Four-change stack: entry recommendation and evidence

Enter stack-b first, identified by its strategy-code SHA-256 1d2db2a895bf0f9e427f6f15f6be20734e4e11643200d195bcb154c7534ee581. It is a pure four-change stack, with reported exact source replay in every spread. Prefer it provisionally because it explicitly preserves the spread-2 source handler and has fewer flagged sources than stack-a. This is a recommendation from the supplied results, not an independent certification of exactness. Stack-a is also reported as an exact stack, with broader replay coverage. [B, A]

Demand with floors is worth a subsequent manual entry as a spread-2-only alternative, preserving the other spread policies. Its positive paired gain, lower reported standard error than Cash or Demand, and larger comparison sample than Demand make it a reasonable experiment. Its superiority to the original spread-2 change is not established; retain stack-b as the known-score target. None of these items supplies a Paradigm score for its assembled file. [F, C, D, B]

Results

Each row cites its input result; numbers below are copied without rounding. “Exact” means the item reports equality of every simulation-result field on the tested runs, not a proof for every possible run. For A and B, spread 2's reference is the original flow-stock source. For D, C and F it is respectively Demand, Cash and Demand with floors: those final files are modified stacks, not exact copies of the original four-source combination. All-spread differences are paired candidate-minus-base local holdout means. [A–F]

Item / source	Exact replay by spread: S1; S2; S3; S4	Spread-2 policy and unseen spread-2 gain over base (mean; standard error; runs)	All-spread mean difference from base (standard error; runs)	Seconds: candidate / base; timing basis
stack-a [A]	Yes, 300; Yes, 300; Yes, 300; Yes, 300 in exact; tried additionally reports a second set of 300 per spread	Original own-flow-scaled one-cent stock target; no separate spread-2 gain measured: fields are mean 0.0, SE 0.0, runs 0	0.9499860332040997 (0.42163309602312105; 2000)	107.61346984887496 / 108.3585469298996; 200 runs each, serial, one process
stack-b [B]	Yes, 150; Yes, 150; Yes, 150; Yes, 150; second set not run	Original own-flow-scaled one-cent stock target; no separate spread-2 gain measured: fields are mean 0.0, SE 0.0, runs 0	1.4760373606658999 (0.379637775706742; 1000)	156.6596 / 152.46959999999999; scaled to 200 runs from 500 measured runs each
stack-demand2 [D]	Yes, 150; Yes, 150 against Demand; Yes, 150; Yes, 150	Unmodified Demand policy only in spread 2; 0.9757539952248745; 0.48213013327879434; 993	1.4124942557192997 (0.39742820148574076; 1000)	130.18419093219563 / 130.79801824502647; 100 runs each, sequential, includes initial factory load
stack-cash2 [C]	Yes, 300; Yes, 300 against Cash in exact (also all 1987 in tried); Yes, 300; Yes, 300	Cash only in spread 2: inventory/cash parameters, lower ask loss term, reserve for uncovered asks; 0.7870275225599397; 0.3485094601712951; 1987	0.8751301380108996 (0.39509970575058967; 2000)	39.55992725910619 / 39.35929244500585; 200 runs each, one process
stack-floor2 [F]	Yes, 300; Yes, 300 against Demand with floors; Yes, 300; Yes, 300	Complete Demand-with-floors policy only in spread 2, including its grid and probe handling; 0.6365839910112733; 0.15359022930502778; 1987	1.0147209778444495 (0.41549948352459654; 2000)	141.41617071192013 / 143.52286202501273; 200 runs each, alternating, one warmup each

The zero-run fields for A and B mean not measured, not a measured zero effect. Their local all-spread gains include the original spread-2 change. Each item reports zero failures in its holdout. A, C and F use holdout seeds [70000,72000); B and D use [70000,71000). Timing protocols and machines differ, so use each candidate/base pair locally rather than ranking absolute seconds across rows. [A–F]

Why the score target is conditional

The assignment gives the base's Paradigm mean as 106.804593, and the four spread-specific gains as +0.775951, +0.065331, +0.164311, and +0.574129. The corresponding policies are: a larger front tick-2 ask below one cent with front-ask retention; own-flow-scaled one-cent stock; the 20 dollars ahead rule for a low front ask at the rival tick; and own-fill retail-flow estimation for ask and stock sizing. Because each run has one fixed spread, exact per-run preservation would make the assembled score 108.384315 on those same 200 Paradigm simulations. This is the assignment's arithmetic target, not a fresh score measured by these items. The goal describes progress from 107.58 to 108.38, with first place 109.93. [G]

All five items report base_mean_edge 108.63974289615767, but their calibration notes identify that as the parent result, not the base's Paradigm score. Their actual unseen-base means are 92.95847243878374 for A/C/F and 92.612021147951 for B/D. These are different local samples; none substitutes for 106.804593 or validates the projected leaderboard score. [A–F, G]

The assignment also reports Demand +0.62 and Cash +0.43 on Paradigm, and Demand with floors +0.55 on unseen runs, not entered. Those changes were built on the parent and acted across all spreads. These figures do not establish the gain from using them only in spread 2, nor can they simply be added to the exact-stack target. [G]

How much weight to put on the alternatives

C and F compare against the original flow-stock policy on the same 1987 spread-2 runs selected from [60000,68000). That original policy gains 0.6075141093718673, SE 0.18318063780805277, over the base. Cash's paired advantage over flow is 0.1795134131880722, SE 0.401422244113473: its own result explicitly says the advantage is not significant. F retains floors because its mean exceeds the original policy's mean, but supplies no standard error for the paired floors-minus-flow difference. Its 0.15359022930502778 is the SE against the base, not the SE of improvement over flow. [C, F]

D uses only the first 993 eligible spread-2 seeds in the same interval. Its original-policy baseline gain is 0.41189538722034263, SE 0.19859171263231082, versus Demand's 0.9757539952248745, SE 0.48213013327879434. D provides no paired Demand-minus-flow standard error. Its larger candidate mean cannot establish superiority to Cash or floors measured on a different sample. [D]

The untouched all-spread holdout also gives no decisive replacement ranking: F's mean exceeds A's on their shared holdout, while C's is lower than A's; D's is lower than B's on their shared shorter holdout. No item supplies the paired replacement-minus-exact-stack SE on that holdout. The recommendation to try floors is therefore an experimental priority, not a claimed proven improvement on Paradigm. Cash remains plausible but noisy; Demand has a smaller comparison sample and no demonstrated all-spread holdout advantage over B. [A–F; interpretation of their reported means and methods]

Exactness and source limitations

A reports two exactness sets, starting at seeds 60000 and 64000, each taking the first 300 matching seeds per spread and comparing every result field. B reports only the first 150 per spread from 60000 and explicitly omits the second set from 64000. D reports an initial pure-stack check and final replacement check at 150 per spread; F reports both at 300 per spread; C reports an initial pure-stack check at 300 per spread and final Cash equality as described in the table. These are item-authored replay claims; this write-up did not rerun strategies or inspect their raw replay logs. Preserving numerical operation order matters: C specifically warns that Python's sum() and explicit loops can round differently, while B says it copied the spread-2 handler verbatim. [A–F]

The assignment flags 4 sources for stack-a, 1 source for stack-b, and 3 sources for stack-cash2 because rereads failed or did not support the quoted claims. The supplied result JSONs contain no source lists, failed quotations, or mapping from each flag to a particular assertion. Thus the affected claims cannot be identified more narrowly from these inputs. Treat A/B/C's source-backed assertions about faithful grafts and reproduction as qualified, rather than dismissing their numerical reports or treating them as independently verified. A's wider replay claim does not erase its larger source-verification gap; C's promising gain does not erase its flags. B's own flag also remains unresolved. [G; inspection of A, B, C]

D and F have no source flags named in the assignment; that is not independent validation. F's claim that all source quotes were fetched and checked is itself an item claim. The public simulator repository was readable during this write-up and describes a passive strategy in a binary FIFO order book, but the base artifact URL was inaccessible through the web reader. That attempt neither verifies the artifact nor proves it unavailable by every method. No source-level or runtime reproduction is claimed here. [F; public repository; retrieval observation]

Source key and file identity

The five authoritative inputs for the table are the supplied JSON results below. Their links point to the provided input locations; they are not additional output files. Field references are approach, exact, spread2_gain, diff, seconds, holdout, and tried. All reported experiment statistics above come from these fields. The provided input digests identify the records even if these local links are unavailable to a later reader.

• A: item-stack-a, logical source artifacts/goal/items/stack-a/result.json; input SHA-256 70d4d492e3250eccb55c7f067b982191e265f6563bc29b1ad4440ab6b7b4187d; strategy-code SHA-256 b4bfd62893176844120f69c7c6b4957706141cba50b8971b18c6faaae71dc174.
• B: item-stack-b, logical source artifacts/goal/items/stack-b/result.json; input SHA-256 ceae4e41f3555400ea8dd71cfcabdf3698613142560717eca64acc9e029d0c3d; strategy-code SHA-256 1d2db2a895bf0f9e427f6f15f6be20734e4e11643200d195bcb154c7534ee581.
• D: item-stack-demand2, logical source artifacts/goal/items/stack-demand2/result.json; input SHA-256 380f0c7016025fd87f4116daba0c7e561d3e3c2435df734c3a21043d7a7be102; strategy-code SHA-256 ca57e0ae3e3454f0c72e6713c58df86b96fb8683a0e81e79ed68880c0b4d474c.
• C: item-stack-cash2, logical source artifacts/goal/items/stack-cash2/result.json; input SHA-256 151e0b1b16f90a1384b823bc528d6d1407f7b4bbfcc7966c12119fe483bfd326; strategy-code SHA-256 4bbf720ff34be09b441e84a7e0c1bf47cba036b5211c4118e44a597f941453db.
• F: item-stack-floor2, logical source artifacts/goal/items/stack-floor2/result.json; input SHA-256 e5edea54786d1a0dced7326bf986d5bfb5975d58b59bf155e47287ef81842984; strategy-code SHA-256 4cda6894bb98e7e4bcd2cdece8ae65e89413211d93a36c9fca64e5d8ad5862e9.

G denotes the requester-supplied goal and source-warning notice, the source of the Paradigm figures and flag counts. Its linked strategy artifacts are the base, spread 1, spread 2, spread 3, spread 4, parent, Demand, Cash, and Demand with floors. These links identify the supplied provenance; they are not independently verified score records. The public simulator repository supplies the challenge description only.