# Merge 01 — per-spread flips of T's `ON` line: cancel s1–s4, cut s1 (seeds 1,000,000–1,001,999)

Sign: d = CANDIDATE.total_edge − T.total_edge per seed (positive = flip helps). All figures below are computed from the 20 item results in `.imd/reads/artifacts/item-*` (each listed in the brief with its SHA-256); nothing here is quoted from outside sources. Pooling: sum_d and sum_d2 added across the four 500-seed blocks (equal weight per seed); mean = Σd/n; SE = sample SD/√n with SD from (Σd² − n·mean²)/(n−1).

T: SHA-256 `9fba1eb5…af70192` (all 20 items report this hash and say it matched). Simulator: danrobinson/prediction-market-challenge @ `b244530ee2a6247b5f930830eeca9dd950da0e7e` (all items say unmodified).

## Result per flip (pooled over 2,000 seeds)

| Flip | Spread | n on spread | Σd (spread) | Σd² (spread) | Gain on spread (mean d) | SE | t | Overall mean d (n=2000) | Overall SE | Worst single run (seed, d) | Best single run (seed, d) | Runtime ratio C/T* |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| cancel-s1-on | 1 | 533 | -1614.5168 | 306497.7850 | **-3.0291** | 1.0313 | -2.94 | -0.8073 | 0.2763 | 1001629, -291.3374 | 1000998, 307.5245 | 0.9927 |
| cancel-s2-off | 2 | 482 | -233.4912 | 8446.8891 | **-0.4844** | 0.1896 | -2.56 | -0.1167 | 0.0459 | 1000194, -45.4122 | 1001915, 50.4834 | 0.9869 |
| cancel-s3-off | 3 | 538 | -510.8911 | 5038.2211 | **-0.9496** | 0.1255 | -7.56 | -0.2554 | 0.0350 | 1001181, -20.2586 | 1001794, 13.2587 | 1.0009 |
| cancel-s4-off | 4 | 447 | -18.6585 | 2568.0799 | **-0.0417** | 0.1135 | -0.37 | -0.0093 | 0.0253 | 1000501, -26.3426 | 1001235, 10.1852 | 1.0016 |
| cut-s1-off | 1 | 533 | 542.1676 | 2069142.0258 | **+1.0172** | 2.7010 | +0.38 | +0.2711 | 0.7194 | 1000631, -313.4653 | 1001661, 1377.7205 | 0.9980 |

*Runtime ratio = mean of the four blocks' candidate seconds/simulation ÷ mean of T's. Items measured timing very differently (some under 4–12-way CPU contention, some including strategy loading; per-block T times range 0.24 s to 10.6 s), so only the within-block ratio is meaningful; every block's ratio is within about ±1.6% of 1.0 (see block table). No flip has a material runtime cost.

## Reading

- **No flip in this group shows a reliable gain.** Four of the five have negative pooled means on their spread: cancel-s1-on (t −2.94), cancel-s2-off (t −2.56) and cancel-s3-off (t −7.56) clearly hurt; cancel-s4-off is indistinguishable from zero (t −0.37). cut-s1-off has a positive point estimate (+1.02) but t = +0.38. Three of its four blocks are negative, and the positive pool comes from one seed (1001661, d = +1377.72, in block 4). Without that seed the spread-1 sum would be about −835.6. Recommendation: keep T's current setting for all five entries.
- `cancel-s1-on`: -3.0291 ± 1.0313 on spread 1 (t = -2.94, n = 533); overall -0.8073 ± 0.2763. Worst run -291.3374 (seed 1001629).
- `cancel-s2-off`: -0.4844 ± 0.1896 on spread 2 (t = -2.56, n = 482); overall -0.1167 ± 0.0459. Worst run -45.4122 (seed 1000194).
- `cancel-s3-off`: -0.9496 ± 0.1255 on spread 3 (t = -7.56, n = 538); overall -0.2554 ± 0.0350. Worst run -20.2586 (seed 1001181).
- `cancel-s4-off`: -0.0417 ± 0.1135 on spread 4 (t = -0.37, n = 447); overall -0.0093 ± 0.0253. Worst run -26.3426 (seed 1000501).
- `cut-s1-off`: +1.0172 ± 2.7010 on spread 1 (t = +0.38, n = 533); overall +0.2711 ± 0.7194. Worst run -313.4653 (seed 1000631).
- Distributions are heavy-tailed (single seeds of −291 and +307 in cancel-s1-on), so SEs computed from sums of squares are approximate; for cancel-s1/s2/s3 the negative sign holds in every block; cancel-s4-off is split by block (two negative, two positive); cut-s1-off's sign hinges on one outlier.

## Per-block detail (on the flip's spread)

| Item | Seeds | n | Σd | Σd² | mean d | SE | overall mean d | sec/sim T / C | flagged sources |
|---|---|---|---|---|---|---|---|---|---|
| cancel-s1-on-1 | 1000000–1000499 | 144 | -399.4694 | 36694.9658 | -2.7741 | 1.3146 | -0.7989 | 0.2361 / 0.2350 | no |
| cancel-s1-on-2 | 1000500–1000999 | 120 | -60.3318 | 127720.6783 | -0.5028 | 2.9903 | -0.1207 | 0.2703 / 0.2696 | yes (1) |
| cancel-s1-on-3 | 1001000–1001499 | 129 | -327.5777 | 24260.1603 | -2.5394 | 1.1912 | -0.6552 | 0.7919 / 0.7832 | yes (1) |
| cancel-s1-on-4 | 1001500–1001999 | 140 | -827.1379 | 117821.9805 | -5.9081 | 2.4090 | -1.6543 | 0.5933 / 0.5899 | no |
| cancel-s2-off-1 | 1000000–1000499 | 118 | -102.4619 | 2548.9460 | -0.8683 | 0.4221 | -0.2049 | 0.8094 / 0.8111 | yes (1) |
| cancel-s2-off-2 | 1000500–1000999 | 122 | -51.9241 | 1286.4353 | -0.4256 | 0.2927 | -0.1038 | 6.7250 / 6.6170 | no |
| cancel-s2-off-3 | 1001000–1001499 | 109 | -40.8573 | 268.1775 | -0.3748 | 0.1466 | -0.0817 | 0.2480 / 0.2487 | no |
| cancel-s2-off-4 | 1001500–1001999 | 133 | -38.2479 | 4343.3303 | -0.2876 | 0.4968 | -0.0765 | 1.0523 / 1.0419 | yes (1) |
| cancel-s3-off-1 | 1000000–1000499 | 128 | -153.0256 | 1724.9445 | -1.1955 | 0.3080 | -0.3061 | 1.3509 / 1.3608 | yes (1) |
| cancel-s3-off-2 | 1000500–1000999 | 140 | -125.9121 | 1054.6967 | -0.8994 | 0.2200 | -0.2518 | 10.1640 / 10.1640 | no |
| cancel-s3-off-3 | 1001000–1001499 | 142 | -182.8286 | 1477.9543 | -1.2875 | 0.2491 | -0.3657 | 0.5939 / 0.5941 | yes (2) |
| cancel-s3-off-4 | 1001500–1001999 | 128 | -49.1249 | 780.6256 | -0.3838 | 0.2165 | -0.0982 | 0.3616 / 0.3627 | yes (1) |
| cancel-s4-off-1 | 1000000–1000499 | 110 | -34.1208 | 500.0689 | -0.3102 | 0.2021 | -0.0682 | 1.4266 / 1.4282 | no |
| cancel-s4-off-2 | 1000500–1000999 | 118 | -42.3004 | 1351.9572 | -0.3585 | 0.3112 | -0.0846 | 0.8431 / 0.8437 | yes (1) |
| cancel-s4-off-3 | 1001000–1001499 | 120 | 17.8554 | 437.4947 | 0.1488 | 0.1745 | 0.0357 | 1.2033 / 1.2046 | no |
| cancel-s4-off-4 | 1001500–1001999 | 99 | 39.9072 | 278.5590 | 0.4031 | 0.1645 | 0.0798 | 0.7265 / 0.7298 | yes (1) |
| cut-s1-off-1 | 1000000–1000499 | 144 | -39.2270 | 8174.8519 | -0.2724 | 0.6297 | -0.0785 | 0.8034 / 0.8017 | no |
| cut-s1-off-2 | 1000500–1000999 | 120 | -410.0888 | 109255.3992 | -3.4174 | 2.7482 | -0.8202 | 10.5800 / 10.5580 | no |
| cut-s1-off-3 | 1001000–1001499 | 129 | -134.5867 | 25750.5693 | -1.0433 | 1.2454 | -0.2692 | 0.4112 / 0.4110 | yes (1) |
| cut-s1-off-4 | 1001500–1001999 | 140 | 1126.0702 | 1925961.2055 | 8.0434 | 9.9250 | 2.2521 | 0.2476 / 0.2471 | no |

## Consistency checks done in this merge

- All 20 items: complete = true, blocked = null, no failures, 500 seeds each, contiguous blocks covering 1,000,000–1,001,999 for every flip.
- Candidate hashes agree across the four blocks of each flip: cancel-s1-on `d54b2518…`, cancel-s2-off `1bf151a1…`, cancel-s3-off `e37b254a…`, cancel-s4-off `f20f0f8f…`, cut-s1-off `a23040e3…`. This merge did not recompute these hashes (T source not re-fetched); they are as reported.
- Re-summed every item's per_seed list and compared to its reported sum_d; checked d = 0 on every seed outside the flip's spread; checked each seed's spread label agrees across all 20 items (the same seed must map to the same sampled config).
  Result: no discrepancies found.
- Per-spread seed counts are identical across flips (spread 1: 533, 2: 482, 3: 538, 4: 447 over the 2,000 seeds) — consistent with every item sampling the same configs.

## Flagged sources

The brief flags re-read source failures for: cancel-s1-on-2 (1), cancel-s1-on-3 (1), cancel-s2-off-1 (1), cancel-s2-off-4 (1), cancel-s3-off-1 (1), cancel-s3-off-3 (2), cancel-s3-off-4 (1), cancel-s4-off-2 (1), cancel-s4-off-4 (1), cut-s1-off-3 (1). The result files handed to this merge contain no sources list, so which sources failed could not be identified here. The cited sources would plausibly be the T artifact or the simulator repository; the numbers themselves are local simulation measurements, and each flagged item's per-seed data passed the internal checks above and is consistent with the unflagged blocks of the same flip (e.g. cancel-s3-off-3 and the other cancel-s3-off blocks all show negative spread-3 means). Their hash claims rest on the item's own report. Excluding all flagged blocks would not change any recommendation: the unflagged blocks of cancel-s1/s2/s3 are all negative on their spread; cancel-s4-off and cut-s1-off stay inconclusive either way.

## Missing

Nothing missing: all 20 named items were present and non-empty.