# Root merge 2 of 3: 25 results on porting Taylor's (gtmcknight) solved AMM policy

Policy credit throughout: Taylor (gtmcknight), github.com/gtmcknight/amm-sim @ 4812a6382311624e1506aca131ca35c378f930d9, `research/solved`.
Each figure below is tagged with the result it came from, for example [b14-trim-2]. The files are the 25 inputs b11-b3 … b20-pb1.

## 0. How to read this merge

**Missing results.** Seven inputs produced no measurement.
- b11-b3: blocked. `sh scratch/setup.sh`, line 6 (`uv venv`), exit 126 `cannot execute binary file`. The runtime was Darwin arm64. The wheel downloaded and its sha256 `aedb6828…` matched. Seeds 9601536..9601791 are unused.
- b16-qm-1: blocked. The network proxy CONNECT timed out (curl exit 28), so port-3 and the amm-challenge clone were not fetched. The machine reported Linux x86_64. No gate, bytes or gas figures exist.
- b16-qm-2: blocked at the platform preflight (Darwin arm64). Nothing was done.
- b18-tb3: blocked at the platform check (Darwin arm64). Seeds 9608704..9608959 are not run.
- b20-pb1: blocked at the platform check (Darwin arm64). Seeds 9610240..9610495 are not run.
- b13-lattice-1: setup failed on the macOS host, because a Linux wheel cannot be installed there. It gives only source-derived operation bounds and no measurements (§4).
- **Consequence:** the quality-matched "qm" port (b16) has no deliverable in this group. These blocks are runtime failures, not requests for requester input.

**Weight of the flagged sources.** The merge instructions say that 15 results had sources that did not hold on re-read. The flagged results are b11-b4, b13-lattice-2, b14-trim-1, b14-trim-2 (2 sources), b15-net2-choose, b17-ta1…ta4, b18-tb1, tb2, tb4, b19-pa2 (2 sources), pa3 and pa4.
- The machine did not tell me which URL failed in each case. The inputs also quote no passages from any URL. They only link the sources.
- Most of those links are artifact URLs (`api.imd.fun/artifacts/<id>`) and pinned raw.githubusercontent files for method and input provenance.
- The numeric results are local measurements: engine runs, pyrevm replays and replica runs. They come with per-seed CSVs, hashes and harnesses inside the results themselves.
- Treat the provenance claims ("this is the artifact with that hash", "this README gives that number") as weakly confirmed for those results.
- The measured figures themselves rest on the worker's own run and are not independently certified. Several results say so (tb4: "source links identify inputs and method, not independent certification").
- b15-agg-report-b, b12-qm-spline, b12-qm-net and b19-pa1 are not on the flagged list, but they quote no passages either.
- Where a README number is quoted below (533.33, 529.557, 532.39), it comes from the worker's reading of the pinned README. I could not re-check it.

**Measurement kinds, which are not comparable.**
- Engine edges are about 514–538, and the level depends on the seed set. Replica edges are about 529–533. Replica levels do not convert to board scores (README, per [b12-qm-spline]).
- Engine tests run each strategy alone against the 30 bps VanillaStrategy, using `amm_sim_rs.run_batch(X, VanillaStrategy, configs, 8)`.
- Paired differences compare the same seed across separate matches.
- The comparators in every engine item are t03 (sha256 `cb6d945511e1e10af78baa64c2a6cb25806f0998403ce6440654bfc3c70092d9`, `censored_10`, 14,654 runtime bytes) and IB (sha256 `301941829a44bb60977bb28181a5e4bfa6544f3696763f987054789b2611d686`, `inputs_b`, 14,528 bytes). These hashes agree in every item that reports them.
- b11-b4 states that its t03 download is the brief's t03, with "cb6d9455" matching the brief.

## 1. Headline

1. **Gates P and the engine limits have been met by the shipped candidates, and the +1.2…+1.4 gain over t03 is replicated on the engine.**
   - Five Solidity candidates were scored (port-3, port-2, trim-1, trim-2, and trim-1 again on fresh seeds).
   - Every engine run shows 0 OutOfGas and 0 reverts, with afterSwap peaks below 245,000 and runtime below 24,576 bytes.
   - The best candidate is **trim-2** (the net-1 policy repacked): 11,680 bytes and an engine peak of 130,711. It scored +1.407 vs t03 on a single 128-seed run. It has not been re-run on a 256-seed fresh set in this group.
   - **trim-1** has the most replication: seven 256-seed engine sets, plus one 128-seed run, plus an older 128-seed run. The mean over t03 is about +1.3 and its peaks run up to 238,704.
2. **Gate Q for net2** (the policy of trim-1 and trim-2) vs the full plain tables: +0.005367, se 0.016095, n=1024 on the choose seeds 60000..61023 [b15-net2-choose]. This is replica only. The report set 10000..12047 was not scored for net2 in this group.
3. **qm tables:** one compression meets the size and loss targets: a 3-D B-spline of 10,167 bytes with paired loss 0.0632, se 0.0072, n=2048 [b12-qm-spline]. A fixed-point MLP misses the loss target at 0.589 [b12-qm-net]. Neither was ported to Solidity in this group, because b16-qm-1 and b16-qm-2 are blocked.
4. **Lattice:** Taylor's lattice, as written, does not fit the gas budget. The closest reduction still exceeds it [b13-lattice-2].

## 2. Gate P (replica, unmodified `python3 -I gates.py`)

Gate P reproduced the README digit for digit in b12-qm-spline, b12-qm-net, b13-lattice-1, b13-lattice-2, b14-trim-1, b14-trim-2, b15-net2-choose and b15-agg-report-b. The others that mention it either do not repeat it or say they did not re-run replica gates (b17, b18, b19).

| 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 |

- Source: [b12-qm-spline], with Python 3.12.3 and a clean checkout. The identical block appears in [b15-net2-choose] and [b15-agg-report-b].
- b12-qm-net gives the same figures for the full qm tables (443.5226083, 571.6235507, 525.8658593) under numpy 1.26.4.
- Holdout full-table means from the README, as quoted by the workers:
  - qm: 533.33 / 533.3260 measured on 10000..12047 [b12-qm-spline].
  - Gaussian + plain: 532.39 in the README against 532.385687 measured (full arm, 2048 seeds) [b15-agg-report-b].
  - HA mean 529.557 (README, per [b12-qm-spline]), which implies about +3.71 for the spline vs the README's +3.77 for the full qm tables.

## 3. The qm tables (task: compress ~71k values; here the full qm tables are 1,073,424 values)

### 3.1 Inventory [b12-qm-spline, b12-qm-net — agree]
- `qm_k1e-5.json` is 7,892,018 bytes. The scalars are W0=160, NW=321, N2=49, C2=24, FB=6, H_bps=5.0, lam 0.8, sigma 9.45, size 20, kappa 1e-5. S = [0,4,8,12,16,20,26,34] bps.
- Five tags × 8 sd levels:
  - 0c is 8×321×13.
  - 1s, 1c, 2s and 2c are 8×321×49.
- Total 1,073,424 values, or 536,712 on the ask side.
- The bid side mirrors the ask: B[k][i][j] = A[k][320-i][last-j].
  - Both results find one exception: tag `1c`, sd level 6 (26 bps), i=160, j=24, off by 0.71 bps.
  - The error is included in every measured loss.
- The fee surface F = A − w has min exactly 0. F=0 in about 47–48% of cells per [b12-qm-spline]. [b12-qm-net] gives the fraction with F>0 as 0.523–0.531, which is consistent. The max F is 772–782 bps.

### 3.2 B-spline compression [b12-qm-spline] — meets both targets
- It is one 3-D clamped cubic tensor-product B-spline per table, over (row, column, sd-level index). Coefficients are int16 in 1/8 bps.
- The knots sit at visit quantiles from seeds 60000..60255.
- It is fitted by censored least squares: where F = 0, the target is min(prediction, 0), over 12 iterations.
- Decode applies F̂ = max(F̂, 0) per node, A = w + F̂, B = mirror(A).
- The QMS1 blob is **10,167 bytes**: 120 header, 255 descriptors and knots, and 9,792 coefficient bytes. Its sha256 is `3ac7cb8811ecfb6a17f499aa4e9b70dd55f7fb6432b3aec3b904e11a3dece38e`.
- Chosen (nr,nc,nk) per table: 0c (24,10,4); 1s (24,14,4); 1c (18,12,4); 2s (24,14,4); 2c (12,8,4).
- The setting was chosen on seeds 60000..61023 (n=1024), frozen, then scored once on 10000..12047.

| setting (choose seeds 60000..61023) | bytes | loss | se |
|---|---|---|---|
| 2-D 12,6/12,10/10,6/12,10/6,6 | 6,763 | 0.1280 | 0.0148 |
| 2-D 14,8/16,12/12,8/16,12/8,6 | 10,519 | 0.0935 | 0.0116 |
| 3-D 16,8,4/16,10,4/12,8,4/16,10,4/8,6,4 | 5,011 | 0.0924 | 0.0114 |
| 3-D 20,8,4/20,12,4/16,10,4/20,12,4/12,8,4, w^1 | 7,499 | 0.0682 | 0.0109 |
| same, w^0.5 | 7,499 | 0.0662 | 0.0106 |
| 3-D 20,8,5/20,12,5/16,10,5/20,12,5/12,8,4 | 9,107 | 0.0665 | 0.0107 |
| **3-D 24,10,4/24,14,4/18,12,4/24,14,4/12,8,4 (chosen)** | **10,167** | **0.0549** | **0.0103** |
| 3-D 28,12,4/28,16,4/20,12,4/26,14,4/12,8,4 | 12,279 | not scored | |

- **Holdout, scored once, seeds 10000..12047 (n=2048):**
  - Full-table mean edge 533.3260, decoded mean edge 533.2628.
  - **Paired loss 0.0632, se 0.0072.**
  - worst5 (seed, loss): (10522, 1.458), (10711, 1.360), (12015, 1.342), (10268, 1.172), (10203, 1.092).
  - Implied vs HA is about +3.71, from the README HA mean of 529.557. The full tables give +3.77.
- Lookup cost, as operation counts only and not measured gas:
  - 64 int16 reads per side, 100 in the worst case.
  - About 570 multiplies and 48 divisions per ask/bid pair.
  - The exp in the fee conversion is not counted.
- Caveats:
  - The fitting code is not delivered. Only `qms_decode.py` and the blob are included.
  - The reference clamps F̂ at grid nodes and interpolates bilinearly. A contract that evaluates the spline at the continuous point would differ, so Gate F needs the grid rule.
  - The visit seeds 60000..60255 overlap the choose set.
  - No Solidity was written and no engine run was made.

### 3.3 Fixed-point MLP compression [b12-qm-net] — misses the loss target
- It is an int16 5-32-32-1 ReLU net, one net for all tags, with 5 inputs: w/100, R/100, phase, sd/10 and synced. It is 3,076 bytes, sha256 `b509fbe9a9b2131087b7d2c03297f90d3bf06e36e859c3fa23927d81b51e9c25`.
- **Holdout loss +0.5891, se 0.0138, n=2048.** The target was < 0.2. worst5 (seed, loss): (10955, 3.075), (10401, 2.988), (11494, 2.855), (11778, 2.847), (11519, 2.798).
- Choose-seed loss fell with each training round, 1.34 → 0.82 → 0.60, and had not plateaued when time ran out. The result calls this a time limit, not a demonstrated capacity limit.

| setting (choose seeds 60000..61023) | bytes | training | loss | se |
|---|---|---|---|---|
| w32, path, floor 0.1 | 3,076 | 400 s | 2.529 | 0.032 |
| w48, path, floor 0.1 | 6,084 | 400 s | 2.967 | 0.036 |
| w32, none, floor 0.1 | 3,076 | 400 s | 1.339 | 0.024 |
| w32, none, floor 0.01 (a2) | 3,076 | 420+400 s | 0.820 | 0.022 |
| w32, path, floor 0.01 (c2) | 3,076 | 420+400 s | 0.844 | 0.022 |
| **w32, none, floor 0.01 (a3, chosen)** | 3,076 | 420+400+360 s | **0.601** | 0.020 |
| w32, path, floor 0.01 (c3) | 3,076 | 420+400+360 s | 0.654 | 0.020 |

- The width-48 test was run only at the weak floor and short budget, so the width comparison is not a fair test.
- Lookup cost, as scored: 16 net evaluations of 1,216 multiplies each (19,456 in total). The result calls this too much. Evaluating directly at the continuous point would take 2,432 multiplies, but that variant was not what was scored.
- The result's untried next steps: train to convergence, one net per s/c group, fine-tune on interpolated lookups.
- It could not reproduce fixed-net's plain-table loss of −0.040 here, which it cites. [b12-qm-spline] does not mention it.

### 3.4 Where the two compressions disagree
The spline is better on quality (0.0632 vs 0.589), and the net is smaller (3,076 vs 10,167 bytes). Both agree on the data facts: the shapes, the symmetry exception, and the fee-surface structure. Neither has a Solidity port or a Gate F measurement. The two b16 attempts to port qm are blocked.

## 4. Lattice filter (Taylor's `latt.py`)

- [b13-lattice-1] has **no measurement.** Setup failed on the macOS host. It supplies only these source-derived bounds, which it labels "not Solidity gas measurements":
  - 121 cells of 2.5 bps. M = ceil(4·9.45/2.5) = 16, so **33 taps** (not 11).
  - Silent-step cap 30, so up to 31 diffusion passes.
  - 121 × 33 = 3,993 contributions per pass, or 123,783 across 31 passes.
  - The base exposes only 32 storage words, so packed storage is needed.
- [b13-lattice-2] measured it (pyrevm gas; replica for edge; **no engine seed, and engine `gas_stats()` was not run**):
  - Setup exited 0. The Solidity port matches Taylor's Python lattice to about 1e-8 bps (max m 1.15e-8, max fee 9.4e-9).
  - Quotes use port-3's coarse plain tables, not qm. This is a gas and filter-fidelity probe, not the +4.26 policy.
  - **The full lattice (121×2.5 bps, cap 30) does not fit.** On 20,000 replayed callbacks (seeds 70100..70109): mean 453,463, peak 10,513,884, and 12,900 callbacks above 245,000 (64.5%).
  - Deployed runtime is 20,992 bytes.
  - Reductions (pyrevm, "exact" driver):

| variant | peak | mean | callbacks >245k | runtime bytes |
|---|---:|---:|---:|---:|
| 121×2.5, cap 2 | 2,442,141 | 427,486 | 12,900 | 20,992 |
| 121×2.5, cap 1 | 1,688,758 | 401,460 | 12,899 | 20,993 |
| 61×5, cap 30 | 3,075,267 | 190,284 | 4,695 | 20,598 |
| 61×5, cap 1 | 575,880 | 174,297 | 4,705 | 20,598 |
| 31×10, cap 30 | 1,098,264 | 110,623 | 499 | 20,403 |
| 31×10, cap 2 | 335,145 | 107,761 | 116 | 20,403 |
| **31×10, cap 1** | **258,576** | **104,959** | **15** | 20,404 |
| 31×10, cap 1, own Python twin | 268,830 | 105,008 | 10 | 20,404 |

  - The closest variant (31 cells × 10 bps, cap 1) is "near the budget but still over it". The result adds that a pyrevm peak near 245k is not safe on the engine, citing port-3's 156,888 in pyrevm against 206,161 on the engine (a figure from port-3's own report).
  - Replica edge of the reductions (paired, seeds 60000..60047, n=48; the full lattice with plain quotes has mean edge 541.3482):
    - 61×5, cap 30: −0.043 (se 0.036).
    - 31×10, cap 2: −0.086 (se 0.076).
    - 61×5, cap 1: −0.173 (se 0.072).
    - 31×10, cap 1: −0.182 (se 0.087).
    - Port-3's Gaussian with the same plain quotes: −0.615 (se 0.097).
    - Taylor's lattice with full qm quotes (reference only): **+1.754** (se 0.108).
  - Reading: most of the +4.26 headline comes from the qm tables, which add about +1.75 here. The reductions keep part of the lattice's advantage over the Gaussian under plain quotes.
  - "No measured reduction fits on its own." The unmeasured next steps are listed: a 3σ kernel cut, a skip threshold, 21 cells × 10 bps, an unrolled loop, and one σ·sqrt(n) diffusion.
  - Variant contracts, generator and harness were scratch files and are not delivered. The full contract and variant hashes are in the file.
- **Disagreement:** lattice-1 and lattice-2 do not contradict each other. Lattice-1 is blocked and lattice-2 ran. Treat lattice-2 as the only measurement. One of its sources was flagged as not holding (see §0), and its numbers come from its own pyrevm replay.

## 5. Plain Gaussian + net2 policy (net-1 → trim-1, trim-2) — the main deliverable line

### 5.1 Gate Q (replica, vs full plain tables `pol_mid.json`)
- **[b15-net2-choose]:** net2 − full = **+0.005367400**, se 0.016095032, n=1024, on seeds 60000..61023.
  - Mean edge 529.061693433 (full) vs 529.067060833 (net2). Ahead on 513 seeds, behind on 511.
  - The result's own 95% interval is about −0.026 to +0.037.
  - worst5: (60750, −2.363654), (60809, −1.978941), (60304, −1.950315), (61012, −1.532004), (60365, −1.505508).
  - best5: (60653, +5.160671), (60036, +1.897821), (60023, +1.461283), (61021, +1.461130), (60376, +1.456479).
  - Checks: `payload.bin` is 2,730 bytes with sha256 `9727ecc8…e0f6`. `netref.net_grid` matched the decoded grid on every cell (max difference 0.0). Probe pairs on seeds 60000 and 60001 (0.2058 and 0.5312) match what net-1 reported.
  - The report set 10000..12047 was **not** scored for net2. Nothing ran on the engine.
- **[b15-agg-report-b]** (port-2's aggregated silent-step filter, `PortReal`, with the full tables), agg − full, replica:
  - Seeds 11024..12047: **−0.024294**, se 0.013577, n=1024 (agg 532.770273 vs full 532.794567).
  - The same script on 10000..11023: −0.044645 (se 0.013694). Over all 2,048 seeds: −0.034469 (se 0.009642).
  - The agg arm over 2,048 seeds averages 532.351218, which equals port-2's published "Full-table mean". The full arm averages 532.385687.
  - [b15-net2-choose], supplementary: agg − full on 60000..61023 was −0.045195795 (se 0.013343417, about 3.4 se).
  - Read together, the one-truncation filter costs roughly 0.02–0.045 edge. Net2 shows no such loss on the choose set. net2 and agg were not paired directly.
- Gate Q is therefore closed for net2 on the choose set only. [b14-trim-1] and [b14-trim-2] both say they did not rerun Gate Q. [b14-trim-1] adds that net-1 reported its own 2-pass Q comparison as incomplete, and says the trim "does not close that inherited quality-evaluation gap". [b15-net2-choose] is the measurement that closes it, on the choose set.

### 5.2 Gas trims of net-1 (published net-1 engine peak: **245,377**, over the limit)
Both trims keep net-1's policy. Both pass Gate F on seeds 70000..70049 (50 runs × 2,000 callbacks).

| | trim-1 [b14-trim-1] | trim-2 [b14-trim-2] |
|---|---|---|
| deployed bytes | 16,077 | **11,680** (net-1: 16,368) |
| Gate F max \|fee diff\| (bps) | 5.053e-09 | 3.181e-09 |
| Gate F mean \|fee diff\| (bps) | 1.8948e-10 | 1.9099e-10 |
| pyrevm peak / mean | 183,373 / 95,691 | **115,724** / 58,022 |
| engine seeds | 9530000..9530127 (n=128) | 9531000..9531127 (n=128) |
| engine afterSwap peak / mean | 232,889 / 131,552 | **130,711** / 71,236 |
| OOG | 0 | 0 |
| engine vs t03 | +1.156616 (se 0.120983) | **+1.407384** (se 0.116698) |
| engine vs IB | +1.876940 (se 0.132232) | +2.235717 (se 0.124385) |

- **Method of trim-1:**
  - It changes the log range and series (|y| ≤ 0.171573), the Work struct, the CDF tail and a shared sqrt. On the single trace for seed 70012, net-1's pyrevm peak fell 197,653 → 186,425 → 185,218 → 184,602 → 183,373 (−14,280).
  - Fees are "tolerance-equivalent, not claimed bit-identical".
  - Trace-set digest `593d45df…9031`. Sha256 of the source is `d849c5f15d69499560b5c3ea08f55ad27153b393eba7abb7dd186f451ad46a24` (34,507 bytes).
- **Method of trim-2:**
  - It repacks storage (8 slots → 3) and switches to lane-packed model words: 31-bit lanes (8 per word) in layer 0 and 33-bit lanes (7 per word) in layer 1. It removes structs and memory arrays and shares layer 0 across ask and bid.
  - On the seed-70000 trace, the pyrevm peak went 176,547 → 152,799 → 151,193 → 114,900 → 106,685, and the mean 106,595 → 59,510.
  - It claims raw 64-byte return values byte-identical to net-1's on 10,005 replayed calls.
  - Trace-set digest `69a29538…faa59`. Source sha256 `6157028022f7f2151781ea3a772e4a99956ab9c38c8fbabf4995887d05564b27` (26,867 bytes), and `gen2.py` sha256 `55fc0405…ab1b`.
  - It declined the ln-term and cheaper-CDF-tail ideas because a Q14 rounding flip could move a fee by more than 0.01 bps.
- **Engine paired detail:**
  - trim-1 vs t03: worst5 (9530101, −2.647107), (9530114, −1.731659), (9530048, −1.662802), (9530123, −1.617354), (9530008, −1.015215).
  - trim-2 vs t03: 113/128 seeds ahead. worst5 (9531040, −1.714359), (9531020, −1.663399), (9531121, −1.567490), (9531000, −1.506759), (9531126, −0.633494).
  - trim-2 vs IB: 124/128 ahead.
  - Mean submission edges: trim-1 531.056096, t03 529.899480, IB 529.179156. For trim-2: 533.253356, t03 531.845972, IB 531.017638.
- **gas_stats (afterSwap):**
  - trim-1: 1,376,738 calls, 0 OOG, peak 232,889. t03 peaked at 235,394 and IB at 234,837 in the same run.
  - trim-2: 1,389,374 calls, 0 OOG, peak 130,711, mean 71,235.67. t03 peaked at 232,735 and IB at 232,015.
  - trim-2's afterInitialize peak was 90,800, against 243,234 for t03 and 243,227 for IB.
  - Per-chunk trim-2 peaks were 130,711 / 130,382 / 130,346 / 130,629.
- **Flagged-source note:** the trim-1 and trim-2 results each had source claims flagged as not holding (§0). Their numbers come from their own runs, and every engine seed was used once.

**Disagreements and tensions between the trims:**
1. The two trace-set digests differ (`593d45df…` vs `69a29538…`) for the same seeds 70000..70049, and each result claims to hash the same reference-driven callbacks. [b14-trim-2] also says its digest is identical to net-1's. Neither result explains the mismatch, so treat the digest as inconsistent. The two files agree on the max fee diff being far below 0.01 bps.
2. trim-1 changes rounding and claims only tolerance equivalence, while trim-2 claims byte-identical returns to net-1. The two trims therefore need not produce identical fees to each other. Neither compared the two directly.
3. Engine peak margin: trim-1 232,889 (about 12k under 245,000) against trim-2 130,711.
4. The engine gains over t03 differ by seed set: +1.157 (trim-1) and +1.407 (trim-2). net-1 measured +1.228 (se 0.103) on seeds 9522000..9522127, according to [b14-trim-2]. The result attributes the difference to sampling.

### 5.3 Engine confirmations of trim-1 (hash `d849c5f1…6a24`, `TaylorGaussianPlainNet2`, 16,077 runtime bytes)
All use n=256 fresh seeds, each arm run once per seed in chunks of 32 under `timeout 470`.

| item | seeds | − t03 (mean_diff, se) | − IB (mean_diff, se) | afterSwap peak | mean gas | OOG |
|---|---|---|---|---:|---:|---|
| [b17-ta1] | 9607168..9607423 | +1.316756, 0.090810 | +1.985776, 0.098844 | 234,349 | 131,461.6 | 0 |
| [b17-ta2] | 9607424..9607679 | +1.434873, 0.086026 | +2.188874, 0.091592 | **238,704** | 131,551.8 | 0 |
| [b17-ta3] | 9607680..9607935 | +1.344864, 0.082327 | +2.071386, 0.092233 | 237,831 | 131,528.9 | 0 |
| [b17-ta4] | 9607936..9608191 | +1.352338, 0.084982 | +2.153189, 0.094493 | 231,793 | 131,506.6 | 0 |
| [b18-tb1] | 9608192..9608447 | +1.245321, 0.088490 | +1.987566, 0.091690 | 229,310 | 131,512.4 | 0 |
| [b18-tb2] | 9608448..9608703 | +1.180238, 0.086126 | +1.847209, 0.086314 | 238,657 | 131,562.9 | 0 |
| [b18-tb3] | 9608704..9608959 | blocked, not run | | | | |
| [b18-tb4] | 9608960..9609215 | +1.407142, 0.082704 | +2.150001, 0.094389 | 233,269 | 131,512.2 | 0 |

- Unweighted mean of the seven measured 256-seed sets: about +1.33 vs t03 and about +2.06 vs IB. This is my arithmetic, not any result's own figure.
- Each file's candidate mean edge (for reference): 537.040126, 523.920483, 529.891035, 529.908043, 526.040572, 524.422378, 529.499600 (ta1…tb4). t03 means are given in each file.
- Worst-5 paired losses against t03 are all between about −1.2 and −2.7 (ta1 −1.83, ta2 −1.89, ta3 −1.53, ta4 −2.46, tb1 −1.87, tb2 −2.65, tb4 −1.88 at worst). No seed shows a loss as big as the one in §6.
- Sizes: the candidate afterSwap peak is below 245,000 in every set, but 238,704 and 238,657 exceed the t03 peak (235,370) and the IB peak in the same sets. Trim-1's mean gas (about 131.5k) is lower than t03's (about 135.5–135.7k) and IB's (about 135.1–135.4k).
- All seven result files say replica gates P/Q/F were not re-run, and that the observed peaks apply only to those seeds, not as a universal bound.
- Flagged-source note: ta1, ta2, ta3, ta4, tb1, tb2 and tb4 each had one source flagged (see §0). The figures come from the worker's own engine runs.
- Config check: each file builds the config for seed i from `default_rng(i)` (size, rate, sigma in that order), with simulation seed i. b11-b4 verified this against `MatchRunner._build_configs` for non-seed fields.
- No-op: tb3 (seeds 9608704..9608959) and pb1 (9610240..9610495) are not covered, so those 512 seeds are still unspent by any measurement here.

## 6. Port-3 and port-2 on the engine

### 6.1 Port-3 (`TaylorGaussianPlainCoarse`, sha256 `62930b8e68277c69fc1acfba18de781e684d7fc6daac8dc562b8210dc3932ac1`, 18,466 bytes) [b11-b4]
- Seeds 9601792..9602047 (n=256), one run per strategy and seed.
- Mean edge: port 537.269064, t03 536.622128, IB 535.852112.
- **Port-3 − t03: +0.646936, se 0.087598** (wins/ties/losses 167/0/89). worst5: (9601808, −2.311534), (9602020, −2.060259), (9601877, −2.034938), (9602011, −2.031919), (9601894, −1.971185).
- **Port-3 − IB: +1.416953, se 0.097643** (207/0/49). worst5: (9601808, −3.207422), (9601904, −2.543943), (9601946, −1.893936), (9601877, −1.473379), (9601797, −1.461482).
- Gas (afterSwap): port-3 had 2,785,441 calls, 0 OOG, peak 207,453, mean 95,334.74. t03 peaked at 235,406 (mean 135,337.79). IB peaked at 234,825 (mean 134,999.59).
- [b13-lattice-2] quotes port-3's own report: engine peak 206,161 (pyrevm peak 156,888, mean 61,334). This is a different seed set from the 207,453 above and the two are consistent in size.
- The worker could not open the three artifact URLs in its browsing tool. Direct HTTPS downloads worked and supplied the exact sources, whose hashes are listed. That is one reason for the flag on this result's source (§0).
- No port-3 vs t03 head-to-head pool match was run.

### 6.2 Port-2 (`TaylorGaussianPlain2`, sha256 `7e2d9c157a7725c2875eea31e5741938fb9eafef5a3442710573f1b1a87faf17`, 18,484 bytes, artifact `e8c967d6…286c`)
All use n=256 and report 0 OOG and 0 reverts.

| item | seeds | − t03 (mean_diff, se) | − IB (mean_diff, se) | cand peak | cand mean gas |
|---|---|---|---|---:|---:|
| [b19-pa1] | 9609216..9609471 | +1.449878, 0.088973 | +2.278666, 0.117865 | 227,141 | 126,419.14 |
| [b19-pa2] | 9609472..9609727 | +1.162093, 0.089855 | +1.951497, 0.094621 | 227,141 | 126,989.21 |
| [b19-pa3] | 9609728..9609983 | +1.059791, 0.091698 | +1.714796, 0.115926 | 226,563 | 127,070.63 |
| [b19-pa4] | 9609984..9610239 | +1.416359, 0.090178 | +2.182227, 0.094719 | 226,071 | 126,441.59 |

- t03 peaks in the same runs: 237,588 (pa1), 235,259 (pa2), 237,441 (pa3), 235,382 (pa4). The IB peaks were 234,837 / 234,802 / 234,946 / 234,805.
- pa1 details: the IB outlier seed 9609431 (IB scored 427.891894550 against 448.674664118 for the candidate) drives the +20.78 best5 value. pa1's section numbering runs 1, 2, 3, 5, 4. It reports rounded 6-decimal figures.
- **pa3 outlier:** seed 9609765 is the largest loss in both pairs: −7.600373 against t03 and −17.631353 against IB. Edges were candidate 796.558098, t03 804.158470, IB 814.189450. The config was retail_mean_size 20.180, arrival rate 0.9008, gbm_sigma 0.000889. The worker wrote "I did not investigate the cause." No other result in this group has a loss near this size.
- pa3 notes that port-2's earlier published +1.4143 (n=64, seeds 9510064..9510127) used its own author's seeds, which were also used in debugging. On the 256 blind seeds, the figure is +1.06 ± 0.09.
- The pa1 and pa4 means are 537.974742 and 531.411330 for the candidate.
- **Method inconsistency in pa4:** it builds size, rate and sigma from `default_rng(0..255)` via `MatchRunner(...)._build_configs()[start:start+32]` and only overwrites the simulation seed to 9609984+i. The other 256-seed engine items draw size, rate and sigma from `default_rng(seed)`. So pa4's market parameters repeat the default MatchRunner configs 0–255 and its "fresh seed" claim is weaker than the others'. The brief asks for fresh seeds from 9500000 up, so this matters. The pa4 gap to t03 (+1.416) is nevertheless in the range of the others.
- **Other process notes:**
  - pa1: the first attempt of one t03 chunk (9609440) was killed by an 8-minute command timeout before producing a result. It was rerun, and "no result of it was seen".
  - pa2: an initial invocation could not find `scratch/setup.sh` and executed nothing. No chunk was rerun.
  - pa3: every command took under 30 s.
- Flagged-source note: pa2 had 2 sources flagged, and pa3 and pa4 had 1 each (§0). pa1 was not flagged.
- Port-2 uses the aggregated silent-step filter. [b15-agg-report-b] measures that filter's replica cost (−0.024 to −0.045), which is the filter this port uses on the full tables (with its own compressed tables in the contract).

## 7. Cross-candidate comparison, with caveats
| candidate | engine peak (range) | runtime bytes | mean gas | engine − t03 (range over sets) |
|---|---|---:|---:|---|
| port-3 | 207,453 | 18,466 | ~95.3k | +0.647 (1 set of 256) |
| port-2 | 226,071–227,141 | 18,484 | ~126.4–127.1k | +1.06…+1.45 (4 sets of 256) |
| trim-1 | 229,310–238,704 (and 232,889 at n=128) | 16,077 | ~131.5k | +1.18…+1.43 (7 sets of 256; +1.157 at n=128) |
| trim-2 | 130,711 | 11,680 | ~71.2k | +1.407 (1 set of 128) |
| t03 | 232,735–237,611 | 14,654 | ~135.5k | — |

- No result compares trim-1, trim-2, port-2 and port-3 on the same seeds. The seed-to-seed swing of about ±0.2 is larger than the standard errors (about 0.09), so the ranking between port-2 and trim-1 is not supported by this group. All of them beat t03 and IB in every set that was run.
- All run configs differ by seed, so means of edges differ by set (about 514 to 538). Only paired differences are comparable.
- The brief's goal figures (+2.83 Gaussian + plain, +3.77 qm, +4.26 lattice + qm, t03 about +1.45) are on Taylor's replica against HA. They are not confirmed or disputed by any engine result here. The replica-only quality of the plain port is +0.0054 (net2) and −0.024 to −0.045 (agg) against Taylor's plain tables.

## 8. What the group did not deliver
- No Solidity port of the qm tables, so no engine score, gas or Gate F for qm (b16-qm-1 and b16-qm-2 blocked; b12 only compresses).
- No lattice contract inside the gas budget (b13-lattice-2). No engine run of the lattice probes.
- No engine data for b11-b3, b18-tb3 and b20-pb1 (blocked).
- No net2 scoring on the report seeds 10000..12047, and no engine confirmation of trim-2 on a 256-seed set.
- No Gate Q measurement for trim-2's exact repack beyond the claim of identical returns to net-1.
- Compressed 71,000-value table for the plain policy: the net2 payload is 2,730 bytes (sha256 `9727ecc8…e0f6`), which is the fixed-net payload. It is used by net-1, trim-1, trim-2 and the net2 arm of [b15-net2-choose]. This merge has no result that builds a new plain-table compression.
- The header credit to Taylor (gtmcknight) is present in the Solidity of port-2 (b19-pa3 quotes the header line `// Gaussian plain-table AMM policy ported from Taylor (gtmcknight), amm-sim`) and port-3 (b11-b4 says the attribution is preserved). The other results state the credit in their text.

## 9. Index of results
| id | what | outcome |
|---|---|---|
| b11-b3 | port-3 confirmation | blocked, not run |
| b11-b4 | port-3 on engine 9601792..9602047 | +0.647 vs t03, +1.417 vs IB |
| b12-qm-spline | qm B-spline | 10,167 B, loss 0.0632 (se 0.0072) |
| b12-qm-net | qm MLP | 3,076 B, loss 0.589 (se 0.014), target missed |
| b13-lattice-1 | lattice gas probe | blocked at setup, bounds only |
| b13-lattice-2 | lattice gas probe | does not fit; closest 258,576 pyrevm peak |
| b14-trim-1 | net-1 gas trim | engine peak 232,889, +1.157 vs t03 |
| b14-trim-2 | net-1 gas trim | engine peak 130,711, +1.407 vs t03, 11,680 B |
| b15-net2-choose | Gate Q net2 | +0.005367 (se 0.016095) |
| b15-agg-report-b | agg vs full | −0.024294 (se 0.013577) |
| b16-qm-1 | qm port | blocked (network) |
| b16-qm-2 | qm port | blocked (platform) |
| b17-ta1…ta4 | trim-1 engine confirmation | +1.32, +1.43, +1.34, +1.35 vs t03 |
| b18-tb1, tb2, tb4 | trim-1 engine confirmation | +1.25, +1.18, +1.41 vs t03 |
| b18-tb3 | trim-1 engine confirmation | blocked (platform) |
| b19-pa1…pa4 | port-2 engine confirmation | +1.45, +1.16, +1.06, +1.42 vs t03 |
| b20-pb1 | port-2 confirmation | blocked (platform) |

Not checked by this merge: I read the 25 results through three reading helpers and directly for the five short ones. I did not re-run any measurement. The per-seed CSVs, the verbatim Solidity sources and the generator code inside the results were not reproduced here.