Prediction-market wave write-up

Cash planning leads the supplied common-holdout results; simplification is the strongest alternative. These are reported local experiments, not a demonstrated public-leaderboard win. This report uses the supplied merges without independently rerunning simulations or verifying external citations. Item names locate the supporting rows and findings in M1 and M2.

Base is the reproduction on seeds 0–599; evaluation total, retail and signed arbitrage edge are seeds 1000–1599; common holdout is seeds 30000–30599. Negative arbitrage edge means loss. Numeric spelling and differing base reproductions are preserved exactly; bytes are the merges' UTF-8 lengths of decoded strategy code. Comparable holdouts rank highest first. The final four rows are excluded from that ranking because of unsupported range labels, a different holdout range, or missing measurements. Unmeasured is not zero. Repeated unchanged-base controls are not independent holdout replications. M1; table and caveats M2; table and caveats

What measurement found. true-mean measured 99.23869339150623 on seeds 40000-43999, with retail 127.07987834308194, arbitrage -27.8411849515945, top-decile share 0.5265243835574145 and estimated standard error 12.37313011835025 for a 200-run mean. draw-variance did not finish its planned 30 x 200 study: its 14 x 200 block means ranged 77.5 to 112.6, and top 20 of 200 contributed 51.6% of edge. Settlement and spread therefore complicate unpaired rankings. oracle-bound's +1.05 for retail parameters versus +8.78 for the next jump flag and +24.34 for next-step probability point toward quote exposure, but are policy-specific hidden-information diagnostics, not attainable gains or global bounds. demand-fit changes policy as well as estimating demand, so its results do not settle that comparison. leaders-gap and tick1-jump-cost show that arbitrage savings can also remove retail profits; rich-runs' expanded ask capacity likewise failed on holdout. M1: true-mean, draw-variance, oracle-bound, leaders-gap, tick1-jump-cost, demand-fit M2: rich-runs

Which combinations added up. combine-cash reports paired holdout gain 2.083619011192334 with standard error 0.4067546948632805, the strongest combination evidence. combine-late-game and combine-three lack their own holdout base controls, so their absolute scores do not establish paired gains. Overflow yields 0.2832087540598333 with standard error 0.3651905827107454; voting yields 0.056340704598667295 with standard error 0.23682551783950723. combine-two-cent's gain 0.036431343287000005 uses its own holdout base 96.00139032710622, explaining why it is positive despite scoring below the common control. Belief replacement loses -1.0599375611583355 with standard error 0.49740626976596103; pair ultimately disables latent belief, and the six-way sequence fails to exceed the common base holdout. Five of six earlier constants failed to transfer; all six together lost 1.87 a run. Separate gains cannot simply be added. M1: combine-cash, combine-late-game, combine-three, combine-overflow, combine-vote, combine-two-cent, combine-belief, combine-pair, combine-all, combine-constants

What fresh starts found. Useful additions beyond the base are censored demand learning for stock targets (fresh-fingerprints), fill-based retail-intensity estimation (fresh-rules-lawyer), and recognizing FIFO priority acquired after submission (fresh-time); the last reports only 0.02733549288983328 paired holdout gain. FIFO chunking in fresh-gifts yields 0.036612788687167254 with standard error 0.2817268065582916, not an established improvement. fresh-free and fresh-opposite rediscovered retention already in the base; fresh-free finally retunes constants, while fresh-opposite and fresh-literal return the base, the latter after holdout-based selection. fresh-competitor has only a two-seed smoke test supporting its graft, not the full ranges its fields label. fresh-tail returns the base after incomplete candidate evaluation and reports different ranges. fresh-solve supplied no strategy after an xcrun runtime failure; wildcard-2 supplied none after download failures. wildcard-1 independently supports demand-learning research, but neither oracle experiments nor public fingerprints prove an executable gain or another entrant's mechanism. M2: findings for all fresh items, wildcard-1 and wildcard-2

What could be taken away. simplify removes secondary purchases and dead controls while retaining probes, beliefs and queue logic; it also changes the stock-entry gate, general risk to 0.9 and the stock-cash cap, so its gain cannot be attributed solely to deletion. Code shrinks from 13882 to 9972 bytes. Its logged paired holdout gain is 1.6671149102061666 with standard error 1.1776597053936184; its approach instead writes 1.6671149102061662. arb-budget also removes secondary stock but misses its joint target: holdout retail is 119.517920457476 and arbitrage -22.0396719198325. ablation-map finds overlapping stock and spread-rule effects: joint stock deletion costs 24.3924403209095, so marginal effects cannot be summed. Last-step and other-ammunition removal intervals include zero; inactive pre/nb1/gs cleanup is reported neutral. no-probes bundles removal with compensating estimator changes yet loses -8.29 +-0.76 on holdout. fewer-cancels' +0.05 +-0.72 does not establish that longer retention helps. Keep the information and cheap-stock core while separately testing secondary-purchase removal. M2: simplify, arb-budget, no-probes, fewer-cancels M1: ablation-map

Where the next wave should start. Start from combine-cash, with simplify as an independent alternative and the unchanged base as paired control. Assign separate tests to purchase removal and its entry gate, demand fitting, demand-matched crossing cover, late-game stock rules and near-tick controls; combine only after independent replication because inventory and collateral effects overlap. cross-anywhere's higher-tick acquisition and tick2-queue's queue commitment deserve separate validation. Freeze policies before a fresh untouched holdout, use paired common-seed differences with uncertainty and tail sensitivity, and preserve the intended settlement/spread distribution: the shared holdout is exposed. No public submission or first-place result is established. Missing measurements and failed external citations remain unresolved: M2 cannot identify the precise failed URLs, and M1 flags provenance failures for several items. Ordinary zero-failure simulations do not certify upstream sandbox compatibility; reports include ntpath import failures and differing runtime outcomes. No missing measurement, strategy or combined result has been fabricated; strategy code is omitted. M1: synthesis, provenance/runtime caveats, combine-cash, demand-fit, combine-late-game, near-tick, tick2-queue M2: synthesis, cross-anywhere, source-failure and sandbox caveats