Prediction-market variants: five supplied results

This is the requested write-up, with strategy code omitted. It covers four one-spread variants and the supplied two-spread combination s14. It does not deliver five entry files or establish five one-spread entries. The assignment names these five results and expressly restricts this output to a report; inventing a replacement fifth experiment would not be supported by the inputs.

The table is a self-contained transcription of the measurement fields in the supplied item JSONs. Each row's source points to its provenance record below, not to a file that might disappear from the handoff. All edge measurements are variant minus T, not absolute earnings. mean ± SE is a mean and its standard error, not a confidence interval. Worst is the minimum paired difference, with its seed. “Richest 50” means the mean paired difference on the 50 runs ranked richest by T, not the variant. Seconds show T then variant, not a computed overhead. Equality is the item's reported exact equality of every result field on the tested untouched runs; it is not an exhaustive test of all seeds.

Item / source	Exact ON line	Untouched-spread equality	Changed-spread measurements	Seconds: T → variant
s4-dc	ON = dict(cancel=(2, 3, 4), cut=(1,), postjump=(1,), premium=(), deadband=(3, 4))	spread1: 100 runs, true; spread2: 100 runs, true; spread3: 100 runs, true	spread4: runs 500; mean 1.4200869211069989; SE 0.5107506860654305; retail 1.7708466736051995; arbitrage -0.35075975251139974; first half 1.0685887407683996; second half 1.771585101445598; richest-50 mean 4.600859521091996; worst seed 2200466, difference -35.004646379099995	161.74336862564087 → 171.03542351722717
s4-c	ON = dict(cancel=(2, 3, 4), cut=(1,), postjump=(1,), premium=(), deadband=(3,))	spread1: 100 runs, true; spread2: 100 runs, true; spread3: 100 runs, true	spread4: runs 500; mean 0.10368416204399977; SE 0.4102594749228982; retail -0.5366064790973997; arbitrage 0.6402906411357999; first half 0.4789823494736003; second half -0.27161402538560075; richest-50 mean 0.4387771822139985; worst seed 2200009, difference -31.456282262299965	154.9151418209076 → 148.39432978630066
s4-d	ON = dict(cancel=(2, 3, 4), cut=(1, 4), postjump=(1,), premium=(), deadband=(3, 4))	spread1: 100 runs, true; spread2: 100 runs, true; spread3: 100 runs, true	spread4: runs 500; mean 0.9981752193062001; SE 0.5771814079018966; retail 2.2236039819422; arbitrage -1.2254287626489997; first half 0.16905083531000012; second half 1.8272996033024003; richest-50 mean 4.968733236584003; worst seed 2200820, difference -175.92424872140003	209.56698005087674 → 205.67333119222894
s1-d	ON = dict(cancel=(2, 3, 4), cut=(1, 4), postjump=(1,), premium=(), deadband=(1, 3))	spread2: 100 runs, true; spread3: 100 runs, true; spread4: 100 runs, true	spread1: runs 1000; mean -0.6169039386181995; SE 1.0245981660820762; retail 0.5400407193614997; arbitrage -1.1569446579905005; first half 0.46620377442200045; second half -1.7000116516583994; richest-50 mean 0.39882605011800254; worst seed 2202068, difference -923.1194120767	59.803158670663834 → 52.450689886230975
s14	ON = dict(cancel=(2, 3, 4), cut=(1,), postjump=(1,), premium=(), deadband=(1, 3, 4))	spread2: 100 runs, true; spread3: 100 runs, true	spread1: runs 1000; mean -0.6169039386181995; SE 1.0245981660820762; retail 0.5400407193614997; arbitrage -1.1569446579905005; first half 0.46620377442200045; second half -1.7000116516583994; richest-50 mean 0.39882605011800254; worst seed 2202068, difference -923.1194120767spread4: runs 1000; mean 0.7365028589834997; SE 0.3593500760511514; retail 1.6063239172819999; arbitrage -0.8698210583067998; first half 1.4200869211069989; second half 0.05291879686000031; richest-50 mean 6.993010559773995; worst seed 2203506, difference -119.07101432770001	45.05994053604081 → 62.47025144100189

s4-dc and s4-d gained on both tested halves. s4-c gained in the first half and lost in the second. s1-d also gained in the first half and lost in the second; its full-sample mean is negative. s14's spread-4 part gained in both halves, while its spread-1 part gained then lost. These are signs of the supplied half means, not claims of statistical significance. Sources: the corresponding rows and provenance records.

s14's spread-1 measurement object is exactly equal to s1-d's, including all reported statistics and best/worst lists, on the reported same runs. For spread 4 the samples differ: s4-dc uses the first 500 matching seeds, while s14 uses the first 1000. s14's first-half mean, 1.4200869211069989, exactly matches s4-dc's whole-sample mean on their shared first 500 runs. Their full-sample means, standard errors, components, richest-50 means and worst runs should not be equated across different sample sizes. The supplied aggregate JSONs contain no raw paired-run series, so per-run equality between those two experiments cannot be independently checked here. However, direct code comparison confirms that s14 and s4-dc have identical switch memberships for spread 4, and s14 and s1-d have identical memberships for spread 1. Sources: s14, s4-dc, s1-d, and the code verification below.

The item diff fields are estimated score changes obtained by dividing a single spread's mean and SE by four; s14 combines its two spread means and combines their SEs in quadrature before dividing by four. They are fresh-run estimates, not measured changes on the board's fixed 200 runs, and the division assumes equal spread weights rather than the board's actual undisclosed composition. Exact supplied mean ± SE values:

• s4-dc: 0.3550217302767497 ± 0.12768767151635763.
• s4-c: 0.025921040510999943 ± 0.10256486873072455.
• s4-d: 0.24954380482655003 ± 0.14429535197547416.
• s1-d: -0.15422598465454987 ± 0.25614954152051905.
• s14: 0.029899730091325044 ± 0.2714468040768672.

The strongest positive mean among the isolated supplied tests is s4-dc, with both halves positive; this is a candidate for manual consideration, not a demonstrated board improvement. s4-c's second half loses. s1-d's negative overall mean and worst run make its tail exposure material; the same exposure is present in s14. s4-d also has a large negative worst run despite positive halves. Positive retail contributions do not remove arbitrage losses: all cases except s4-c show negative mean arbitrage contributions in the table. The small s14 spread-4 second-half gain gives weaker support than its first half. Sources: the table's exact item fields.

Entry limitations and code verification. This report supplies ON lines, not uploadable strategies. The complete code exists in the input items' code fields, but those input files are not promised as part of this report-only handoff. No entry should be represented as available from this report alone. The high-severity review correctly identifies that absence; its demand for five additional code files conflicts with this task's explicit “Leave the code out” and “Write nothing else” requirements. No credentials or manual board action are required for this write-up.

I fetched T directly during this revision and verified its code SHA-256 as 9fba1eb5659d5b0ac9b7273a968a520ef9a95551c697be22020993005af70192. Each supplied code hash matches its item's code_sha256. Replacing its sole ON line with T's line yields exactly T's bytes for every item. Thus there are no other code changes or added seed-recognition logic relative to T. This is static verification; I did not rerun the simulator or reproduce the board score.

T's ON line is ON = dict(cancel=(2, 3, 4), cut=(1, 4), postjump=(1,), premium=(), deadband=(3,)). Relative to it, s4-dc removes cut and adds deadband in spread 4; s4-c removes cut in spread 4 while retaining cancel; s4-d adds deadband in spread 4; s1-d adds deadband in spread 1. s14 removes cut in spread 4 and adds deadband in spreads 1 and 4: the union of changed memberships is {1, 4}. It is a combination entry, unsuitable if the requirement is that every entry isolate exactly one spread. This reproduces the advisory finding without replacing an assigned result with an unmeasured alternative. Sources: fetched T and each item's ON line above.

Evidence limits. The task flags one failed external source validation for each of s4-dc, s4-d, s1-d and s14; it does not identify which URL or passage failed. The supplied JSONs contain no source URL/quote records identifying those failures. Their tried statements, simulation statistics and equality checks therefore remain contributor-reported measurements, not externally authenticated or independently reproduced results. Directly checking T repairs confidence in the baseline and one-line-change claims only; it does not validate their claimed runs. s4-c is not flagged, but its measurements likewise have not been independently rerun here. The pinned simulator is the task's reference implementation, not a source for these private measurements. No raw per-seed outputs, full sampled configurations, or execution logs were supplied. In particular, the claimed untouched equality cannot be audited run by run from the aggregates.

The review's missing-measurement-file finding describes a real handoff risk even though all five inputs exist on this revision's machine. This report preserves their requested measurement fields and provenance below rather than relying on absent relative links. Hashes identify the exact inputs; they cannot recover missing raw runs or authenticate contributor execution claims. No absent measurements have been invented.

Provenance and supplied method statements. Each record below identifies the original input (available while preparing this revision), its task-specified SHA-256 independently checked here, and its verified code hash. The table above is the report's embedded measurement excerpt from that source. The following method details are attributed to each item's tried and next fields.

Source s4-dc. Input artifacts/goal/items/s4-dc/result.json, also supplied as .imd/reads/artifacts/item-s4-dc. Input SHA-256: 33eb9771ca81b1698131b13be4029e9eba2c356ec85040b8b51762c173786369. Code SHA-256: d88d2c72d2478ed7b9fc702432c2227e06eeef8d3bf980be511b63428b4d6337. Reports first 100 matching seeds per untouched spread starting at 2100000; first 500 spread-4 seeds starting at 2200000. Its halves divide that ordered sample. Timing uses seeds 1–200. Earlier-file calibration: 600 runs from seed 1000, mean 108.63974289615767; reported 7m54s led to halving the changed-spread sample.

Source s4-c. Input artifacts/goal/items/s4-c/result.json, also supplied as .imd/reads/artifacts/item-s4-c. Input SHA-256: e14ada102ac780bac6f8fd0fb5dfed6fbb01def4b3248aa1b81a6965828f50af. Code SHA-256: 8fc15ca7148427ccb5f49ec5cf7ab2a8d34e0c594306969524792aaa0a448d34. Reports first 100 matching seeds per untouched spread starting at 2100000; first 500 spread-4 seeds starting at 2200000. Timing uses seeds 0–199. Earlier-file calibration: 600 runs from seed 1000, mean 108.63974289615767; reported 7m53s led to halving the changed-spread sample.

Source s4-d. Input artifacts/goal/items/s4-d/result.json, also supplied as .imd/reads/artifacts/item-s4-d. Input SHA-256: 8c65399227ed9922614f91e653e67751f681c4630ecbc63febcaecacfd6310aa. Code SHA-256: 2231b337bb0ab3a01d8909be0630541f8cb59fb81d98bee806c34dd51cecb27b. Reports first 100 matching seeds per untouched spread starting at 2100000; first 500 spread-4 seeds starting at 2200000, in seed order. SE is sample standard deviation of paired total-edge differences divided by sqrt(500); halves contain 250 pairs each. Timing uses seeds 0–199, excludes strategy loading, and reports zero failures. Earlier-file calibration: 600 runs, seeds 1000–1599, mean 108.63974289615767, elapsed 371.84397927019745 seconds; sample halved.

Source s1-d. Input artifacts/goal/items/s1-d/result.json, also supplied as .imd/reads/artifacts/item-s1-d. Input SHA-256: e91fc9154813ad184803b43191501aa317b591fc3447ff8e3955b8067446efbe. Code SHA-256: a4907178f3ed12f1825290beb5316d1cbe81de2d960a188a8f96482923b36e56. Reports first 100 matching seeds per untouched spread starting at 2100000; first 1000 spread-1 seeds starting at 2200000 through runner._run_single_simulation, with no failures. Timing uses seeds 0–199. Earlier-file calibration: 600 runs from seed 1000, mean 108.63974289615767, took 33 s; sample not halved. Reports T mean 34.6864449842615 on the changed-spread runs and mean 0.3065219954539545 when worst seed 2202068 is omitted; that omission is not the full-sample estimate.

Source s14. Input artifacts/goal/items/s14/result.json, also supplied as .imd/reads/artifacts/item-s14. Input SHA-256: 5cc906284cff6ecee62dcaca2d781a6d27a4bf40127c4f5f73c4083254e0c8f2. Code SHA-256: 12f0e24c0b95f170e1f8ec595f58f23af6c2580972bdc31c4650cc887267e461. Reports first 100 matching seeds per untouched spread starting at 2100000; first 1000 seeds of each changed spread starting at 2200000 through runner._run_single_simulation, with no failures. Timing uses seeds 0–199. Earlier-file calibration: 600 runs from seed 1000, mean 108.63974289615767, took about 33 s; sample not halved. Its timing statement explicitly notes shared-machine noise.

Timings were sequential T-then-variant within each item, on different contributor machines and sometimes different seed ranges; they do not establish a controlled cross-variant speed ranking. Source: the five method statements and seconds fields above.