog.py` and `twocomp.py` succeeded.
4. **Replica replay (stdlib, plain `python3 -I`, about 0.7 s per seed),** using the README's startup imports (`reference/combo`, `reference/taylor` on `sys.path`): ComboG edges 444.3066107552986 / 571.8761229207034 / 526.1507459263532 on seeds 10000/10001/10002, and QMRef 443.1017411146794 on 10000. All equal D's per-seed table (12 dp). `ComboG.GATE == 1`.
5. **Import resolution:** `tools/check.py` adds `reference/{taylor,qt1,twocomp,mix}` but not `reference/combo`. The README F command therefore sets `PYTHONPATH=$TASK_ROOT/reference/combo` and does not edit any reference or tool. The `twocomp` it picks up is byte-identical (`cmp`).

## Evidence recorded (attributed to D, not re-run here)

Q choose n=1024 +0.6370844005364295 se 0.019226759996279156; report n=2048 +0.6028942685272641 se 0.013049022593029348; F max 3.0968576748141574e-6 bps on 100,000 callbacks, pyrevm peak 203,056; probe mixed dt=1 peak 203,136; runtime 22,058 bytes; 0 OOG on those samples, and no universal gas bound is claimed. D's engine diagnostic on 9800000..9800007 reused earlier reservations and is marked **EXCLUDED** from fresh confirmation.

## Explicitly NOT run (left for separate workers)

- Solidity compile / `amm-match validate` of `Combo.sol`, and its runtime size. No engine setup was done in this item ("no engine").
- Gate F replay (`tools/check.py … combog.ComboG 70000 70050`) and the gas probe.
- Gate P and the full Q paired runs (only the 3-seed spot check above).
- Engine confirmation on Combo A/B and the TC-A gap. These ranges are reserved in README and not consumed.

Dependencies: replica needs only python3 stdlib. Compile, F and engine need the pinned tools setup (uv, solc 0.8.24, amm_sim_rs wheel, pyrevm 0.3.7) per `tools/README.md`. Scratch state was only in `test/scratch/` and `$TMPDIR`.