Per-spread prediction-market pilot

The local evidence supports a modest additive gain, conditional on preserving each branch when merging; it does not yet establish a gain on Paradigm’s unseen simulations. This report transcribes the five supplied results, rather than rerunning their experiments. [S1–S4]

Sources: [S1], [S2], [S3], [S4] and [T] below identify the supplied JSON results; field names make each figure traceable. All edges are signed: a less negative arb_edge means less arbitrage loss. Numbers in the tables preserve the items’ numeric values without rounding. Code size is computed as the UTF-8 byte length of each item’s code string, not the JSON file size. No strategy code is reproduced.

The four spread items report baseline reproduction on seeds 1000–1599, tuning scores over 4000–7999, and frozen-policy holdout scores over 50000–53999. Each reports zero failures in the final score and holdout batches and exact equality with the base outside its target spread. [S1–S4: approach, score, holdout, tried] The table item instead reports 600 runs over 4000–4599 and 600 over 50000–50599, citing runtime limits; its requested full 4000–7999 score is missing, and its baseline and spread breakdown must not be compared as if they used the larger holdout. [T: approach, score, holdout]

The task flags one external source cited by spread-4 as failing a later source check, without identifying the failed URL or passage. The supplied S4 JSON has no source list that resolves this. Its simulator measurements are therefore attributed as item-reported results, not independently verified measurements; its external-source claims receive no additional evidentiary weight here. This limits confidence in the spread-4 contribution and hence the proposed aggregate. No missing measurements or source verification are fabricated.

Item / route	Reproduced base (base_mean_edge)	Base on its unseen runs (base_holdout_mean_edge)	Own seeds 4000–7999: mean; retail; arbitrage	Own unseen: mean; retail; arbitrage	Code bytes
Spread 1: ordinary-quote margin and boundary probability floor [S1]	109.27254289017199	98.28063388702982	98.9689746680761; 124.29013580906907; -25.321161141007252	98.52619137752235; 125.95607618418198; -27.4298848066588	14835
Spread 2: coarser belief grid [S2]	109.27254289017199	98.28063388702982	98.83942460660653; 124.34260207298652; -25.503177466397474	98.30682016589887; 126.04095539629678; -27.73413523039735	14777
Spread 3: guarded demand estimate for tick-one stock [S3]	109.27254289017199	98.28063388702982	98.84950498965385; 124.4013873695152; -25.551882379876773	98.47028967306635; 126.1276183775221; -27.657328704453427	16586
Spread 4: immediate-cover horizon and queue threshold [S4]	109.27254289017199	98.28063388702982	98.84752384692902; 124.52530289979967; -25.6777790528859	98.40769532192527; 126.21554314516477; -27.807847823237275	14974
Offline table: per-spread arbitrage-loss weight [T]	109.272542890172	97.06959828018033	Full range missing; 4000–4599 only: 100.62054626478633; 127.584287384187; -26.9637411193865	97.03279922204433; 123.6284762471255; -26.59567702508067	14753

The following table copies by_spread: each item cell is mean; difference ± paired standard error, relative to that item’s matching base. S1–S4 share the base column and run counts. T has its own base and counts because its unseen sample is smaller. Zero differences outside the target spread are reported equality, not a fitted estimate. [S1–S4, T: by_spread]

Spread	S1–S4 base (runs)	S1	S2	S3	S4	T base (runs)	T
1	35.099921878704336 (1016)	36.06668365229695; 0.9667617735926181 ± 0.24996913452108194	35.099921878704336; 0.0 ± 0.0	35.099921878704336; 0.0 ± 0.0	35.099921878704336; 0.0 ± 0.0	43.66184288153885 (157)	43.521209538343946; -0.14063334319490464 ± 0.07301492695071997
2	83.51205221971222 (998)	83.51205221971222; 0.0 ± 0.0	83.61700724523948; 0.10495502552725416 ± 0.11415515237099481	83.51205221971222; 0.0 ± 0.0	83.51205221971222; 0.0 ± 0.0	76.95587352704295 (156)	76.95587352704295; 0.0 ± 0.0
3	122.97345583671554 (1017)	122.97345583671554; 0.0 ± 0.0	122.97345583671554; 0.0 ± 0.0	123.71939796468614; 0.7459421279705998 ± 0.2441846966634406	122.97345583671554; 0.0 ± 0.0	141.71566301263334 (144)	141.71566301263334; 0.0 ± 0.0
4	153.82041508580312 (969)	153.82041508580312; 0.0 ± 0.0	153.82041508580312; 0.0 ± 0.0	153.82041508580312; 0.0 ± 0.0	154.34492049300826; 0.5245054072051594 ± 0.1582480941681672	132.69005518649442 (143)	132.69005518649442; 0.0 ± 0.0

The gains add algebraically if the merged policy reproduces each frozen target branch: sum each target diff × runs / 4000 from the second table. This gives a derived gain of approximately 0.58846099029355 per run and a combined local mean of approximately 98.86909487732337, against the reported base 98.28063388702982; neither is a measured merged-policy score. Spread 2’s difference is smaller than its standard error, while the other target differences exceed twice theirs. The original constants-copying plan is insufficient: spread 3 needs its guarded flow estimator and spread 4 its price-dependent queue rule. No merged run or aggregate paired standard error is supplied, and this arithmetic is not a forecast of leaderboard improvement. [S1–S4: by_spread, approach, next; aggregate calculated here]

The table route argues that fair-value fill scoring makes inventory value flat when cash constraints do not bind, reducing its proposed dynamic program to the base’s marginal quote-sizing rule. It shipped a small loss-weight table, choosing 0.9 for spread 1 and 0.7 elsewhere, rather than a large state-dependent policy. Its unseen spread-1 difference was -0.14063334319490464 ± 0.07301492695071997, and its whole-file mean was 97.03279922204433 against 97.06959828018033. This gives no demonstrated improvement. The claim of a solved optimum remains the item’s modeling argument: a limited weight search and the abbreviated evaluation do not prove optimality for cash-constrained inventory and queue dynamics. [T: approach, constants, holdout, by_spread, tried, next]

A wide goal should first merge the frozen branches, verify exact preservation of other spreads, and test the combined policy on fresh shared seeds; the reported holdout is now spent. Preserve separate retail and arbitrage attribution and rare-jump losses. If that test supports expansion, pursue spread-specific belief calibration, demand-aware stock replenishment and inventory-aware queue exposure; use a calibrated jump mixture and a value function where cash binds. Complete the table route’s missing full-range evaluation before comparing it fairly, and resolve spread-4’s failed source check before treating its supporting claims as verified. The present evidence supports that validation stage, rather than funding a large constants-only sweep on the assumption that the additive gain has already transferred. [S1–S4, T: next, approach; validation sequence is this report’s recommendation]

Source keys (the supplied task inputs; no external page is used as independent confirmation):

• [S1]: item-spread-1, supplied logical path artifacts/goal/items/spread-1/result.json; SHA-256 5073494da3b5de0c65a054e476ecbfec02f31d29462de470587b675429b1e101.
• [S2]: item-spread-2, supplied logical path artifacts/goal/items/spread-2/result.json; SHA-256 9b34e5921027c0ef5f8236b1bc81a622c602af34172ce31ef9d353cfcc29a36e.
• [S3]: item-spread-3, supplied logical path artifacts/goal/items/spread-3/result.json; SHA-256 98e4760387ce818687aebc79b9ecb937360636d66e42cdcdf85fcc70e1f63d11.
• [S4]: item-spread-4, supplied logical path artifacts/goal/items/spread-4/result.json; SHA-256 2e3943f3f77127c66a331c96328e4036b6cc44a1b058448c29ba3cc3fbe450e2.
• [T]: item-solve-table, supplied logical path artifacts/goal/items/solve-table/result.json; SHA-256 27f30cc086acf1ff266b5fd3d29c724ec11c9bcaeb0184dfa8fc43aa8a907923.