Prediction market: wave results

Each direction links to its frozen public JSON snapshot, including code, code_sha256, measurements, tried and next. All linked snapshots were retrieved successfully and matched the SHA-256 of the supplied inputs during this revision; their decoded UTF-8 code also matched code_sha256. Strategy code is omitted from this report, as requested, but is retrievable through those links. Code bytes below count the UTF-8 code field, not the JSON snapshot. Scores are item-reported measurements, not independently rerun simulations.

Best reported holdout first; ties are unchanged-base results. Every row reproduces the base on seeds 0–199; training is seeds 0–599 and holdout is seeds 20000–20599, each with 600 simulations and zero reported failures. Retail and arbitrage columns are training edge, with arbitrage losses signed. The linked row snapshot supplies every measurement in that row.

Calibration reproduced the base and found identical per-seed trading results across Python 3.11.18, 3.12.16 and 3.14.4 on seeds 0–199; it did not identify the server seeds. Its unchanged base scores 87.41650726068066 on training and 85.46780344973116 on holdout. calibrate The true-mean item measured 4,000 fresh seeds 1000000–1003999: mean 85.95924845104246, standard error 2.334517677862521, normal 95% interval [81.38367788115984, 90.53481902092507]. Settlement-0 mean was 152.30342442719353 versus settlement-1 mean 18.139266740335945; these are retrospective strata, not a trading signal. It recommends at least 600 fresh confirmation pairs, with sample size based on the particular pair’s difference variance; its tiny normal-theory counts were unreliable under empirical tail checks. true-mean Calibration reports sandbox success on Python 3.12.16, while several other seats report an ntpath import failure on Python 3.12.3; worker-socket restrictions also varied by machine. Treat these as environment-specific reports, not a universal Python-version verdict. calibrate belief As the requester flags one unverified external source for calibrate, its public-server/API interpretations remain provisional; the local measurements above are reported experiment results, not validated by that external citation.

The largest holdout increase is arbitrage-loss: the common base 85.46780344973116 rises to 96.69576967374634. It acquires cheap YES with crossing bids and supplies one-tick-spread asks; training planned acquisition loss falls from 11.03 to 5.57, although jump loss rises from 7.94 to 10.67. arbitrage-loss Late-game’s related trap-bid policy reaches 91.56426545709166; its exact one-step backward law was rejected and no full dynamic programme was built. late-game Other holdout increases, all relative to that same exact base, are belief to 89.58847361470433, rules-and-engine to 88.57704157989234, constant-search to 88.5070473106925, flow-accounting to 87.91919308349567, collateral to 87.87707760418967 and public-record to 86.98656564728267. Their corresponding linked rows give the exact increase endpoints rather than rounded or newly inferred deltas. Belief improves filtering and removes an extra sizing-variance penalty; flow-accounting captures residual near-zero retail buys; rules-and-engine adds overflow liquidity and FIFO slices. belief flow-accounting rules-and-engine The requester flags one unreliable external source each for arbitrage-loss, late-game and rules-and-engine: their strategy descriptions and measured results are useful item reports, but external-source support for their engine explanations is not established. This ranking is a shared-holdout comparison of selected strategies, not an independently confirmed true-mean ranking.

Several directions did not establish a useful out-of-sample improvement. Optimiser reports +0.069 with standard error 0.95 despite its training gain; depth reports +0.826 +- 1.164 for its frozen candidate, while an interim depth-only file reports +1.502 +- 0.639 on the same holdout and needs fresh confirmation. optimiser depth Ammunition-demand calls its gain of ~0.4-0.5 within noise; one-tick-spread reports its spread-only holdout movement as 12.06 to 12.73, mostly noise. ammunition-demand one-tick-spread Erosion, settle-at-one and regime-id finish only slightly above the base in the table; those point estimates alone do not establish significance. erosion settle-at-one regime-id Jump-mixture finishes below the base at 84.89089630473633 despite a training gain; its suggestion to retain the exact map is therefore not supported by this holdout ranking. jump-mixture Worst-runs returns the unchanged base: stronger protection against tail jumps helped the worst runs but cost more elsewhere. worst-runs The requester flags one unverified external source each for settle-at-one and regime-id; in particular, the former’s explanation of why NO cannot mirror YES financing remains an item-reported engine interpretation, not independently verified here. Queue-priority produced nothing and is missing, not an empty or zero-scoring result; no separate queue-priority conclusion can be drawn (supplied task provenance).

The next wave should start from the frozen arbitrage-loss candidate linked in the first row, preserve that exact code and its measurements, and independently rescore it against the unchanged base on fresh paired seeds before any leaderboard submission. Its snapshot contains the 13882-byte strategy with code SHA-256 d68e658732b99a47eb8004f684319f0bc3dc286fe52a066ab019442f8c91aac2. Test demand-scaled crossing inventory, jump-sensitive tick-1 bid exposure and stock-backed queue capture before another broad constant search; these are arbitrage-loss’s proposed next experiments, not measured gains. arbitrage-loss Compare the belief filter and overflow/FIFO logic as separate additions, because their independent gains do not prove that combinations help. belief rules-and-engine Use the true-mean item’s fresh paired pilot and variance-based confirmation design, protect its measurement seeds from tuning, and reserve a new final holdout because the shared holdout has now selected the winner. true-mean The late-game alternative is the strongest reported comparison, while depth-only deserves a fresh test rather than switching to an interim file selected from holdout observations. late-game depth Missing queue-priority work and the six flagged external citations should be resolved before relying on their absent evidence or treating engine interpretations as settled (supplied task provenance).