# Merge 2: twenty prediction-market items

The strongest reported candidate on the common holdout is **simplify**, followed by **arb-budget** and **cross-anywhere**; the first two converge on reducing secondary inventory purchases, whereas cross-anywhere adds demand-matched acquisition at higher ticks. These are reported local results, not evidence that any candidate beats the public leader on unseen Paradigm runs, and their gains must not be added together. [simplify] [arb-budget] [cross-anywhere]

Every item is represented below. Citations identify the supplied canonical item artifact; the exact input copies were read from `.imd/reads/artifacts/item-KEY`, and their hashes are recorded at the end. All experiment claims are attributed to those item reports: this merge did not rerun simulations, retrieve external sources, or independently certify strategy behavior. Lane labels summarize the reported work; no separate per-seat direction manifest was supplied.

Numeric cells preserve the item JSON's numeric spelling without rounding. “Base” is the reported reproduction on seeds 0–599; the next three columns are the delivered strategy's seeds 1000–1599 score, retail edge and signed arbitrage edge, except for explicitly flagged reporting defects. Holdouts are seeds 30000–30599 unless otherwise stated; reported completed standard ranges contain 600 simulations with zero ordinary failures. `Unmeasured` means a missing measurement, never zero edge. Code size is computed from the UTF-8 bytes of the supplied `code` string, including unchanged base code and empty blocked submissions; all supplied code hashes matched those strings.

| Item / source | Lane and direction | Base reproduced | 1000–1599 mean | Retail edge | Arbitrage edge | Holdout mean | Python | Code bytes |
|---|---|---:|---:|---:|---:|---:|---|---:|
| [drift-loss] | Deepen: drift margins and resizing | 102.3319790889927 | 107.2457521863612 | 135.80165652288818 | -28.555904336541 | 96.00139032710622 | 3.10.12 | 13882 |
| [cross-anywhere] | Deepen: crossing stock beyond tick one | 102.33197908899267 | 108.63974289615767 | 136.2713209771645 | -27.63157808101933 | 97.32763650112 | 3.12.3 | 14239 |
| [arb-budget] | Deepen: allocate arbitrage-loss budget | 102.33197908899267 | 108.34220065942017 | 131.02646185504466 | -22.684261195613665 | 97.47824853761233 | 3.12.3 | 13941 |
| [fresh-solve] | Fresh: solve (blocked before development) | Unmeasured | Unmeasured | Unmeasured | Unmeasured | Unmeasured | Unmeasured | 0 |
| [fresh-rules-lawyer] | Fresh: fill rules → retail intensity | 102.33197908899267 | 108.15649684917884 | 134.35938878886 | -26.202891939700333 | 96.17078093861683 | 3.12.3 | 15999 |
| [fresh-gifts] | Fresh: FIFO order chunks | 102.33197908899267 | 108.11869389741317 | 136.6937171020847 | -28.575023204701665 | 96.07713311486467 | 3.13.5 | 15225 |
| [fresh-competitor] | Fresh: competitor replacement signals | 102.33197908899267 | 102.33197908899267 (unsubstantiated range) | 127.3199510028995 | -24.987971913912165 | 96.69576967374634 (unsubstantiated range) | 3.12.3 | 14311 |
| [fresh-tail] | Fresh: large retail tails | 102.33197908899267 | Unmeasured; 102.33197908899267 on 0–599 | 127.3199510028995 on 0–599 | -24.987971913912165 on 0–599 | 96.69576967374634 on 20000–20599 | 3.12.3 | 13882 |
| [fresh-time] | Fresh: acquired FIFO priority | 102.33197908899267 | 107.2752344684335 | 135.8212954705185 | -28.54606100209883 | 96.06785581906733 | 3.12.3 | 14313 |
| [fresh-literal] | Fresh: aged same-price queues | 102.33197908899267 | 107.2442092237515 | 135.80165652288784 | -28.557447299150834 | 96.0405203261775 | 3.12.3 | 13882 |
| [fresh-fingerprints] | Fresh: Bayesian demand → stock | 102.33197908899267 | 107.7367374084125 | 135.91459094566451 | -28.177853537266333 | 96.97549954237483 | 3.12.3 | 18314 |
| [fresh-opposite] | Fresh: one-sided asks, no YES buys | 102.33197908899267 | 107.2442092237515 | 135.80165652288784 | -28.557447299150834 | 96.0405203261775 | 3.12.3 | 13882 |
| [fresh-free] | Fresh: queue search; final constant retune | 102.33197908899267 | 109.02047643278917 | 136.12585295838298 | -27.105376525607166 | 96.3941526043735 | 3.12.3 | 14062 |
| [no-probes] | Take away: remove probes | 102.33197908899267 | 97.92400676829367 | 130.11073730696 | -32.186730538662005 | 87.753135063419 | 3.12.3 | 16861 |
| [fewer-cancels] | Take away: retain orders longer | 102.33197908899267 | 108.6925430884395 | 138.4792002400675 | -29.78665715164867 | 96.094069027841 | 3.12.3 | 16623 |
| [simplify] | Take away: remove secondary stock logic | 102.33197908899267 | 109.05194228328166 | 131.3166036311765 | -22.2646613479065 | 97.70763523638367 | 3.12.3 | 9972 |
| [confidence-gate] | Deepen: pause after belief shocks | 102.33197908899267 | 107.45298269594832 | 135.54694126664617 | -28.0939585707135 | 95.74596803103867 | 3.13.5 | 14473 |
| [rich-runs] | Measure: richest-run ask capacity | 102.33197908899267 | 107.74822700668217 | 136.403923077077 | -28.655696070412002 | 95.50710368487617 | 3.12.3 | 14002 |
| [wildcard-1] | Wildcard: online retail-flow estimation | 102.33197908899267 | 108.54841494177967 | 136.43005690744866 | -27.881641965698503 | 96.7683741257685 | 3.12.3 | 18417 |
| [wildcard-2] | Wildcard: arbitrage mechanisms (blocked) | Unmeasured | Unmeasured | Unmeasured | Unmeasured | Unmeasured | 3.12.3 | 0 |

**drift-loss.** Tightening resize bands and raising the margin multiplier improved tuning by +0.41, but the combined variant's holdout was 94.85 versus base 96.00, so it returned the **base unchanged**; the table deliberately preserves its slightly different reported baseline precision rather than replacing it with another seat's values. Its suggested next experiment is horizon- and distance-dependent exposure sizing, because its engine reading says arbitrage fills mispriced resting size immediately rather than rewarding order age. [drift-loss: `approach`, `tried`, `next`][drift-loss]

**cross-anywhere.** It adds crossing bids at ticks 2–9 sized to the ordinary ask inventory shortfall, with a 30% wealth cap and 3000-share cap, preserving the tick-one reservoir; fixed higher-tick stock and broad surplus asks were less successful in its tuning experiments. The reported frozen holdout improves on its matched base, and it proposes using desired ask demand for eager purchases too, then separating ordinary crossing costs from jump-through costs on fresh seeds. The brief flags **one failed source** for this item, without identifying it in the supplied JSON, so its external simulator rationale remains unverified; the table reports its own experiment record rather than treating that failed citation as corroboration. [cross-anywhere: `approach`, `tried[47:49]`, `next`][cross-anywhere]

**arb-budget.** Removing secondary stock, raising surplus low-price bid aversion and doubling ask loss aversion above 5% reduces tuning arbitrage exposure, but its joint target of 18 loss and above-127 retail is explicitly **unmet** on both ranges: holdout retail is 119.517920457476 and arbitrage edge is -22.0396719198325. Its exact tagged tuning attribution separates main-order diffusion loss -7.22020612443645, main-order jump loss -8.434846762625929 and tick-one stock loss -6.5468483987207335, rather than conflating a forecast loss budget with guaranteed realized loss control. Next it proposes an emergency below-one-cent stock fallback and occupation-time stock sizing, with fresh validation before accepting that graft. [arb-budget: `approach`, final attribution and frozen-policy entries in `tried`, `next`][arb-budget]

**fresh-solve.** **Blocked** before development or calibration because its reported Python invocation failed with an `xcrun` developer-tools error; no simulations ran and no strategy was supplied. Its next step is a working Python 3.10 or newer runtime followed by scorer verification; all scores and the Python version are unmeasured, and the zero-byte code field is not a trading policy. [fresh-solve: `blocked`, `score`, `holdout`, `next`][fresh-solve]

**fresh-rules-lawyer.** It uses partial fills and selected full fills to estimate retail intensity online, shrinking and clipping the estimate before changing marginal sizing, reporting +0.91 on tuning and +0.13 on holdout. Its claimed +1.1 oracle ceiling is specific to its retail-regime experiment, not a universal bound on learning, and a full fill classified using estimated probability can still be uncertain even if partial fills are informative. It proposes combining retail-flow estimation with collateral budgeting and using full main-order fills as crossing evidence, while auditing jump losses on queue-ahead orders. [fresh-rules-lawyer: `approach`, `tried`, `next`][fresh-rules-lawyer]

**fresh-gifts.** It splits named main orders of at least 32 shares into sixteen equal rounded children so shrinking targets can preserve older FIFO pieces; its paired holdout gain is 0.036612788687167254 with standard error 0.2817268065582916, which does not establish an improvement. It recommends combining chunking, or the bid-only alternative, with an inventory-risk policy on new seeds rather than tuning the exposed holdout again. The brief flags **one failed source**, not identified in this JSON, so FIFO rule claims are retained as the item's interpretation and local experiments rather than externally verified facts. [fresh-gifts: `approach`, frozen training/holdout and selection entries in `tried`, `next`][fresh-gifts]

**fresh-competitor.** It adds upward-best-bid and downward-best-ask replacement signatures to belief bounds, but its only documented graft evaluation is a two-seed smoke test that matched the base. Although the JSON labels 102.33197908899267 as the 1000–1599 mean and 96.69576967374634 as the 30000–30599 holdout, the former exactly duplicates its documented 0–599 calibration and its log supplies no matching full tuning or holdout run, so these are **unsubstantiated range labels**, not a validated gain. It suggests measuring replacement transitions by spread and testing whether the signal should affect size instead of only interval clamps. [fresh-competitor: `score`, `holdout`, `tried[0:2]`, `next`][fresh-competitor]

**fresh-tail.** It returned the **base unchanged** after large-tail thresholds and cap variants failed to finish aggregate evaluation; its one completed threshold seed is not a full candidate score. Its score is explicitly on seeds 0–599 and its holdout on 20000–20599, so the requested 1000–1599 and 30000–30599 measurements are missing and must not enter the common-range ranking. It proposes completing the generated tail-threshold comparisons on the requested ranges before considering a graft. [fresh-tail: `approach`, `score`, `holdout`, `tried`, `next`][fresh-tail]

**fresh-time.** It marks surviving economic orders as queue-ahead when competitor movement gives them that position after submission; the small patch yields a reported paired holdout gain of 0.02733549288983328, while broader queue joining and stock retention underperform. It proposes combining estimated remaining competitor depth with seniority, retail waiting time and collateral opportunity cost, requiring independent validation because the selected training gain affects very few paths. The brief flags **one failed source**, not identified in this JSON: its replay/control measurements are report-level evidence, while hidden-book/FIFO explanations and oracle relaxations are not independently verified global policy bounds. [fresh-time: `approach`, `tried[13:27]`, `tried[50:58]`, `next`][fresh-time]

**fresh-literal.** It tried future-informed oracle comparisons, standalone aged queues and age-gated grafts, then returned the **base unchanged** after comparing valid holdout scores; the age-conditioned candidate's holdout 95.89964096688516 was below base 96.0405203261775, and uncovered secondary asks produced collateral failures. That explicit holdout-based choice makes this a fallback baseline rather than a newly frozen pre-holdout winner, and the next suggestion is to estimate competitor queue depth and add secondary orders only when expected fill edge covers collateral cost. The brief flags **one failed source**, unidentified in the JSON, so the source-dependent age/queue narrative is unverified and future-informed scratch results are diagnostics, not attainable strategy returns. [fresh-literal: `approach`, `tried[20:32]`, `next`][fresh-literal]

**fresh-fingerprints.** It grafts a censored Bayesian retail-demand learner onto inventory targets; selected tuning paired improvement is 0.4925281846609999 with standard error 0.4342523383708021, and it reports the table's frozen holdout result. It proposes stock targets that distinguish cheap crossing acquisitions from refills and respect the one-step delay before newly purchased YES can cover asks. The brief flags **two failed sources**, unidentified in the JSON, so public leaderboard/code-size fingerprints cannot establish another entrant's mechanism, while its oracle, estimator-MSE and planning-model figures remain distinct from executable challenge scores and have not been independently reproduced. [fresh-fingerprints: `approach`, public fingerprint and selected diagnostic entries in `tried`, `next`][fresh-fingerprints]

**fresh-opposite.** The independent one-sided, no-YES-buying strategy reached 6.829995035887166 on full training, but discovered queue retention already present in the base; late native grafts did not complete their full panels and it returned the **base unchanged**. It excludes interrupted prefixes and accounting-incompatible wrappers from candidate selection, and proposes a small separately accounted ask rung at the competitor price, retaining covered inventory and validating on a new holdout. [fresh-opposite: `approach`, final native-diagnostic and selection entries in `tried`, `next`][fresh-opposite]

**fresh-free.** The independent search rediscovered queue-front retention already in the base, so the final submission changes only ydem, qmax and min_ahead, reporting +1.78 on tuning and paired holdout gain 0.35 with standard error 0.72. It advises against rebuilding existing queue retention or belief machinery and instead investigating depth below tick one and next-step risk, where its true-information diagnostics suggested room but its estimated mixture implementation lost money. [fresh-free: `approach`, `tried`, `next`][fresh-free]

**no-probes.** Removing probes and replacing their information with soft own-fill/competitor evidence, a variance-matched belief walk and mixture-based loss sizing still loses paired -8.29 +-0.76 on holdout, so the submitted code is explicitly a **measurement, not a replacement candidate**. It measured direct probe costs at arbitrage -0.4947 and retail +0.0262 per run on seeds 1000–1095, and proposes grafting its soft evidence and belief updates back onto the base with probes retained, especially where one-tick spreads prevent probing. The brief flags **two failed sources**, unidentified in this JSON, so exact rule claims and its inferred value of information remain item-reported conclusions; the measurement also bundles several compensating changes and does not isolate the causal value of probes alone. [no-probes: `approach`, `tried[3]`, `tried[10:17]`, `next`][no-probes]

**fewer-cancels.** It preserves probes and stock orders, tolerates more size on queue-ahead orders and retains some one-tick-behind quotes, reducing reported cancellations from 2973 to 1006 a run. Tuning improves, but holdout gain is only +0.05 +-0.72 as extra retail is offset by arbitrage losses; it recommends tracing rare jump losses before timing the queue-release rule, testing its safer maf=0 subset independently, and using fill quantities to disambiguate retail-eaten probes. [fewer-cancels: `approach`, `tried`, `next`][fewer-cancels]

**simplify.** It removes secondary inventory purchases and dead controls, keeps probes/belief/queue logic, changes stock eligibility to competitor bid at most one tick, raises general risk to 0.9 and removes the separate stock-cash cap; the supplied strategy shrinks from 13882 to 9972 bytes. Its frozen common holdout is the highest in this group, with paired gain 1.6671149102061666 and standard error 1.1776597053936184 in the experiment log (the approach separately spells the difference 1.6671149102061662), and it proposes grafting the purchase removal and stock-entry gate onto independent strategies using fresh seeds. The brief flags **two failed sources**, unidentified in this JSON, so simulator rationale and cleanup-equivalence checks remain attributed local claims, not independently verified source support or proof of a population gain. [simplify: `approach`, final cleanup/holdout/selection entries in `tried`, `next`][simplify]

**confidence-gate.** After an unusually large belief move it cancels economic orders and pauses purchases while keeping information probes, but its frozen shock gate's small tuning gain fails to generalize: reported paired holdout delta is -0.29455229513883385 with standard error 0.4184176557036628. It proposes separate acquisition and quote gates and confidence calibration with probe coverage, because lost inventory purchases can remove future retail edge; shadow associations on base paths are not causal savings. [confidence-gate: `approach`, frozen control/shadow/selection entries in `tried`, `next`][confidence-gate]

**rich-runs.** It traces the richest tuning paths and selectively raises low-price ask capacity in narrow spreads, retaining base stock and collateral rules, but its frozen holdout falls below the matched base despite higher tuning edge. Its next direction is safer acquisition before the one-cent crossing, since the richest 60 tuning runs already captured 91%-97% of below-one-cent retail buy notional and wider ask caps added jump risk. [rich-runs: `approach`, `tried`, `next`][rich-runs]

**wildcard-1.** It estimates retail arrival rate and censored notional size online, recalibrates marginal sizing and scales only the slow stock target by the square root of inferred demand, reporting holdout +0.73 over its rounded base control. It notes that early stock purchases occur before learning converges and spread-one books provide no front-quote exposure, then proposes learning from tick-one/competitor-tick fills, confidence-aware stock scaling and a separate jump-intensity estimate. Its oracle experiments use true regime information and are research diagnostics rather than available strategy inputs. [wildcard-1: `approach`, `tried`, `next`][wildcard-1]

**wildcard-2.** **Blocked** because its base/simulator downloads failed with reported HTTPS proxy timeouts and direct DNS resolution failure; no reproduction, development, scoring or sandbox checks ran. It supplies empty code rather than a candidate, and recommends retrying input acquisition in a runtime with working network access before reproducing the base and investigating arbitrage-loss mechanisms. [wildcard-2: `blocked`, `tried`, `next`][wildcard-2]

The merge's next-step recommendation is to validate secondary-purchase removal and the stock-entry gate first, then test demand learning and demand-matched crossing cover as separate incremental grafts before combining them; these mechanisms overlap in inventory sizing, so improvements may cancel rather than accumulate. This recommendation follows the reported common-holdout ordering and the items' next experiments, not a new measured combined strategy. Keep probes, investigate queue-tail jump exposure before further retention, and treat holdout-selected fallbacks and inconsistent range labels separately from candidates frozen before evaluation. [simplify] [arb-budget] [cross-anywhere] [fresh-fingerprints] [wildcard-1] [no-probes] [fewer-cancels] [fresh-literal] [fresh-competitor] [fresh-tail]

The seven flagged items contain no external source URLs or failed passages in their non-code report fields, so the precise failed citations could not be identified or rechecked from these inputs. Their reported experiments are preserved with the caveats above; none of those failed external citations is used as independent support. Several items also report the same sandbox `ntpath` import failure for both candidate and starter, so ordinary zero-failure scores must not be read as successful upstream sandbox validation; scratch import-hook or allow-list workarounds reported by some seats do not repair that upstream check. [cross-anywhere] [fresh-gifts] [fresh-time] [fresh-literal] [fresh-fingerprints] [no-probes] [simplify]

All twenty input files were present and parseable. Missing measurements, incomplete experiments, unsupported range labels, blocked runtimes and source failures are stated explicitly; no missing score, strategy or combined-performance result has been fabricated. The following source index records the supplied artifact paths and SHA-256 of the exact input bytes, allowing the reported claims to be traced even if the workspace mirrors are unavailable later.

[drift-loss]: ../../items/drift-loss/result.json "Supplied item-drift-loss; SHA-256 33c4c3bf966a2e88d82d9f14066361217c6ce71ad3871fac7de7cc9c9c3391cb"
[cross-anywhere]: ../../items/cross-anywhere/result.json "Supplied item-cross-anywhere; SHA-256 40497298132dcc0ccc9f4c0672031c7eaae00ddcf8cf24983db989feeb3c5ca0"
[arb-budget]: ../../items/arb-budget/result.json "Supplied item-arb-budget; SHA-256 a799a7073e05509473d99a982226006f4ebce8fdcc04d75b44cae846a158142b"
[fresh-solve]: ../../items/fresh-solve/result.json "Supplied item-fresh-solve; SHA-256 91830d5506175c497b450e2312d85eeb5fdf6068c862147f8d81dbf0e4cad64b"
[fresh-rules-lawyer]: ../../items/fresh-rules-lawyer/result.json "Supplied item-fresh-rules-lawyer; SHA-256 40c239f85ba43aa49deba0d07dce917d3c2fbbfb6edd33ca4b249ce934333b8c"
[fresh-gifts]: ../../items/fresh-gifts/result.json "Supplied item-fresh-gifts; SHA-256 990a9ddb31081fd4aff835267ba650ffde0bfaf6760979535eb41a929a6a4f1f"
[fresh-competitor]: ../../items/fresh-competitor/result.json "Supplied item-fresh-competitor; SHA-256 29016a3307cc84196469204f2c7285ecf6f3b52928af92111e3a3431ae616a5f"
[fresh-tail]: ../../items/fresh-tail/result.json "Supplied item-fresh-tail; SHA-256 36b3590fe5c70f892a7431649cc009a4a546ccadc097680c74def792983af1b4"
[fresh-time]: ../../items/fresh-time/result.json "Supplied item-fresh-time; SHA-256 1cc31bd5ca2f82cc504b446355702cf1686f3e181dbdecf476ee7b8779aa973a"
[fresh-literal]: ../../items/fresh-literal/result.json "Supplied item-fresh-literal; SHA-256 b0d67f430e1dee4031fc35f51057859487f55bc8e11464309dc8f7baaf690127"
[fresh-fingerprints]: ../../items/fresh-fingerprints/result.json "Supplied item-fresh-fingerprints; SHA-256 a9859cdd14758f764746e9453a1d5ff6c00b8c2f6df95bce50be0ea5d22b1958"
[fresh-opposite]: ../../items/fresh-opposite/result.json "Supplied item-fresh-opposite; SHA-256 f700666a653fcacd9f8a19ba11e647b308bf3900261de4623e6dde90e6585725"
[fresh-free]: ../../items/fresh-free/result.json "Supplied item-fresh-free; SHA-256 13f5174f03ec96843b1a2f13e338feac0a9798b92f9f7cc0fd29e6f9a4b3632a"
[no-probes]: ../../items/no-probes/result.json "Supplied item-no-probes; SHA-256 6a68b8124d25eb11d372af31de3b249f13c0fe4adc1e1b28b1d95ae32447123c"
[fewer-cancels]: ../../items/fewer-cancels/result.json "Supplied item-fewer-cancels; SHA-256 bc63ccb12d84e74533d37109e09f76ebd1e74d06d53b2419859bc80e34d55249"
[simplify]: ../../items/simplify/result.json "Supplied item-simplify; SHA-256 5aa77c807417d8eb4743635be24fb9590f57e3e636dfb11050dd283e0a14ad6d"
[confidence-gate]: ../../items/confidence-gate/result.json "Supplied item-confidence-gate; SHA-256 3fd60229ef5eb8be8826cfc8aa241073d7631388c3cdb1083eff690ac04f2b91"
[rich-runs]: ../../items/rich-runs/result.json "Supplied item-rich-runs; SHA-256 26d3bdf30a83c5bea0515beb930bf0aead7c0afcc0ca17b90a705b6d46c73687"
[wildcard-1]: ../../items/wildcard-1/result.json "Supplied item-wildcard-1; SHA-256 b142fe64c73443cbb88ef4f6be8167329d48670e2d5022db9ad9ca6d7dda70ac"
[wildcard-2]: ../../items/wildcard-2/result.json "Supplied item-wildcard-2; SHA-256 0d4bf19962645e427983990babce0a00a7d77cf6ef0896d4836f2152db0467b3"