The best supported selection is q1-ahead for spread 1, the unchanged base for spread 2, q3-ahead for spread 3, and q4-ahead for spread 4. These files report eligible improvements, but the implied combined gain does not close the stated gap to first place. Merge 02 supplies a qualifying spread-4 improvement that supersedes Merge 01's spread-4 base recommendation. Neither merge tested the combined strategy or established a new leaderboard score. M1, “Recommended stack”; M2, opening and “Comparable results and eligibility” M2

The requester's task brief reports that on 2026-11-06 the network is 3rd of 186 at 107.17, versus first at 109.59, on a challenge scored by mean edge over 200 simulations. It supplies the earlier measurements motivating this round: first place at the rival's price while quoting there was worth 47.86 more per run at every spread; perfect probability estimation was worth 1.9; first place while still quoting one tick inside was worth 1.98. Keeping small front positions previously gained 0.32, whereas resting all stock at tick 2 near 1 cent at spread 1 lost 40 more per run to arbitrage. These are requester-provided historical figures, not measurements reproduced by either merge or this write-up; no independent source for them or the dated leaderboard standings was supplied.

The evidence below is the two supplied merges: M1, artifacts/goal/merges/01/result.md, SHA-256 ac647882b4179e0ffc4931eb4647d2c36ed10360c784a578b13d929cba07d4ad; and M2, artifacts/goal/merges/02/result.md, SHA-256 07816380f9e8475eb3d64e24df77b1036efc815672d1df01e57a9a0906b5ef26. The supplied inputs matched those hashes on local inspection. M1 summarizes 20 item records; M2 summarizes the remaining spread-4 items and preserves their experiment ledger. Their measurements are worker reports, not independent replications. The original item records and seed-level outcomes are not available among these write-up inputs. M1, introduction and “Missing”; M2, “Evidence and provenance” and “Complete experiment ledger” M2

The public base JSON was retrieved for this write-up. Its strategy text hashes to dcf68e0c9a7b06f90cf70a833357b60e2208b5b3e2971e7dddf02986d4747076, matching M2's unchanged fallback. The base caps its ordinary size rule at 3000.0 shares and stays at the rival's tick only while tracked orders placed ahead there retain at least 2.0 shares; its manager can cancel an entire front order when the current size target shrinks. Public base M2, “What to retain, and what failed”

Results, one row per item. Mean difference and standard error refer to paired unseen runs of that item's spread against its base. Retail and arbitrage columns are the signed components of that difference; a positive arbitrage difference means less arbitrage loss. Seconds are base → delivered file, comparable only within that pair. “NR” means not reported in the supplied merge, not zero. M1's figures are copied at its supplied precision: its underlying unrounded item values could not be recovered. M2's full decimal strings are preserved. Independently rounded components need not sum exactly to the displayed mean. A returned-base row describes the delivered fallback, not a rejected experiment. M1, “Per-spread outcomes”; M2, both results tables M2

Item / part and source	Spread	Unseen mean difference	Standard error	Retail difference	Arbitrage difference	Failures	Other spreads identical?	Seconds, base → file	Disposition
q1-ahead — ahead	1	1.6416	0.2151	+1.604	+0.037	0	900/900	521.4 → 517.7	Select
q1-bids — bids	1	0.7918	0.0945	+0.781	+0.011	0	900/900	235.1 → 219.4	Eligible alternative
q1-life — life	1	0.4439	0.1489	+0.307	+0.137	0	300/300	181.3 → 180.3	Eligible; scorer check not run
q1-demand — demand	1	0.3543	0.1476	−0.417	+0.772	0	900/900	438.0 → 430.7	Eligible alternative
q1-keep — keep	1	0.3137	0.0742	+0.294	+0.020	0	900/900	69.3 → 68.6	Eligible alternative
q1-lead — lead	1	0.1238	0.3596	+0.780	−0.656	0	900/900	86.1 → 85.7	Gain unresolved; M1 says not significant
q2-ahead — ahead	2	0	0	NR	NR	0	900/900	NR	Returned base
q2-bids — bids	2	0.1485	0.3154	+0.917	−0.768	0	900/900	163.4 → 167.7	Exclude: slower; gain unresolved
q2-life — life	2	0	0	NR	NR	0	900/900	NR	Returned base
q2-demand — demand	2	0	0	NR	NR	0	900/900	NR	Returned base; experiment slower
q2-keep — keep	2	0	0	NR	NR	0	900/900	NR	Returned base
q2-lead — lead	2	0	0	NR	NR	0	900/900	NR	Returned base
q3-ahead — ahead	3	0.1361	0.0407	+0.189	−0.053	0	900/900	114.6 → 113.9	Select
q3-bids — bids	3	0.1699	0.2411	+1.061	−0.892	0	900/900	290.4 → 305.6	Exclude: slower; gain unresolved
q3-life — life	3	0	0	NR	NR	0	NR	NR	Returned base
q3-demand — demand	3	1.8604	0.6402	−1.146	+3.006	0	900/900	265.5 → 302.2	Exclude: slower
q3-keep — keep	3	0	0	NR	NR	0	NR; only 180 runs checked	NR	Returned base; scorer discrepancy
q3-lead — lead	3	0	0	NR	NR	0	NR	NR	Returned base
q4-ahead — ahead	4	0.17255959975495247	0.05609359681034755	0.2216695056799397	-0.0491099059248383	0	900/900	294.15425222413614 → 291.83430307451636	Select
q4-bids — bids	4	0.051792014328919854	0.012644173743880268	0.09115884220389603	-0.0393668278740158	0	900/900	67.5692147470545 → 66.84266982902773	Eligible alternative
q4-life — life	4	0.0	0.0	0.0	0.0	0	900/900	595.4729134468362 → 575.7384612760507	Returned base; timing variation
q4-demand — demand	4	0.5909248230146836	0.19743783993870737	-1.2474939821301518	1.8384188051388293	0	900/900	144.19885997008532 → 145.2658611133229	Exclude: slower
q4-keep — keep	4	0	0	NR	NR	0	900/900	NR	Returned base
q4-lead — lead	4	0	0	NR	NR	0	900/900	NR	Base behavior; trailing newline differs

The part labels retain the items' names: ahead concerns replenishment priority; bids concerns bid policy; life concerns lifetime valuation/sizing; demand concerns inferred retail flow; keep concerns retaining existing orders; lead concerns waiting to acquire priority. The table is a selection among delivered files, so its zero-gain fallbacks must not be replaced with their experimental gains. All rows report zero delivered-file failures, but the finite other-spread checks establish identity only on their tested runs. M1 does not give identity results for q3-lead or q3-life, or an explicit identity count for q3-keep's 180 checked runs. M1, “Per-spread outcomes” and “Findings that hold”; M2, “Comparable results and eligibility” M2

How the selected files keep the queue lead. At a fixed tick, the simulator sorts by placement sequence. Depletion schedules the rival's replacement on the opposite side, separated by its spread; the replacement receives a new sequence. An order already resting at that destination can therefore precede it. Arbitrage consumes the entire mispriced quantity, making a larger front order vulnerable to a price jump. Those mechanics support preserving useful earlier orders while controlling exposure. Simulator, _best_order, _schedule_competitor_replenishment, _create_competitor_order, execute_arbitrage

Spread 1: select the q1-ahead item file, identified by strategy SHA-256 d5a76ce3ac426619d1e4fe4e3864afb7b88acbea5b4d63b35f59c09ae3dff3bb. M1 attributes its improvement to front-order retention, specifically ask-side retention in its proposed combination with q1-bids, and raising the below-1-cent cap from 3000 to 15000 while retaining the risk-based size rule. Keeping a resting front ask through a shrinking one-step target preserves its earlier sequence when the rival replenishes at that tick; relaxing the cap permits more quantity only where the risk rule allows it. Its unseen gain is predominantly retail, as the table shows. M1 reports independent tuning validation of +1.307 on the next 640 tuning seeds and +2.018 on the remaining 1236, then holdout mean 43.4265 against 41.7849 over all 1969 spread-1 seeds in 60000–67999. This is the best reported eligible spread-1 mean, not proof that it beats each alternative in a direct paired comparison. M1 identifies the source as .imd/reads/artifacts/item-q1-ahead; it does not preserve an original delivery pathname, exact retention thresholds, or the complete strategy. Those details cannot be specified from these inputs. M1, introduction, code hashes, “Recommended stack”, and findings 1 and 3

This is consistent with the requester's boundary warning: below 1 cent the rival's lowest ask is at tick 1 + spread. At spread 1, selling at tick 1 after buying at tick 1 nets nothing and diverts the retail that could deplete the rival at tick 2. The useful seat is at the replenishment price; it must still be sized for distance from the true price. That warning and the rival lots' 24–72 dollar size are supplied by the task brief, rather than newly measured here.

Spread 2: retain the public base, strategy SHA-256 dcf68e0c9a7b06f90cf70a833357b60e2208b5b3e2971e7dddf02986d4747076. No delivered improvement qualifies. It continues quoting inside until an order previously placed ahead at the rival's tick retains at least 2.0 shares, then uses that existing queue lead; it supplies no new method of winning first place. q2-bids is slower and its uncertainty exceeds its mean. q2-demand's rejected experimental policy gained +0.408 with SE 0.212 on 1987 unseen seeds, but took 151.39 seconds against 149.99 for the base; its delivered result is the base, as recorded in the table. The experimental retail/arbitrage split is not supplied. M1 reports that q2-ahead's retention plus larger sizing was consumed by arbitrage and q2-life's whole-lot valuation failed its reserved gate. Public base, top and order manager M1, “Per-spread outcomes” and findings 1 and 5

Spread 3: select the q3-ahead item file, strategy SHA-256 a0c13edd9e6917d067d6fc10c1a49aa45c2f091e8e8fa9e1ebc101c65256427d. Its reported change splits the order at the rival ask tick minus 3, the future rival bid destination, into 25%/75% FIFO tranches. Separate tranches allow a later piece to be cancelled when the size target shrinks while an earlier piece retains priority over the replacement rival lot. This queue explanation follows from the reported split and the simulator's FIFO mechanics; it was not independently replayed. M1 identifies the source as .imd/reads/artifacts/item-q3-ahead but gives no original delivery pathname, further placement guards or minimum sizes. The eligible gain is the table's 0.1361 with SE 0.0407, mostly additional retail edge with some additional arbitrage loss. q3-demand's larger gain fails the speed condition, and q3-bids is both slower and unresolved. M1, introduction, code hashes, “Per-spread outcomes”, and finding 2 Simulator

Spread 4: select artifacts/goal/items/q4-ahead/result.json, strategy SHA-256 04ef81f760b2ae1c3802ba120998ae4621d8e9c8e04478395671ecbebd36ef3d. It splits a new destination bid into two equal, cent-quantized tranches only when a real rival ask exists at tick at most 99, the new bid is above the rival bid and exactly the rival ask tick minus 4, quantity is at least four shares, and total notional is at least $20. Aggregate quantity and its sizing rule stay unchanged. When the target shrinks, the manager can cancel the tail while keeping the earlier tranche ahead of the replenishing rival. This preserves a seat without broadly increasing inventory. The reported diagnostic raises front-bid step share from 0.166671875 to 0.174 while average reserved cash falls from 22.336927876562495 to 22.23691695. Independent tuning validation was 0.16580731308913887 over 488 runs in 14000..15999, before freezing for unseen scoring. All four M2 items used 2009 spread-4 seeds in 60000..67999 and the same unseen base mean, 163.47813945343543. M2, “What to retain, and what failed”, selected strategy, and q4-ahead trials 74, 86–89

q4-bids is a smaller eligible alternative: it keeps an existing front bid using a keep target multiplied by four, capped at 3000 shares, only with estimated probability at least two ticks higher; it never adds behind the rival. No supplied seed-level candidate-to-candidate differences establish that q4-ahead's advantage over q4-bids is statistically significant. q4-demand reduces ordinary sizes after >=64 observations when fitted BUY flow is <0.45 dollars/step, sacrificing retail edge to avoid arbitrage loss, but remains 0.74% slower. q4-life's experimental policy lost 3.3483234647361373, SE 1.510332187739084, over 2009 unseen runs, with retail +2.2084078648600793 and arbitrage -5.55673132959786. Restoring original arithmetic reproduced those results; the zero-gain row is its unchanged-base fallback. Its validation range 9000..10999 overlapped tuning 8000..15999. M2, “Comparable results and eligibility”, “What to retain, and what failed”, and q4-life trials 45–48

Combined gain, calculated from the supplied figures. The brief assigns each spread a quarter of runs. Selecting only one policy per spread gives

(1.6416 × 1/4) + (0 × 1/4) + (0.1361 × 1/4) + (0.17255959975495247 × 1/4) = 0.4875648999387381175 additional mean edge over all runs.

This is exact decimal arithmetic on M1's already rounded selected means and M2's supplied mean, not additional measurement precision. M2 gives its spread-4 contribution rounded as 0.04313989993873812. M1's approximately 0.44 stack omitted the subsequently supplied q4-ahead improvement. The updated expectation remains below the 2.42 gap between the requester's 107.17 and 109.59 scores. It does not establish a score improvement over the network's current submission: M1 explicitly distinguishes the experimental base from the 107.17 file and says the demand ports may already be in that file. M1, “Recommended stack” and finding 5; M2, opening; task brief for weights and standings M2

Per-spread gains add because each selected policy is confined to a different spread, provided a combined implementation preserves each source policy's behavior. Changes competing for the same spread's orders cannot be summed. No combined file was evaluated, and no combined timing or uncertainty estimate is supplied. The finite other-spread checks and paired timing samples support selection but do not guarantee equality or speed on every run. M2 reports alternating-order, warmed-up timing on 200 target runs for q4-ahead; q4-bids has no reported timing protocol; q4-life's apparent speed improvement is variation in identical code. M1, “Recommended stack”; M2, timing table and accompanying discussion M2

The evidence has additional limits that affect interpretation. M1 reports the reference base scorer result as 98.4694596619638, reproduced by the selected q1-ahead and q3-ahead items, among others; q1-life and q3-demand did not run that check. q3-keep instead reports 114.49265606862934 and an unseen base 115.98 versus approximately 118.82 in other spread-3 items. Its returned-base result does not improve the selection, but its experiments should not be pooled uncritically. M1's spread-4 base means also differ: 159.705 for q4-keep versus 163.478 for q4-lead, with seed-set/count differences unresolved there. M2 gives its own common baseline precisely above. M1, “Scorer check”; M2, “Comparable results and eligibility” M2

M2's q4-ahead, q4-bids and q4-life report reproducing the 4000-run reference result 98.4694596619638 on seeds 4000..7999. q4-demand's 113.9414113871885 comes from a different 200-run sample starting at 4000; it explicitly did not reproduce the reference check. Its scoring used engine-loop patches whose equality with the unpatched engine was checked on only eight seeds; its final timing used the unpatched engine and did not time other spreads. These are additional comparability limits beyond its disqualifying slowdown. M1 also records that q4-lead differs from the base by one trailing newline: equivalent behavior does not mean byte-identical text. M1, anomaly; M2, “Scorer discrepancy” and timing discussion M2

M1 says its item records contain no source URLs, and flags source recheck failures for q1-lead, q2-life, q3-ahead, q4-keep, q1-ahead, q1-keep, q1-demand, q2-lead, q2-keep, q3-lead and q3-life. M2 flags unidentified external sources behind q4-life and q4-bids. Neither merge can identify the disputed source passages. Those citations cannot be repaired from the supplied materials. Their numbers remain attributed worker measurements, and their causal explanations remain reported interpretations. Reading the public simulator for this write-up corroborates the queue mechanics described above; it does not independently validate any item's gains. M1, “Sources”; M2, “Evidence and provenance” M2

What the seats say to try next:

• Spread 1: test q1-ahead's ask retention together with q1-bids' risk-limited tick-1 bid banking; bound dormant fragments or require a minimum retained lot. The two gains are not additive without testing. q1-life proposes fill-versus-crossing valuation; q1-lead proposes jump-aware lead sizing; q1-demand proposes separating fitted bid and ask sizing. M1 says seeds 60000–67999 are now consumed for spread 1, so use fresh unseen seeds. M1, “Recommended stack” and “Next steps proposed by items”
• Spread 2: make demand inference fast enough, with the proposed 3×3 grid or deferred refreshes; condition extra size on imminent rival depletion and jump-crossing risk, bounded by the rival lot's 24–72 dollar value. q2-life proposes compound-Poisson queue-tail survival and q2-lead rate-gated waiting. M1's front-size tolerance lead, +0.29 on only 160 runs, remains unconfirmed. M1, finding 2 and “Next steps proposed by items”
• Spread 3: optimize q3-demand to pass the speed condition, restrict posterior sizing near the price, and test replenishment prediction across multiple lots. Compare a retained floor-sized tranche with a separately cancellable remainder and value the incremental tick advantage. q3-keep's negative tuning retention results and q3-bids' noisy positive unseen result disagree; they do not support blanket retention. M1, findings 1 and 5 and “Next steps proposed by items”
• Spread 4: try ask-side tranches after removing their overhead; avoid splitting stock bids likely to be swept. Infer the remaining rival ask notional before guarded bid enlargement. Test a conservative jump-loss mixture bound while retaining the ordinary one-step core, and a tighter low-flow cutoff or posterior-probability gate that retains more retail. Shared posterior tables are a proposed speed repair, not a measured improvement. Broad enlargement, artificial source lots at tick 0/100, and lifetime extension alone did not validate in the tested variants. Any hybrid needs fresh seeds outside consumed tuning/unseen ranges, identity checks against each source file, and paired speed measurements. M2, “What to retain, and what failed” and “Follow-up carried forward”; M1, spread-4 next steps M1