# Merge 3: one-entry flips of T's `ON` line — postjump-s3, postjump-s4, premium-s1, premium-s2, premium-s3

Each flip is CANDIDATE minus T on seeds 1,000,000–1,001,999, in four blocks of 500 (items `*-on-1` … `*-on-4`). Source: the 20 item results in `.imd/reads/artifacts/item-<flip>-on-<k>`. d = CANDIDATE.total_edge − T.total_edge, so positive d helps. Simulator: danrobinson/prediction-market-challenge at b244530ee2a6247b5f930830eeca9dd950da0e7e. T hash 9fba1eb5…0192 was verified by every item.

## Result

**No flip in this group helps.**
- `postjump-s3-on` and `postjump-s4-on` both lose edge on the spread they change. Their standard errors are small enough that the loss is clearly not noise.
- The three `premium` flips (`premium-s1-on`, `premium-s2-on`, `premium-s3-on`) change nothing. Every paired d is exactly 0.0.

Recommendation from this group: keep T's `postjump=(1,)` and `premium=()`. Add nothing from these five flips.

## Pooled figures

Pooling is equal weight per seed: sums of `sum_d` and `sum_d2` over the four blocks. The standard error is sqrt((Σd² − (Σd)²/n)/(n−1)/n). "Spread n" is the number of seeds on the spread the flip changes.

| Flip (line change) | Seeds | Spread changed | Spread n | Gain on spread (mean d) | Σd on spread | SE | Worst single run (seed, spread, d) | Best single run | Overall mean d (n=2000) | Overall SE | Runtime ratio cand/T |
|---|---|---|---|---|---|---|---|---|---|---|---|
| postjump-s3-on: postjump=(1,3) | 2000 | 3 | 538 | **−0.14533** | −78.187 | 0.0603 | 1000052, 3, −21.641 | 1000994, 3, +13.671 | −0.03909 | 0.01628 | 1.002 |
| postjump-s4-on: postjump=(1,4) | 2000 | 4 | 447 | **−0.20956** | −93.674 | 0.1084 | 1000841, 4, −35.687 | 1000874, 4, +17.563 | −0.04684 | 0.02428 | 1.001 |
| premium-s1-on: premium=(1,) | 2000 | 1 | 533 | 0.0 | 0.0 | 0.0 | 0.0 (all ties) | 0.0 | 0.0 | 0.0 | 1.000 |
| premium-s2-on: premium=(2,) | 2000 | 2 | 482 | 0.0 | 0.0 | 0.0 | 0.0 (all ties) | 0.0 | 0.0 | 0.0 | 0.999 |
| premium-s3-on: premium=(3,) | **1707** | 3 | 464 | 0.0 | 0.0 | 0.0 | 0.0 (all ties) | 0.0 | 0.0 | 0.0 | 1.003 |

I derived the line changes from the flip names (add that spread to the tuple). The `on_line` text itself was not in the item files; only the flip name and candidate hash were.

Runtime ratio is the n-weighted mean of the reported per-simulation seconds, candidate over T. All ratios are within ±0.3% of 1, so none of these flips changes runtime. Timings were taken on different machines and under different load, so absolute seconds are not comparable between blocks. The ratio is meaningful only within a block.

## Per-block breakdown

Spread-specific Σd and per-block mean on the changed spread:

| Flip | Block 1 (1000000–499) | Block 2 (1000500–999) | Block 3 (1001000–499) | Block 4 (1001500–999) |
|---|---|---|---|---|
| postjump-s3-on | −32.310 (n=128, mean −0.252) | −4.672 (n=140, −0.033) | −26.552 (n=142, −0.187) | −14.653 (n=128, −0.114) |
| postjump-s4-on | −39.518 (n=110, −0.359) | −21.332 (n=118, −0.181) | −9.448 (n=120, −0.079) | −23.376 (n=99, −0.236) |

All four blocks of both postjump flips are negative, so the sign is consistent across independent seed blocks.

Behavior on seeds:
- `postjump-s3-on`: the results contain 467 non-zero paired differences, 260 of them positive. All are on spread 3. The mean loss comes from a heavy left tail, mainly seed 1000052 at −21.6, together with many small gains. With 260 of the 467 non-zero d positive (about 56%), the losses are not simply driven by one seed. Even so, one seed accounts for about 28% of the pooled spread-3 loss, so the estimate depends on the tail.
- `postjump-s4-on`: 378 non-zero differences, 200 positive. The worst seed (1000841, −35.7) is about 38% of the pooled spread-4 loss, and the best (1000874, +17.6) offsets part of it. Without these two seeds the loss would be smaller but still negative. I did not recompute that figure.
- Spreads other than the flipped one show d = 0.0 exactly in every block for the postjump flips, as expected.

## The premium flips are structural no-ops

Eleven items state this from reading T's code (premium-s1-on-1/2/3/4, premium-s2-on-2/3/4, premium-s3-on-1/2/3/4; premium-s2-on-1 gives no reason but measured zeros). They report that `ON['premium']` is read in exactly one place, line 409, inside the spread-4 branch of `qask`:

`if 4 not in ON['premium'] and P['pL'] <= phat <= 0.95: b *= 1.4`

That test checks for the literal 4, not the current spread. Adding 1, 2 or 3 to the tuple therefore cannot change behavior on any spread, and the measured result agrees: all d are bit-identical zeros. So these three zeros are not a measured "no effect" of premium on spreads 1–3. The flips are inert in this file.

The same reading implies that only a `premium` entry containing 4 could change behavior (it removes the ×1.4 on spread 4). That flip (premium-s4-on) is not in this group, so this merge does not measure it.

## Caveats and disagreements

- **`premium-s3-on-1` is incomplete.** It finished only 207 of 500 seeds (`complete: false`), so `premium-s3-on` pools 1,707 seeds rather than 2,000. Its note says the finished seeds are the first ~26 of each of 8 contiguous 63-seed blocks, not a random subset, and the remaining seeds are unmeasured. Its timing (about 49 s/simulation, inflated by 8 processes on a 1-core quota) is mostly irrelevant to the ratio but is included in the weighted mean. The n-weighted ratio is therefore dominated by the other three blocks. Because the code reading shows the flip cannot act, the missing seeds do not change the conclusion, but the 2,000-seed coverage the goal asked for was not achieved for this flip.
- **Seeds per spread differ across flips.** The seed set is the same for all flips, but spread n differs (e.g. spread 3 has 538 for postjump-s3 and 464 for premium-s3) because premium-s3-on is missing 293 seeds.
- **Source reliability.** Fifteen of the 20 results were flagged because one or two cited sources did not hold on re-reading (a page that could not be read, or that does not say what it was quoted for): postjump-s3-on-1/2/3, postjump-s4-on-1/2/3, premium-s1-on-1/2/3/4, premium-s2-on-2/3/4, premium-s3-on-2/3. I could not see the `sources.json` of any item; only the result files were handed to me. The numbers in this report come from the locally measured per-seed outputs, which the items describe as run against the unmodified simulator, not from quoted pages. The cited sources are the T artifact, the simulator repository and API provenance. What a failed citation could affect is the claim about *what* was run, not the arithmetic. So the flagged sources weaken the provenance claims (T's text and simulator commit identity), and I checked what I could: all 20 items report matching T hash 9fba1eb5… and the same candidate hash within each flip, with no failures. The code-reading claims about line 409 rest on the items' own quotes of T and are consistent across those premium items, and with the exact-zero measured differences. I did not independently fetch T or the simulator.
- **Candidate hashes (as reported by items, all four blocks of a flip agree):** postjump-s3-on 85db38e3…3b44; postjump-s4-on 9d1c8aa3…5406; premium-s1-on b27b1047…e9db; premium-s2-on b06acad9…5880; premium-s3-on 2a707134…f19e.
- **Runtime environment differs between blocks** (Python 3.10.12, 3.12.3, 3.13.5, 3.14.4; serial vs. split processes). Edge values are deterministic per seed. `premium-s2-on-2` notes it ran "on a 2-core machine", and `postjump-s4-on-2` ran 4 concurrent processes. Repeat runs of three seeds in fresh processes in `postjump-s3-on-3` gave identical total_edge, supporting determinism. No block reported failures (`premium-s2-on-2` and `premium-s3-on-4` give `failures: 0` rather than `[]`; same meaning).
- Per-seed d values are rounded to 6 decimals in the items; sums and SEs use the unrounded sums the items reported (I pooled `sum_d` and `sum_d2` as given). Worst and best come from each item's own unrounded lists.

## Bottom line

Neither postjump flip helps (−0.145 ± 0.060 on spread 3, −0.210 ± 0.108 on spread 4, per seed on that spread), and the three premium flips are inert. To combine winning flips into one file, nothing from this group qualifies.