# Merge 1: reduce adverse selection while preserving retail cover The strongest actionable finding is to restrict expensive routine inventory purchases while retaining crossing purchases and the one-cent reservoir. Spread 2 has a broad one-step response to depleted main quotes; spread 3 has an especially well-supported cap on routine ammunition; spread 4 has a broad improvement from restoring the existing low-price bid risk penalty. Spread 1 offers two promising alternatives, with materially different uncertainty. These are candidate decisions from the supplied experiments, not a demonstrated first-place entry. This merge covers all 20 supplied JSON results. None is missing, empty, or marked blocked. It does not create a new strategy or claim to have run a combined backtest. Each item contains its own complete strategy under `code`; the code hashes below identify those exact submissions. Item references refer to the supplied artifact, specifically its `approach`, `found`, `tried`, `holdout`, `diff`, `others_identical`, and embedded `code` fields. Numerical results below are reported experiments unless explicitly called a merge-time calculation or code inspection. ## Decision for an operator | Spread | Preferred first candidate | Why | Alternatives retained | |---|---|---|---| | 1 | a1-free for a conservative purchase-rule change; a1-jumps for broader run coverage | a1-free lowers routine ammo bids one tick: +0.475181, SE 0.189718, positive trimmed result. a1-jumps splits ordinary asks: +0.833990, SE 0.658410, 1357 improved runs and positive trimmed result, but much greater dispersion. These were not compared head-to-head on one complete set. | a1-middle is a small positive structural change; a1-after is weak; a1-ledger fails the breadth criterion. a1-stock and a1-rich correctly return the baseline. | | 2 | a2-after | +0.619103, SE 0.118830, 1378 up / 630 down on all 2011 confirmation runs, with positive trimmed mean. | a2-stock and a2-ledger are the same effective ammo restriction and count once. a2-jumps is a smaller bridge reduction. a2-middle gains retail while slightly increasing arb loss. a2-rich is insufficiently tested; a2-free returns baseline. | | 3 | a3-ledger | +0.486756, SE 0.027370, 571 up / 32 down, 1417 unchanged, on all 2020 confirmation runs. Routine ammo target is capped by current desired ask size. | a3-stock implements an equivalent restriction of routine ammo, with smaller confirmation coverage and an unresolved scorer mismatch. a3-jumps is a different cover-bid split worth retaining as an alternative. a3-after is weak. a3-middle is tiny, has a scorer mismatch, and its narrative timing description disagrees with its code. | | 4 | a4-rich | +0.589009, SE 0.053431, positive trimmed mean, 792 up / 188 down; restores existing `kL` on low-price bids. | No other spread-4 item was supplied. | Sources for every row: the identically named item records in the evidence inventory below. Preferences are merge judgments based on those records, not new measurements. Do not stack changes within a spread and add their standalone gains: inventory, budgets, queue priority, and subsequent competitor replenishment interact. Changes in different spreads can be composed under mutually exclusive guards, but no supplied item tested the resulting multi-spread file. Keep each candidate exactly as frozen when evaluating it. The brief states that first place earns 109.93 (128.26 retail, 18.34 lost), second earns 109.59 (130.77 retail, 21.18 lost), and the network earns 108.47 (130.37 retail, 21.90 lost), on the same 200 Paradigm runs. Its reported final wealth is 1,036 versus the leader's 1,074; its best spread-4 run about 1,362 versus about 1,440. These are requester-supplied figures, not independently verified board observations. The source for these figures is the task brief; none of the 20 results supplies a board evaluation of its candidate. The brief also reports the historical +0.95 unseen-run change that lost about 2.9 on the board. Therefore none of the results establishes a ranking improvement on those 200 runs. In particular, the apparent inventory surplus is a diagnostic clue, not evidence that selling or buying less mechanically improves scored edge. ## Evidence quality and disagreements **External-source warnings.** The task's source recheck flagged a2-stock (1 source), a4-rich (2), a1-stock (2), a1-middle (1), a2-ledger (1), a2-jumps (1), a3-ledger (1), and a3-stock (1). The supplied JSONs do not contain a source-URL/quotation manifest identifying which sources failed. It is consequently impossible to attribute an individual failed quotation to a particular claim. Treat their external-source support as unverified. Their embedded code, hashes, and reported paired experiments remain inspectable primary item evidence, but their simulation logs and source re-fetches are not delivered here. This merge attributes those measurements to the workers and does not turn them into independently reproduced facts. Matching reports such as a2-stock/a2-ledger are corroboration of a mechanism, not proof of independent replication. The provided original strategy and simulator links are reference locations, not replacements for failed source verification: - [Simulator](https://github.com/danrobinson/prediction-market-challenge). - [Start artifact](https://api.imdnotdead.fun/artifacts/151e0b1b16f90a1384b823bc528d6d1407f7b4bbfcc7966c12119fe483bfd326). - [Earlier scorer artifact](https://api.imdnotdead.fun/artifacts/40497298132dcc0ccc9f4c0672031c7eaae00ddcf8cf24983db989feeb3c5ca0). **Scorer discrepancy.** Most completed checks reproduce 108.63974289615767 on the earlier file's 600 seeds starting at 1000. a1-after did not complete that calibration and records null. a3-stock reports 108.64128535876738, a difference of 0.001542462609705808, using Python 3.10.12 versus the artifact's reported Python 3.12.3; it explicitly leaves the cause unresolved. a3-middle reports 108.6412853587675. Do not silently round either mismatch into agreement or assume Python version explains it. a3-middle's choosing baseline 126.39561823742822 also differs from a3-ledger's 126.39937568895706 on the ostensibly same 1938 choosing seeds. Its complete confirmation baseline inferred from holdout minus diff is approximately 127.136850, close to but not exactly a3-ledger/a3-jumps' 127.13684733215113. Paired within-worker gains should be assessed in that context. a3-stock's confirmation mean is much higher largely because it uses only the first 1010 qualifying seeds; compare paired deltas rather than absolute holdout means. Sources: a1-after, a3-stock, a3-middle, a3-ledger, a3-jumps. **Equivalent changes are not additive.** Code inspection shows a2-stock adds `not (self.s == 2 and t == 2)` to routine ammo, while a2-ledger requires `self.s != 2 or t == 1`. Under the unchanged spread-2 ammo threshold, this branch's attainable ticks are 1 and 2, so these are the same effective restriction; their reported full confirmation statistics agree. Likewise a3-ledger sets `target = min(target, qa)` before the stock branches; only routine ammo uses that target. a3-stock instead requires `Y < qa` there and caps its quantity by `qa-Y`. Algebraically both enforce `Y < min(target,qa)` and buy at most that shortfall; crossing/full-reserve branches remain intact. The slight differences in their paired results are not evidence of additive benefits, given differing runtimes and seed coverage. Sources: embedded code of those four items. **Timing-label correction.** a3-middle calls its change “final quarter,” but its submitted code sets `H = max(1, state.steps_remaining)` and applies the reduction when `H <= 250`. On a 2000-step run this is the last eighth, not the final quarter. Preserve the actual frozen code boundary; do not implement a 500-step version under these reported statistics. It also tested several factors and bands (0.50, 0.75, 0.90; 21–80 and 41–80), despite the brief's instruction not to tune constants. a2-rich compares several numerical ammo settings on a very small, shortened-horizon sample. a4-rich explored numerical horizon/cap/risk probes but explicitly discarded numerical settings and submitted an existing coefficient. a2-middle changed quadrature resolution from 20 to 60 nodes after comparing both halves; its validation is consequently less clean than a single frozen numerical-method choice. These are protocol caveats, not newly established performance failures. Sources: corresponding `tried` and `code` fields. **Jump definitions differ.** a1-free/a1-jumps/a2-jumps/a3-ledger/a3-jumps identify actual Poisson jump-count-positive steps in measurement copies. a2-after uses absolute score move above 0.08; a2-free uses score move above four diffusion SDs; a2-middle flags probability moves above 3 cents. Those proxies are not interchangeable with actual jump events. Band definitions also differ: a1-middle primarily uses pre-move probability, others use order limit/fill price, and some secondary tables use post-move probability. A low ask jumped into the middle band must not be reclassified as an ordinary middle-price quote failure. Sources: named items' `found` fields. **One unusable ledger.** a2-free says its logging wrapper accumulated across worker runs, inflating aggregate attributed losses to roughly 72 against true loss roughly 26. Discard those absolute attributed numbers. Even its ranking is only a worker-reported hint because repeated runs may be weighted unequally; the low-price concentration is supported more reliably by a2-jumps and a2-ledger. Its ordinary uninstrumented strategy comparisons are separately reported, but only 240 choosing and 240 confirming runs were used. **Counts and timing are heterogeneous.** Slow-worker reductions differ substantially: some halved prescribed counts, some used much smaller samples. a1-ledger timed 40 runs and scaled to 200; a1-after scaled 6 to 200; a2-free timed 240 in a pool; several calibration runs overlapped exploration. Wall times across machines are not comparable. Reported equality checks cover finite samples only. a2-free's `others_identical` is true with zero tested runs because the submitted file is unchanged. a1-stock/a1-rich/a2-free are exactly the same baseline code by this merge's SHA-256 inspection. All strategy strings parse as Python and all 20 embedded code hashes match their declared values; parsing/hashing does not validate behavior. ## Findings retained across approaches **Inventory acquisition is a subsidy to later asks.** The large direct arb loss on stock orders is real in the ledgers, but direct retail earnings of those bids omit the later ask sales they fund. Blunt reservoir cuts repeatedly lose. Spread 1's tick-1 stock has 8.793 arb cost per choosing run in a1-stock's full ledger, while a1-ledger measures about 11.07 on a smaller seed window; these are different samples and attribution harnesses, not a reconciled universal number. a1-stock measures about 3344 final YES on the full choosing window, yet demand-based stock halving loses about 6.92 mean edge in its first 1000 runs. a1-jumps similarly loses 6.914 by halving the reservoir rate and 14.186 by capping tick-1 stock at 3000. a2-ledger's shorter reservoir horizon loses 8.045 and disabling that stock loses 9.597 on discovery. a3-ledger's shorter horizon loses 2.568. A reserve can look wasteful at settlement while remaining useful insurance for retail flow. Sources: a1-stock, a1-ledger, a1-jumps, a2-ledger, a3-ledger; externally flagged items remain worker evidence under the caveat above. **The expensive repeated above-estimate top-up is more promising than cheap crossing stock.** a2-stock's FIFO map reports routine ammo costing 1.96 arb per run on about 249.5 arb-bought shares, versus q1's 2.78 on about 3042.4. Eliminating all ammo loses 0.957 in discovery, while disabling only tick-2 ammo gains about 0.122 in discovery and 0.123 on unused choosing runs. In spread 3, a3-ledger measures ammo arb loss 1.608 against only 0.00985 direct retail, the worst stock-rule ratio, but full removal loses 0.266 on discovery. The current-ask cap gains 0.576/0.590 on discovery/unused choosing instead. a3-stock's FIFO sample shows about 1010 unsold YES from ordinary bids below five cents versus about 436 from the one-cent reserve, warning against treating all leftover inventory as one reservoir problem. FIFO resale attribution is descriptive bookkeeping, not a causal assignment of future profit. Sources: a2-stock, a2-ledger, a3-ledger, a3-stock. **Jump exposure is concentrated but generic caution can cost too much.** a1-jumps' complete choosing sample has 3.903 actual jump steps/run, with 9.400 arb loss there and 13.180 on other steps. a3-ledger reports 3.879 actual jump steps/run, with 14.706 loss there and 13.597 elsewhere; front asks alone account for 10.304 jump loss against 124.485 total retail edge. a2-after finds old resting main asks on jump steps cost 8.232/run in its sample. These measurements suggest protecting selected marginal orders, not withdrawing every quote. In a3-jumps, splitting main asks loses 5.811 on discovery even though a1-jumps' ask split gains in spread 1. That disagreement supports spread-specific guards. The a3-jumps order-ID diagnostics also find little prior retail edge on many orders later hit by jumps, but that is an ex-post selected subset and does not supply a live predictor. Sources: a1-jumps, a3-ledger, a2-after, a3-jumps. **Post-hit information helps only with a usable signal.** a2-after measures approximately 0.03875 loss per non-jump step immediately after an arb fill versus 0.00461 on plain steps; the same-side newly placed stock and main bids dominate that next-step loss. Its live proxy is disappearance of tracked main-order IDs, not privileged knowledge that an arbitrageur traded. Retail can also fully consume an order. Probe depletion is nearly ubiquitous (about 1994/2000 steps in its sample), so “any probe filled” is not a useful shock label. a1-after also found tiny-fill/probe flags almost always true. a3-after's resets, broadened beliefs, and broad defensive responses mostly lose; its selected narrow stock delay has weak positive confirmation. These are different mechanisms and sample sizes, not an inconsistency requiring one universal hold-off rule. Sources: a2-after, a1-after, a3-after. **Rich-run capacity gaps are not free scored edge.** a1-rich's 259 choosing rich runs end with about 14977 YES and 946.19 wealth. During their low-probability steps, offered retail dollars 468.356 comprise retained fill edge 304.276, fair value delivered 82.113, size/no-quote residual 67.523, and insufficient-cover residual 14.444. Only part of the dollar gap can be edge; fixed-path opportunities ignore later inventory and competitor dynamics. Extending the 5x ask multiplier to tick 1 loses 6.210 overall and 9.897 on its rich discovery subset. Spread 4's corresponding static ask-size gap is about 9.008 per rich run, no-stock gap 0.526, worse-price gap 0.00258, and low-price bid arb loss 14.791; simply raising the ask cap gains only 0.090 in a 309-run discovery diagnostic, all trimmed away. Its actual chosen bid-risk change gains 1.160 on the frozen confirming rich subset of 253 runs. Do not add the static capacity estimates to the measured candidate gain. Sources: a1-rich, a4-rich (the latter has two failed external source rechecks). **Middle-price bands largely earn their keep.** In a1-middle's pre-move mapping, >=5c appears slightly negative directly, but suppressing all middle quotes loses about 0.889 and suppressing late quotes loses about 0.14, showing downstream value. The narrow bridge cutoff at estimated 5c offers only +0.037533 confirmation. a2-middle's exact transformed diffusion distribution produces +0.110224 overall through +0.156757 retail and -0.046534 arb: useful mean edge, but it does not meet the literal aim of reducing arbitrage loss. Broad ordinary quote-risk restoration loses in spreads 1–3 in multiple items, while restoring the special low-price bid penalty gains in spread 4. The coefficient, band, and acquisition alternatives matter. Sources: a1-middle, a2-middle, a1-free, a1-jumps, a2-ledger, a3-jumps, a4-rich. ## Complete candidate ledger All deltas are candidate minus the supplied start strategy on the same worker's paired confirmation runs, per run of that spread. Positive arb delta means less loss; it is signed edge, not a positive cost. The trimmed statistic generally removes the 20 largest absolute paired differences, not a percentage trim. Zero-effect runs are excluded from up/down counts. Rows are rounded for readability; each source retains full precision. This is not an all-spread score table. | Source | n | Delta | SE | Retail delta | Arb delta | Trimmed | Up/down | Other-spread pairs | |---|---:|---:|---:|---:|---:|---:|---:|---:| | [a1-after](#evidence-a1-after) | 64 | 0.126070 | 0.096832 | 0.002189 | 0.123881 | 0.000337 | 28/8 | 36 | | [a1-free](#evidence-a1-free) | 1027 | 0.475181 | 0.189718 | -0.360230 | 0.835411 | 0.467330 | 227/69 | 450 | | [a1-jumps](#evidence-a1-jumps) | 2032 | 0.833990 | 0.658410 | -0.166255 | 1.000244 | 0.722121 | 1357/584 | 900 | | [a1-ledger](#evidence-a1-ledger) | 506 | 0.665594 | 1.818570 | -3.223409 | 3.889003 | -0.715853 | 97/132 | 200 | | [a1-middle](#evidence-a1-middle) | 2032 | 0.037533 | 0.037321 | -0.151401 | 0.188934 | 0.046876 | 422/272 | 900 | | [a1-rich](#evidence-a1-rich) | 2032 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0/0 | 900 | | [a1-stock](#evidence-a1-stock) | 2032 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0/0 | 900 | | [a2-after](#evidence-a2-after) | 2011 | 0.619103 | 0.118830 | -0.438114 | 1.057218 | 0.612694 | 1378/630 | 900 | | [a2-free](#evidence-a2-free) | 240 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0.000000 | 0/0 | 0 | | [a2-jumps](#evidence-a2-jumps) | 1006 | 0.217720 | 0.107436 | -0.142351 | 0.360071 | 0.111533 | 307/194 | 450 | | [a2-ledger](#evidence-a2-ledger) | 2011 | 0.250538 | 0.053333 | -0.114062 | 0.364600 | 0.142697 | 228/52 | 900 | | [a2-middle](#evidence-a2-middle) | 2011 | 0.110224 | 0.035773 | 0.156757 | -0.046534 | 0.111186 | 1177/834 | 900 | | [a2-rich](#evidence-a2-rich) | 8 | 0.307978 | 0.307978 | -0.175143 | 0.483121 | 0.000000 | 1/0 | 900 | | [a2-stock](#evidence-a2-stock) | 2011 | 0.250538 | 0.053333 | -0.114062 | 0.364600 | 0.142697 | 228/52 | 900 | | [a3-after](#evidence-a3-after) | 1010 | 0.060450 | 0.106578 | -0.067267 | 0.127717 | 0.030033 | 204/157 | 450 | | [a3-jumps](#evidence-a3-jumps) | 2020 | 0.231238 | 0.074481 | -0.087722 | 0.318960 | 0.286872 | 685/340 | 900 | | [a3-ledger](#evidence-a3-ledger) | 2020 | 0.486756 | 0.027370 | -0.006709 | 0.493465 | 0.433424 | 571/32 | 900 | | [a3-middle](#evidence-a3-middle) | 2020 | 0.008105 | 0.001627 | -0.023480 | 0.031585 | 0.006554 | 314/156 | 900 | | [a3-stock](#evidence-a3-stock) | 1010 | 0.470072 | 0.037322 | 0.006068 | 0.464004 | 0.358004 | 271/12 | 450 | | [a4-rich](#evidence-a4-rich) | not stated¹ | 0.589009 | 0.053431 | -0.094781 | 0.683790 | 0.622653 | 792/188 | 900 | ¹ a4-rich's scalar fields report changed-run counts but not an explicit full confirmation count; this merge does not infer total n from 792+188 because unchanged runs may exist. Every item reports zero candidate holdout failures. No delivered per-run result arrays allow recalculation of these standard errors or cross-candidate covariance. The n values above follow each item's stated sampling policy; a3-after's every-other subset of the 2020 qualifying seeds implies 1010. a1-after uses only 64 confirming runs and 36 other-spread checks. a1-ledger's 506 confirming runs and 200 other-spread checks fall well short of the full windows. a2-rich's confirmation is eight runs, with only one improved run and a zero trimmed mean; its development used 500 rather than 2000 steps. It is not evidence for submission despite a positive reported mean. ## Item-specific implementation and retained experiments Each subsection identifies its submitted mechanism through the synthesis and source, holdout totals, identity, proposed follow-up, and a compact index of tested alternatives. This retains negative and inconclusive findings as well as positive ones. Routine repeated descriptions of calibration, timing, freezing, and cross-spread checks are consolidated in the tables and caveats; granular price-band exposure arrays, marginal-tail tables, order-ID tables, and a4-rich's per-seed CSV are summarized above rather than reproduced. They remain source diagnostics, not separate selectable gains. No raw strategy is silently promoted into a merged executable. Experiment records below abbreviate paired statistics as m (mean), se, r (retail delta), a (arb delta), tr (trimmed), u/d (improved/worsened), med, fail. Decimal measurements are rounded to five significant digits in the experiment index to fit the bounded merge. These records are attributed entirely to the subsection's input source. The bounded experiment index below retains abbreviated source experiment labels and available mean/uncertainty statistics; mechanisms are explained in the synthesis and original items. Ellipses mark condensed descriptions. Extended repeated discussion, secondary statistics and nested selection objects are omitted explicitly, not treated as absent experiments. Repeated summaries may remain when they document separate selection decisions. ### a1-after Source: [artifacts/goal/items/a1-after/result.json](../../items/a1-after/result.json), supplied as `.imd/reads/artifacts/item-a1-after`. Input SHA-256 `eef93bef6ed137eb972c423808e9e38452603318fddc62dc7d2fd1a4de69177b`. Strategy SHA-256 `29f64fc51c4390fbf67ec1ced247abcb9fda7306eb26561ab6ca7b1075e6733f`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 33.975893331; retail 55.262881865; signed arb -21.286988533; failures 0. Calibration mean: None. - Idea A: halve both sizes for 2 steps after any own fill >0.05 shares: choosing diff -1.16 a run… — m -1.16. - Idea B (kept): halve sizes for 1 step after a >=2 tick rival quote shift: choosing diff +0.484… — m 0.484, se 0.097. Follow-up hypothesis: Distinguish arbitrageur fills from retail fills at the fill step (e.g. fill taken against a quote that was inside the previous belief interval, or probe consumed on the far side) so the hold-back applies only after arb hits; test longer holds on a side-specific basis. Use more runs than the ~5 s/sim machine allowed here. ### a1-free Source: [artifacts/goal/items/a1-free/result.json](../../items/a1-free/result.json), supplied as `.imd/reads/artifacts/item-a1-free`. Input SHA-256 `c7f4f274520f484181ba0e664f87875a63b1b410c8c8fbb66cd5d32448cfbabd`. Strategy SHA-256 `aed6d158efa525beefc4a55d511818df9ba8c2c40cfb9bcc62a4978f3a9b020a`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 42.587042140; retail 65.802751606; signed arb -23.215709466; failures 0. Calibration mean: 108.63974289615767. - Idea before implementation: for spread 1, use the already existing W=200 stock horizon in place… — m -14.12, se 3.2128. - Idea before implementation: for spread 1, place the 1-cent stock bid only when the competitor… — m 0.2212, se 0.40441. - Idea before implementation: spread 1 retains oversized front asks at ticks 1-2 even when the… — m -4.5511, se 1.1055. - Idea before implementation: in spread 1 suppress ordinary qbid quotes below the existing pL=0.05… — m -4.0537, se 0.93152. - Idea before implementation: spread 1 uses full modeled arbitrage cost (kA=1) for ordinary… — m -0.23111, se 0.15087. - Idea before implementation: spread 1 never pays above its probability estimate for the broad… — m 0.24767, se 0.30187. - Pre-selection split: first 300 of the 1014 spread-1 seeds among 30000..33999 were training; the…. - Selected candidate choosing validation — m 0.36086, se 0.091644. Follow-up hypothesis: Test allowing the shifted ammunition bid to rest behind the rival at the lower tick, rather than skipping it when below bid_t. Keep the existing long reserve; shortening it damaged retail edge and did not reliably reduce arbitrage loss. Use fresh choosing validation before any new confirmation. ### a1-jumps Source: [artifacts/goal/items/a1-jumps/result.json](../../items/a1-jumps/result.json), supplied as `.imd/reads/artifacts/item-a1-jumps`. Input SHA-256 `61d3ba0c42cd1e20117ad7e31d1b2c8aad12f63f5075d11cdd5af9d277213b66`. Strategy SHA-256 `790b81e5d092f71e1f5a98b4ff5c7aaff575b39bc62063004afd415ad1d4082c`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 39.923287563; retail 62.329117366; signed arb -22.405829803; failures 0. Calibration mean: 108.63974289615767. - Choosing split fixed before candidate evaluation: 1014 spread-1 seeds in [30000,34000) for…. - stockhalf (halve tick-1 stock target rate) — m -6.9144, se 1.7019. - stockhorizon (halve tick-1 target horizon) — m -0.21003, se 0.21853. - stockcap (cap each tick-1 stock order at existing qmax=3000) — m -14.186, se 2.1212. - stockbudgethalf (halve cash-budgeted tick-1 quantity) — m -0.77953, se 0.7228. - bridgehalf (halve inventory-cover bid quantity at ticks >=2) — m 0.2025, se 0.25662. - lossunit (use full modeled loss rather than 0.7 for ordinary non-low-price sizing) — m -0.17619, se 0.12652. - askhalf (halve the special tick-2 ask multiplier) — m -0.31271, se 0.27127. - asksplit (split ordinary asks at ticks 3..98 evenly, back half one tick higher) — m 0.7664, se 0.17726. - ask2split (split ordinary asks at ticks 2..98 evenly, back half one tick higher) — m 0.72859, se 0.40484. - bridgesplit (split inventory-cover bids at ticks >=2 evenly, back half one tick lower) — m 0.20542, se 0.25628. - bridgesplit_validation reserved choosing paired vs start — m -0.10697, se 0.21666. - ask2split_validation reserved choosing paired vs start — m 0.37728, se 1.2059. - asksplit_validation reserved choosing paired vs start — m 1.284, se 0.87651. - Frozen after all ten fixed training controls and three reserved validation checks, before any…. - Baseline one-cent stock bid last-half direct retail edge 0.12813265507475888 versus jump loss…. - Frozen candidate confirmation subrange [90000,94000) — m 0.41003, se 0.66819. - Frozen candidate confirmation subrange [94000,98000) — m 1.2672, se 1.143. Follow-up hypothesis: Keep the tick-1/2 retail engine intact. Investigate inventory-cover bids using observable stock availability and time remaining; blunt stock cuts lost too much retail. Study terminal/boundary risk separately: including tick-2 asks in the split produced large rare losses and much higher validation variance. Use fresh choosing and confirming ranges for any new decision. ### a1-ledger Source: [artifacts/goal/items/a1-ledger/result.json](../../items/a1-ledger/result.json), supplied as `.imd/reads/artifacts/item-a1-ledger`. Input SHA-256 `826299021f309244740ddd9d6ded3c3957ddb24230dbd70a45e15178c7015121`. Strategy SHA-256 `e542c50ad3bca651d5a9070e8df5d451e08edfdb3aecfae8189a5069d129a38e`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 43.621386245; retail 68.592806384; signed arb -24.971420139; failures 0. Calibration mean: 108.63974289615767. - v1, q1 *= 0.5 when s == 1: +0.5164873674717556 a run on 262 choosing runs (30000-30999; 68 up,…. - v2, q1 = 0 when s == 1 and p above 3 cents: -0.16701916734580138 on 262 choosing runs (7 up, 6…. Follow-up hypothesis: The gain comes from fewer, larger runs (more runs down than up), and Paradigm's runs punished narrow gains before. Instead of halving q1 for all spread-1 runs, try sizing the tick-1 stock from the YES the run's own asks actually sold (its realised cover demand). Also cut main-ask arbitrage losses on jump steps (4.44 a run). ### a1-middle Source: [artifacts/goal/items/a1-middle/result.json](../../items/a1-middle/result.json), supplied as `.imd/reads/artifacts/item-a1-middle`. Input SHA-256 `223247cb4c2eec5b61bac2887d098088516f73b062cd077e5299b10d72b9aae3`. Strategy SHA-256 `233b6a973c724047fa9b022e42b54c4bacede7fd932117819f6ad46c43e6a6dc`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 39.126831245; retail 62.343971197; signed arb -23.217139952; failures 0. Calibration mean: 108.63974289615767. - V1, no bridge buy at floor(p) when s==1 and phat>=5c (bridge kept for 2-5c): on the first 1000… — m 0.061432, se 0.12926. - V2, no tick-1 stock bid (q1) when s==1 and ptick>=5: on the first 1000 spread-1 choosing seeds… — m -0.025903, se 0.1338. - V3, no main bid/ask when s==1 and 5c<=phat<=95c: on the first 1000 spread-1 choosing seeds… — m -0.88854, se 0.19632. - V6, on V1: kA=1.0 (smaller quotes) in the s==1 mid band: on the first 1000 spread-1 choosing… — m -0.24916, se 0.10558. - V7, on V1: kA=0.5 (larger quotes) in the s==1 mid band: on the first 1000 spread-1 choosing… — m -0.46846, se 0.13638. - V8, on V1: dmin 0.6 ticks in the s==1 mid band: on the first 1000 spread-1 choosing seeds… — m -0.75792, se 0.15028. - V9, on V1: dmin 0 in the s==1 mid band: on the first 1000 spread-1 choosing seeds (30000-37999),… — m -0.014663, se 0.013987. - V11, on V1: no main quotes in the s==1 mid band when H<500 (last quarter): on the first 1000… — m -0.1402, se 0.046216. - V12, on V1: no main bid in the s==1 mid band when H<500: on the first 1000 spread-1 choosing… — m -0.14805, se 0.04125. - V14, bridge one tick lower when s==1 and phat>=5c: on the first 1000 spread-1 choosing seeds… — m 0.021593, se 0.1275. - V15, bridge one tick lower only when ptick-floor(ptick)<0.5, s==1 and phat>=5c: on the first… — m 0.045444, se 0.095825. - V16, on V1: no main quotes when s==1 and 80c<=phat<=95c: on the first 1000 spread-1 choosing… — m -0.0065395, se 0.0054959. - V17, on V1: no main quotes when s==1 and 50c<=phat<=95c: on the first 1000 spread-1 choosing… — m 0.00088806, se 0.0056128. Follow-up hypothesis: The mid band is a small slice of spread 1's edge (about 2.3 a run from retail and 2.4 lost, against 60.3 and 21.3 below 5c), and its quotes are tuned: size and distance changes all lost. The gain is small (+0.0375 with SE 0.0373 on 2032 confirming runs), so treat it as a low-risk add-on. The spread-1 arb loss to go after is below 5c: tick-1 stock bids hit on downward jumps, and low asks crossed by upward jumps (-2.1 a run of asks filled at 20-50c in the last quarter). Next seat: try making the bridge's goal depend on the bands below 5c rather than on qa. ### a1-rich Source: [artifacts/goal/items/a1-rich/result.json](../../items/a1-rich/result.json), supplied as `.imd/reads/artifacts/item-a1-rich`. Input SHA-256 `99fe30b13f0a60a7b5e802fb9e9d35534d57b81c47d95d6385c173762e1f1ec6`. Strategy SHA-256 `4bbf720ff34be09b441e84a7e0c1bf47cba036b5211c4118e44a597f941453db`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 39.089297777; retail 62.495372076; signed arb -23.406074299; failures 0. Calibration mean: 108.63974289615767. - Extend the existing 5x low-price ask multiplier from tick 2 to tick 1. Discovery spread-1 seeds… — m -6.2096, se 2.0572. - Spread 1 only: shorten Wf from 2000 to 500 for all stock calculations. Discovery spread-1 seeds… — m -4.1155, se 1.0343. - Spread 1 only: queue at tick 2 instead of fronting tick 1 and permit replenishment at the rival… — m -5.1634, se 2.7261. - Suppress one-cent ammunition when asking one cent; a no-op because the existing ammunition block… — m 0, se 0. - Activate tick-2 queueing only after 500 observed estimated-price-below-2-cent steps. Discovery… — m -1.6917, se 0.46874. - Shorten only the one-cent crossing-stock horizon to 500; retain every other stock horizon.… — m -4.0212, se 1.0165. - Reuse the supplied spread-4 censored-flow estimator, constants and one-cent eligibility for… — m 0.35331, se 0.40229. - Same flow-based stock target, with a half-tick minimum observation distance for spread 1; retain… — m 0.0011439, se 0.39426. - Fixed flowstock candidate on reserved choosing seeds [34000,37999], not used to design it — m -0.095658, se 0.41331. - Return the exact start file. Full discovery contained 1014 spread-1 runs, with 1021 reserved… — m 0.0011439. Follow-up hypothesis: Test a probability- and jump-risk-aware cap on one-cent crossing-cover bids while retaining higher-tick stock horizons and established tick-2 FIFO priority. Do not reuse the rejected blanket size/horizon changes as improvements. ### a1-stock Source: [artifacts/goal/items/a1-stock/result.json](../../items/a1-stock/result.json), supplied as `.imd/reads/artifacts/item-a1-stock`. Input SHA-256 `af8796c1c81547eb5de5b4c8b1ce60ea2e425dca6ecfde94bdf2d6da8a8a5190`. Strategy SHA-256 `4bbf720ff34be09b441e84a7e0c1bf47cba036b5211c4118e44a597f941453db`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 39.089297777; retail 62.495372076; signed arb -23.406074299; failures 0. Calibration mean: 108.63974289615767. - v1 stock demand converted at tick 2 (halves the tick-1 target), first 1000 choosing runs: mean… — m -6.9204, se 1.7241. - v3 rest the tick-1 stock bid only while the belief is under 2 cents: choosing runs 0-999 mean… — m -0.0959, se 0.3911. - v4 rest the tick-1 stock bid only when the belief is within 3 one-step sds above tick 1: first… — m -12.027, se 1.863. - v13 (v1+v3): mean -7.1897 SE 1.7393; v14 (v1+v4): mean -14.4119 SE 2.256, both on the first 1000… — m -7.1897, se 1.7393. - v8 also rest the tick-1 stock bid when the belief is within g_hi (0.15 cent) under tick 1, where… — m -0.0218, se 0.0206. Follow-up hypothesis: Do not spend another seat on the tick-1 stock's price or size: it is already bought at the crossing and the deep jump fills are recovered at tick-1 asks. Two measured losses remain open in spread 1: main bids filled by the arbitrageur below 5 cents (1720 shares, 2.80 a run) and asks filled by the arbitrageur below 5 cents (1640 shares, 6.94 a run, 2.03 of it at tick 2 and 2.90 at tick 1). The tick-1 asks are sized by qsize against a belief that cannot see below tick 1 (no probe there), so a seat could test sizing tick-1 asks from the time spent under tick 1 (self.tb) rather than the belief alone; prove it per run against the start file on the 2035 spread-1 seeds in 30000-37999 before confirming on 90000-97999. ### a2-after Source: [artifacts/goal/items/a2-after/result.json](../../items/a2-after/result.json), supplied as `.imd/reads/artifacts/item-a2-after`. Input SHA-256 `cabede3c4180c49f8c5b5d6746b51f47bff599a891a34d9777555c81ee7dfcb1`. Strategy SHA-256 `963807d6db6ddd2ccf3e594d938edd2c4a7452c7b1172733513b8f013d3688ab`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 83.843677439; retail 109.576113375; signed arb -25.732435937; failures 0. Calibration mean: 108.63974289615767. Follow-up hypothesis: Try a hold-back on the step after the rival's quote shifts 2+ ticks (k1 loses 0.0144 a step against 0.0046 plain). Try shrinking the held side by a fraction instead of setting it to zero, to win back some of the 0.44 a run of retail edge this gives up. Try the same one-step hold in spreads 1, 3 and 4, each proven on its own seeds. Separately, main asks resting from earlier steps and hit on jump steps cost 8.2 a run in spread 2, which is the largest single loss. ### a2-free Source: [artifacts/goal/items/a2-free/result.json](../../items/a2-free/result.json), supplied as `.imd/reads/artifacts/item-a2-free`. Input SHA-256 `7270cc5c994de53f2e1b63280d773928fc38a264411f0be12b060541510ad955`. Strategy SHA-256 `4bbf720ff34be09b441e84a7e0c1bf47cba036b5211c4118e44a597f941453db`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 76.082888672; retail 100.522617075; signed arb -24.439728403; failures 0. Calibration mean: 108.63974289615767. - A: in spread 2, halve the 2-10c bridge goal (0.3 -> 0.15 of wealth) to buy fewer cover bids the… — m 77.702. Follow-up hypothesis: Low asks hit on jump steps are the largest spread-2 arb loss. Try pulling asks below 5c for a step or two after a probe or the competitor's quote shows an upward move, or size low asks by the jump-up tail alone. Use more runs than 240: idea C's effect was smaller than its noise. ### a2-jumps Source: [artifacts/goal/items/a2-jumps/result.json](../../items/a2-jumps/result.json), supplied as `.imd/reads/artifacts/item-a2-jumps`. Input SHA-256 `0e0489ba0325acf8d6044dd8188651048aec3a55f38182f70b150d077bfd22fa`. Strategy SHA-256 `1ce5f0a6183adf0ff040e084e0ed269e26360fe11cffe457df842b82f64ead56`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 83.812523038; retail 109.587006736; signed arb -25.774483698; failures 0. Calibration mean: 108.63974289615767. - Exploratory stockhalf (q1 *= 0.5 before budgeting, self.s == 2 only) on first 150 spread-2 seeds… — m -1.1265, se 0.45037. - Exploratory stockcap (q1 = min(q1, existing qmax=3000) before budgeting, self.s == 2 only) on… — m -3.0112, se 1.4129. - Exploratory bridgehalf (qam *= 0.5 when 0.02 < phat < 0.1, self.s == 2 only) on first 150… — m 0.12686, se 0.17197. - Exploratory askhalf (qa *= 0.5 before stock calculation and management, self.s == 2 only) on… — m -1.8663, se 3.2121. - Exploratory bridgesplit (split qam equally between amt and amt-1 when 0.02 < phat < 0.1,… — m 0.035718, se 0.16003. - Exploratory bridgereserve (halve the placed qam quantity at 0.02 < phat < 0.1 but retain its… — m 0.12733, se 0.17199. - bridgehalf independent choosing validation, alternating spread-2 seeds from 30000..37999… — m 0.13888, se 0.098992. - bridgesplit independent choosing validation, alternating spread-2 seeds from 30000..37999… — m 0.1557, se 0.10189. - bridgereserve independent choosing validation, alternating spread-2 seeds from 30000..37999… — m 0.1357, se 0.099028. - Selection rule fixed before confirmation: positive mean and trimmed choosing-validation…. - Frozen candidate confirming result before accept/reject — m 0.21772, se 0.10744. Follow-up hypothesis: Measure marginal retail liquidation value of bridge stock by current YES inventory and remaining horizon. Try reducing only inventory surplus beyond near-term ask demand; preserve tick-1 stock, since naive reservoir cuts increased whole-run arbitrage loss in these choosing runs. ### a2-ledger Source: [artifacts/goal/items/a2-ledger/result.json](../../items/a2-ledger/result.json), supplied as `.imd/reads/artifacts/item-a2-ledger`. Input SHA-256 `1d3fdd0207683d66c09de9f06c83f6cb8ea9f61f5e1f1143b22a9e740e8cc2e3`. Strategy SHA-256 `e984bb0a2803f6e25072d583cc2b908ea7085b1481799999cd3303f3210a4b85`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 83.475111851; retail 109.900165835; signed arb -26.425053984; failures 0. Calibration mean: 108.63974289615767. - Discovery was fixed to spread-2 seeds [30000,34000); validation fixed to [34000,38000); final…. - Tick-1 stock horizon: use existing W=150 rather than Wf=2000 in spread 2; discovery paired — m -8.0445, se 1.407. - Disable tick-1 stocking only in spread 2; discovery paired — m -9.5968, se 1.4497. - Disable target-stock branch only in spread 2; discovery paired — m -0.95691, se 0.34768. - Low-price main bid: use existing ordinary kA loss weight instead of inventory discount in spread… — m -0.56925, se 0.44909. - Keep target-stock branch only at tick 1 in spread 2; preserve crossing buys and tick-1 stock… — m 0.12185, se 0.078674. - Frozen final structural candidate, target stocking only at tick 1: untouched validation on all… — m 0.1233, se 0.039592. - Selection before confirmation: Selected target_floor_only before examining confirming seeds.… — m 0.12185, se 0.039592. Follow-up hypothesis: Treat stocking losses as inventory acquisition costs, not independently removable losses. Next measure the downstream ask profit and main-bid risk displaced by each stock fill; focus on the front asks on jump steps before making another stock cut. ### a2-middle Source: [artifacts/goal/items/a2-middle/result.json](../../items/a2-middle/result.json), supplied as `.imd/reads/artifacts/item-a2-middle`. Input SHA-256 `8cf0f13e127b020216a1033a1408db3bbe700efcf83213e45c8f03bbc9501521`. Strategy SHA-256 `72235757ce93abcd6898ccf7c9ad7c6d1d2a8aee6893ffd9a86e2da15de340d8`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 83.334797536; retail 110.170984648; signed arb -26.836187112; failures 0. Calibration mean: 108.63974289615767. - Seeds: 1977 spread-2 seeds in 30000-37999 (choosing), split alternately into half A (989) and…. - v1: Loss weight kA 0.7 -> 1.0 (risk-neutral) for spread 2 when 5c <= phat <= 95c: shrinks every… — m -0.45147, se 0.1165. - v2: Per-run jump intensity: uniform prior 0.0008-0.003, Poisson count of rival-quote jumps of >=… — m 0.03403, se 0.055615. - v3: Exact one-step distribution of p through the probit map (20 equal-mass normal nodes,… — m 0.10325, se 0.061611. - v4: Same as v3 with 60 nodes (smoother tails); frozen as the final file. Half A vs start: diff… — m 0.13888, se 0.041919. - v4 minus v3: half A 0.03563411570748235 (SE 0.047172517508056004), half B 0.0534478822078949 (SE… — se 0.047173. Follow-up hypothesis: Shrinking mid-band quotes across the board loses (v1: retail falls more than the arbitrageur's take), and the per-run jump rate cannot be learned from about one visible jump per run (v2). The exact one-step distribution helps, so the next seat could carry the same exact treatment into the jump part of AB (the 6 NODES use a fixed sig_j 0.35 with the linear cdf map) and into spreads 1, 3 and 4 under their own guards. They could also try the exact distribution below 5c for spread 2, where most of the arbitrageur's take sits (bids filled under 5c on the crossing). ### a2-rich Source: [artifacts/goal/items/a2-rich/result.json](../../items/a2-rich/result.json), supplied as `.imd/reads/artifacts/item-a2-rich`. Input SHA-256 `15990676e9113b7842723b1274a10a0c977b800be7ed51e0bfead0e9440c6131`. Strategy SHA-256 `cacb6c7ee88475f0fa98ad2fe70e6504ed690fa7c2bb1014f699a205ec16f1c9`. For spread 2 only, reduced the low-probability YES ammunition reserve (ammo_pmax 0.005 and fy 0.2) so rich runs do not retain as much YES to expire; all other spread logic is unchanged. Reported holdout: edge 28.959767069; retail 38.804670958; signed arb -9.844903888; failures 0. Calibration mean: 108.63974289615767. - On 8 untouched spread-2 seeds 90000-90019, the candidate mean edge was 28.959767069349997 versus…. Follow-up hypothesis: Test a broader inventory-aware reserve schedule on the rich regime, especially whether the remaining YES reserve can be reduced smoothly without affecting non-rich spread-2 runs. ### a2-stock Source: [artifacts/goal/items/a2-stock/result.json](../../items/a2-stock/result.json), supplied as `.imd/reads/artifacts/item-a2-stock`. Input SHA-256 `cf52846fa0106a2687982c19f61bbf24755cf778d969e995b76f0287b7005f9a`. Strategy SHA-256 `fe70f7b991b517bf74687812e54b14bf495c75978e74e108a8bb88e2716b1392`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 83.475111851; retail 109.900165835; signed arb -26.425053984; failures 0. Calibration mean: 108.63974289615767. Follow-up hypothesis: In spread 2, q1 leaves ~916 YES/run unsold at expiry; size it to the run's remaining retail buy flow at the actual ask (e.g. measured fills) rather than 0.6*H/0.02. Also check the bridge rule (0.28c per share later sold to retail) for buying at the tick crossing instead of at floor(p). Try the same tick-2 ammo cut in spreads 1, 3, 4 (each on its own seat). ### a3-after Source: [artifacts/goal/items/a3-after/result.json](../../items/a3-after/result.json), supplied as `.imd/reads/artifacts/item-a3-after`. Input SHA-256 `cc477a8885c267bff72dffbe2ec2b5347d8305cd9224efc96037278e19484fe0`. Strategy SHA-256 `4b355108973846632dfa97b86aec145210a2185176176bae66779c0b1da6e2f4`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 126.583554873; retail 155.398828002; signed arb -28.815273129; failures 0. Calibration mean: 108.63974289615767. - screen reset_shift (reset sub-tick belief to uniform on a rival quote shift of at least two ticks) — m -0.011243, se 0.028105. - screen reset_probe (reset belief after any interior probe depletion) — m -0.88677, se 0.4102. - screen reset_both (reset belief after either quote-shift or interior-probe signal) — m -0.87307, se 0.4118. - screen wide_shift (broaden predictive Gaussian SD to at least half the feasible interval width… — m -0.018418, se 0.011868. - screen wide_both (broaden predictive Gaussian after quote-shift or interior-probe signal) — m -0.27416, se 0.30412. - screen cautious_shift (halve bid, ask, ammunition and tick-1 stock quantities for the callback… — m 0.097847, se 0.071047. - screen cautious_both (halve quantities after quote-shift or any interior-probe signal) — m -1.0915, se 1.8984. - screen wide_persist (broaden predictive Gaussian for three callbacks after either signal) — m -0.28196, se 0.39346. - screen2 reset_probe_surprise (reset belief only after probe crossing more than half a tick… — m -1.2544, se 0.47306. - screen2 reset_both_surprise (reset after quote shift or surprising probe) — m -1.2664, se 0.47296. - screen2 wide_both_surprise (broaden predictive Gaussian after quote shift or surprising probe) — m -0.64215, se 0.40187. - screen2 wide_persist_surprise (broaden predictive Gaussian for three callbacks after quote shift… — m -0.85639, se 0.40681. - screen2 cautious_both_surprise (halve quantities for one callback after quote shift or… — m -1.3885, se 0.4357. - screen2 down_stock1 (suppress stock/ammunition purchases for one callback after downward quote… — m 0.0073482, se 0.036673. - screen2 down_stock3 (suppress stock/ammunition purchases for three callbacks after downward signal) — m 0.024016, se 0.059117. - screen2 down_bid3 (halve regular bids and suppress stock/ammunition for three callbacks after… — m -1.1379, se 0.19676. - screen2 direction1 (halve only the quote side exposed to the observed move and suppress stock… — m -1.3594, se 0.29774. - screen2 direction3 (same directional response for three callbacks) — m -2.0772, se 0.46282. - screen3 full_probe (use completely depleted probe sides as probability bounds, retaining… — m -15.708, se 2.03. - screen3 full_probe_reset (full-depletion bounds plus uniform reset on complete probe-side depletion) — m -15.63, se 1.9666. - screen4 main_half (halve all quote and stock quantities for one callback after a previously… — m -1.1995, se 0.51997. - screen4 main_reset (uniform belief reset on tracked main-order disappearance) — m -0.25968, se 1.0488. - screen4 main_wide (broaden predictive Gaussian on tracked main-order disappearance) — m -1.0259, se 0.53138. - validation cautious_shift (halve bid, ask, ammunition and tick-1 stock quantities for the… — m 0.18416, se 0.11506. - validation cautious_both_surprise (halve quantities for one callback after quote shift or… — m -1.807, se 0.73185. - validation full_probe (use completely depleted probe sides as probability bounds, retaining… — m -21.592, se 3.1481. Follow-up hypothesis: Explain the large negative stock-holdback outliers using only callback-visible inventory, cash and directional probe evidence. Test keeping cover purchases when near-term ask demand would otherwise become uncovered, rather than strengthening a generic post-move pause. Use new choosing and confirming seeds; this small positive mean remains statistically uncertain. ### a3-jumps Source: [artifacts/goal/items/a3-jumps/result.json](../../items/a3-jumps/result.json), supplied as `.imd/reads/artifacts/item-a3-jumps`. Input SHA-256 `ef4b96b2d811a07bb2c6308f6b57fc5f3acf7dc9d337c9d7536e6dce982a73a7`. Strategy SHA-256 `7357e4039b2f130895dd7afda173731249bb055dcad94c53404a4f31fcb15286`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 127.368085219; retail 155.420196529; signed arb -28.052111311; failures 0. Calibration mean: 108.63974289615767. - Halve q1 and qam replenishment quantities before cash caps — m -0.094278, se 0.18225. - Halve main bid desired size below estimated 5 cents — m 0.204, se 0.10796. - Halve main ask desired size below estimated 5 cents — m -1.9948, se 0.72405. - Split main ask size equally between original tick and one tick higher, retaining both levels — m -5.8108, se 0.69775. - Split qam cover bid at tick >=2 equally between original tick and one tick lower — m 0.38108, se 0.062808. - Split main bids below estimated 5 cents between original tick and one tick lower where valid — m 0.064141, se 0.092377. - Halve only main desired bids at tick 1 — m 0.091816, se 0.041546. - Halve the demand-based tick-1 inventory target before subtracting inventory — m -1.8056, se 0.39057. - Combine cover-bid split and main bid size halving below estimated 5 cents — m 0.56146, se 0.14074. - Selected cover split for gains in all four choosing-seed quarters and lower paired dispersion… — m 0.046169, se 0.17836. - Selected cover split choosing quarter 30000:31000, n=241; — m 0.29835, se 0.12249. - Selected cover split choosing quarter 31000:32000, n=245; — m 0.59665, se 0.15057. - Selected cover split choosing quarter 32000:33000, n=241; — m 0.37607, se 0.098612. - Selected cover split choosing quarter 33000:34000, n=232; — m 0.24456, se 0.12379. - Frozen cover split and start scored separately in unmodified engine on all 2020 spread-3 seeds…. Follow-up hypothesis: Test whether inventory already exceeds remaining retail demand before posting tick-1 stock orders; distinguish necessary ask cover from excess stock. Keep a reserved choosing partition and check broad paired gains before confirming. Do not cut profitable low asks wholesale. ### a3-ledger Source: [artifacts/goal/items/a3-ledger/result.json](../../items/a3-ledger/result.json), supplied as `.imd/reads/artifacts/item-a3-ledger`. Input SHA-256 `ad3b2ad91284e519d0c80ecfe18eac37e0853b90703d564341744a9639e2a761`. Strategy SHA-256 `3bcbb5f3d400eb31c7ed5a1f45929987613cea9e96b2df850e5ed1af8335b448`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 127.623603422; retail 155.501209818; signed arb -27.877606396; failures 0. Calibration mean: 108.63974289615767. - Remove routine above-fair ammunition purchases (crossing purchases remain); train n=969 — m -0.26591, se 0.41871. - Shorten tick-1 target horizon to existing ammunition W=200, only spread 3; train n=969 — m -2.5678, se 0.52718. - Remove routine purchases; independent validation n=969 — m 0.020033, se 0.24582. - Ask-size cover cap; independent validation n=969 — m 0.58962, se 0.089325. - Frozen selection: Ask-size ammunition cap chosen on first 969 spread-3 seeds; the same fixed cap…. Follow-up hypothesis: Attribute downstream front-ask retail credit to the stock acquisition rule before pruning more stock: the one-cent-horizon ablation increased whole-strategy arbitrage losses. Next test an observable expected-retail-use cap on stock only after measuring its effect on front-ask availability; prioritize the front-ask jump-loss ledger rather than applying a blanket size cut. ### a3-middle Source: [artifacts/goal/items/a3-middle/result.json](../../items/a3-middle/result.json), supplied as `.imd/reads/artifacts/item-a3-middle`. Input SHA-256 `165519ded198f623cc152bc5221cbdbb9462516f11edea9561e12793a4aa3751`. Strategy SHA-256 `fa4915870fda7a751d1ebc76abbe5ddd788a0177e5476b7a5fb2bb452c4c3fb8`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 127.144954983; retail 155.484438496; signed arb -28.339483514; failures 0. Calibration mean: 108.6412853587675. - All 5-95 cent quotes x0.75, development 152 runs: mean -0.2969885289651334, SE… — m -0.29699, se 0.47008. - All 5-95 cent quotes x0.50, development 152 runs: mean -3.207347074519081, SE… — m -3.2073, se 1.0027. - Final-quarter 5-95 cent quotes x0.75, development 152 runs: mean +0.39744333992565734, SE… — m 0.39744, se 0.32594. - Final-quarter 5-95 cent quotes x0.75, disjoint validation 308 runs: mean +0.12735311769350607,… — m 0.12735, se 0.17029. - Final-quarter 5-95 cent quotes x0.75, remaining validation 1478 runs: mean +0.29489368541677957,… — m 0.29489, se 0.15348. - Final-quarter 5-95 cent quotes x0.75, confirming 2020 runs: mean +0.2754449933940596, SE… — m 0.27544, se 0.17693. - Final-quarter 5-95 cent quotes x0.50, disjoint validation 308 runs: mean -0.048146473453571766,… — m -0.048146, se 0.33988. - Final-quarter mixed bid x0.50/ask x0.75, disjoint validation 308 runs: mean +0.0741006027730513,… — m 0.074101, se 0.1752. - Final-quarter 41-80 cent quotes x0.75, development 152 runs: mean +0.011824018842105079, SE… — m 0.011824, se 0.005876. - Final-quarter 41-80 cent quotes x0.75, remaining validation 1478 runs: mean… — m 0.011065, se 0.0016409. - Final-quarter 41-80 cent quotes x0.50, development 152 runs: mean +0.012941749760526247, SE… — m 0.012942, se 0.012499. - Final-quarter 21-80 cent quotes x0.75, development 152 runs: mean +0.014491935536183956, SE… — m 0.014492, se 0.013801. - Final-quarter 41-80 cent quotes x0.90, development 152 runs: mean +0.0044716468717104045, SE… — m 0.0044716, se 0.0022099. Follow-up hypothesis: Test a smooth size multiplier based on remaining horizon and quote distance inside 41-80 cents, but choose the functional form before scoring and require positive trimmed gains on a fresh large seed range; do not broaden into 5-40 or 81-95 where the mapped retail-to-arbitrage tradeoff was stronger. ### a3-stock Source: [artifacts/goal/items/a3-stock/result.json](../../items/a3-stock/result.json), supplied as `.imd/reads/artifacts/item-a3-stock`. Input SHA-256 `15194b703c89bfe0b5d827be4cb4085dd96419fae397c520f7e065fd399489c2`. Strategy SHA-256 `efc27c4ca872b17a3aa78d4ca28e9cb1e5655daeb042a2f98c29abc3abd3ebc0`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 134.199329504; retail 160.585702702; signed arb -26.386373198; failures 0. Calibration mean: 108.64128535876738. - Pilot no_top (remove repeated above-probability cover top-ups; preserve near-crossing buys),… — m 0.46661, se 0.13928. - Pilot demand (cap planned cover and one-cent reserve jointly at estimated remaining retail… — m 0.35106, se 0.2012. - Pilot cross_demand (combine no_top with demand cap), first 240 spread-3 seeds in 30000-37999,… — m 0.77593, se 0.27236. - Pilot all_stock_cap (also cap ordinary bids above estimated probability by the same remaining… — m 0.77593, se 0.27236. - First-wave no_top validation, 729 previously unused matching seeds 30981-34028 — m -0.50707, se 0.55447. - First-wave cross_demand validation, 729 previously unused matching seeds 30981-34028 — m -1.4585, se 0.70184. - First-wave combined versus no_top validation — m -0.95138, se 0.3487. - Follow-up structural ideas preserve needed top-ups: urgent_cover buys above estimated…. - Fresh second-wave pilot urgent — m 0.42554, se 0.072836. - Fresh second-wave pilot low_tick — m 0.57529, se 0.10637. - Fresh second-wave urgent_pilot — m 0.42554, se 0.072836. - Fresh second-wave urgent_validation — m 0.64363, se 0.11624. - Fresh second-wave low_tick_pilot — m 0.57529, se 0.10637. - Fresh second-wave low_tick_validation — m 0.37717, se 0.24105. - Fresh second-wave low_tick_increment_pilot — m 0.14975, se 0.074548. - Fresh second-wave low_tick_increment_validation — m -0.26647, se 0.20869. - Second-wave pilot/validation are 240/729 matching spread-3 seeds in the unused second half of…. - The low-tick-boundary alternative passed as a standalone candidate but was not chosen: its…. Follow-up hypothesis: Investigate inventory-aware sizing of main bids below five cents, which left the most YES at expiry in FIFO measurements. Forecast the future ask-price path when estimating remaining share demand: a static current-ask cap saved arbitrage cost but lost more retail on separate validation. Preserve urgent cover below the lowest visible tick and use fresh paired validation for any new rule. ### a4-rich Source: [artifacts/goal/items/a4-rich/result.json](../../items/a4-rich/result.json), supplied as `.imd/reads/artifacts/item-a4-rich`. Input SHA-256 `be28481bad801e43bc8216b42bef60a70442b4365aa8252cb49d5208f253206a`. Strategy SHA-256 `41f2bb431d34c009ae17d19e19551fbce150cd1b23166374e759ed4f92f0a570`. Mechanism: see decision/synthesis and linked source. Reported holdout: edge 157.445255854; retail 183.254371664; signed arb -25.809115809; failures 0. Calibration mean: 108.63974289615767. - Remove above-probability ammunition refills: discovery diagnostic, n=309,… — m 0.1629. - Remove low-price bids when above existing 200-step stock target: discovery diagnostic, n=309,… — m -1.3409. - Shorten stock horizon to 1000: discovery diagnostic, n=309, mean=-0.07420303314110051,… — m -0.074203. - Shorten stock horizon to 500: discovery diagnostic, n=309, mean=-1.1020392822527494,… — m -1.102. - Restore existing kL on all low-price bids (same rule as submitted): discovery diagnostic, n=309,… — m 0.59163. - Numeric full-stock risk probe 0.3: discovery diagnostic, n=309, mean=0.49011199923042315,… — m 0.49011. - Use existing normal-price risk only above stock target: discovery diagnostic, n=309,… — m 0.35155. - Raise low-price ask cap to 12000: discovery diagnostic, n=309, mean=0.08961231701747568,… — m 0.089612. - Remove main bids below 5c: discovery diagnostic, n=309, mean=-3.082432580300646,… — m -3.0824. - Remove main bids below 2c: discovery diagnostic, n=309, mean=-0.2100393695731394,… — m -0.21004. - Limit ammunition overpayment to 0.005: discovery diagnostic, n=309, mean=0.009457995982201334,… — m 0.009458. - Limit tick-1 stock target to 10000: discovery diagnostic, n=309, mean=-0.74832610304466,… — m -0.74833. - Restoring existing kL on low-price bids, discovery [30000,34000) — m 0.5395, se 0.085973. - Restore kL only above existing stock target, discovery only — m 0.51433, se 0.076998. - Numeric horizon/cap/risk probes were rejected as a basis for a setting; no new numerical…. Follow-up hypothesis: Measure a structural stock-order cap based on already available expected retail demand, and distinguish ask-capacity misses from competitor ladder movement. Do not enlarge asks based on a static tail-fill estimate alone; the competitor refresh path changes. Confirm on fresh seeds before combining seats. ## What remains unproven The most useful next evaluation is a frozen, mutually guarded composition with one candidate per spread, starting with a1-free or a1-jumps, a2-after, a3-ledger, and a4-rich. Compare the spread-1 alternatives on a common untouched set before choosing if additional runs are available. Do not claim an independent holdout by reusing the already inspected 90000–97999 results to select new combinations. Same-spread combinations require new paired evaluation and separate attribution; equivalent ammo restrictions should be deduplicated first. This is a recommendation for subsequent work, not a claim that it was performed in this merge. The source records do not deliver the Paradigm 200 seeds, their candidate scores, candidate-specific board retail/arb decomposition, an executed combined strategy, or a reproducible archive of worker simulation logs. Consequently no exact board score, first-place claim, or independently validated combined gain can be supplied. The final result preserves the strongest measured mechanisms, the baseline-preserving failures, and the disagreements that constrain an entry decision.