PaperlayerView exact revision on arXiv

Source-linked AI summary

OOM-RL II: Reality Is an Oracle, Not a Debugger Provenance-Constrained Diagnosis in Continually Evolving Agent-Engineered Systems

Kun Liu, Liqun Chen

arXiv:2610.10256v1cs.AIcs.SEq-fin.PM

TL;DR

In evolving production systems, external outcomes establish what happened without identifying the procedure or mechanism that produced them. This paper analyzes that distinction in a trading-system archive, separates outcome, identity, and causal claims, and finds that the account gained and outperformed CSI300 even though the archive does not establish a single frozen instance or unique cause.

  • Problem

    In production systems whose code, artifacts, configuration, and operational rules change, external outcomes do not identify which internal procedure or mechanism produced them.

    Source: §1 · pp. 1–2

  • Method

    The paper analyzes outcome evidence alongside evaluated-object identity and causal diagnosis, distinguishing frozen instances, adaptive procedures, and per-execution binding.

    Source: §2.2 · p. 5, §2.3 · p. 5

  • Results

    Across the annual window, the account gained and outperformed CSI300, but the archive does not establish one frozen model or a unique cause.

    Source: §11 · p. 25

  • Takeaways & Limitations

    Reality may establish that an outcome occurred without identifying which evolving procedure produced it or why.

    Source: §11 · p. 26

    The archive does not establish a single frozen instance, and this case’s archive-relative criterion was not evaluated by enumerating its compatible-history set.

    Source: §1 · p. 2, §2.4 · p. 6

Abstract

from arXiv · show

Reality may establish that an outcome occurred without identifying which evolving procedure produced it or why. This distinction matters in production ML systems whose code, configuration, and artifacts change while external feedback accumulates. We examine it in a human-directed, agent-engineered quantitative trading system, using oracle to mean an external source of realized outcomes rather than a complete correctness specification. Across one year, the account gained and outperformed a broad market index, while annual alpha was not statistically distinguishable from zero under the main retrospective specification. Retrospectively selected subperiods include adverse relative performance and conditional candidate-level weakness under declared approximate references. Engineering records document changes during the episode, and complete recommendation-to-runtime binding is unavailable. The archive does not establish a common frozen instance or a unique cause. The case motivates an outcome--diagnosis gap: outcome evidence, evaluated-object identity, and causal explanation support distinct claims. We distinguish frozen instances, pre-specified adaptive procedures, and ad-hoc development; organize archive-relative claim identifiability and an evidence hierarchy; and propose a prospective production-binding protocol. An illustrative compatible-history example shows how factual binding can resolve a recommendation's referent without supplying its counterfactual effect. The protocol is proposed rather than prospectively validated. External feedback constrains outcome claims, while provenance and additional identification structure determine the resolution of diagnosis.

1 Introduction

External outcomes constrain what happened but do not identify the evolving procedure or mechanism that produced them. The paper examines this gap in a human-directed, agent-engineered trading system and develops distinctions for interpreting its evidence and preserving future provenance.

1 Introduction

  • An external outcome establishes what happened, but does not identify the internal procedure or mechanism responsible for it.Losses are stochastic outcomes, not built-in software-correctness labels; failure thresholds, comparators, and analysis windows must be specified.Source: §1 · pp. 1–2
  • The 2026 account’s poor performance and rebound could not be reconstructed as a single frozen model entering and leaving a market regime.Repository and configuration records document changes, but complete per-action deployment identity was not persisted.Source: §1 · p. 2

1.1 From one non-stationarity to two

  • Changing environments and changing decision procedures create distinct sources of non-stationarity, so longitudinal claims depend on whether the evaluated object is frozen, pre-specified adaptive, or ad hoc.Outcomes are stochastic, delayed, and reference-dependent, and historical markets cannot literally be reset for debugging.Source: §1.1 · p. 2

1.2 Research questions

  • The research questions ask which outcome and output symptoms the data support, whether the archive identifies one frozen function or root cause, and what provenance longitudinal claims require.The questions condition symptoms on explicit references, scopes, and horizons.Source: §1.2 · pp. 2–3

1.3 Contributions

  • The case reports one year of live account performance and conditional symptoms, but establishes neither a common frozen instance nor a unique technical cause.Development and configuration changes overlap the episode while action-level binding remains incomplete.Source: §1.3 · p. 3
  • The paper distinguishes evaluated-object identities, proposes archive-relative claim identifiability and an evidence hierarchy, and proposes a production-binding protocol that has not been prospectively evaluated.Pre-specified adaptation can preserve a prospective record across state updates, while exact factual identity alone does not supply counterfactual identification.Source: §1.3 · p. 3

2 Conceptual Framework: Reality as Oracle, Not Debugger

The framework separates realized outcomes from increasingly specific claims about symptoms, subsystems, and causes. It distinguishes evaluated-object identities and treats archive-relative identification as invariance across evidence-compatible histories, not as a causal method.

2.1 Restricted oracle and the outcome–diagnosis gap

  • An external outcome is evidence of what occurred, not a complete correctness specification or an explanation of the procedure that generated it.Even a reference-relative symptom depends on comparator, timing, horizon, and scope.Source: §2.1 · p. 3

2.2 One action, outcome, and update model

  • The outcome model treats realized outcomes as depending on the executed action path, execution-time context, external environment, and residual stochasticity.Recommendations precede human substitutions, cash allocation, and feasible fills; candidate forward returns are not realized account P&L.Source: §2.2 · p. 5

2.3 Evaluated-object identities

  • A frozen instance fixes its procedure and relevant artifacts, a pre-specified adaptive procedure fixes allowed updates, and ad-hoc changes create new procedure-level claim boundaries.Per-execution binding to realized artifacts, data-as-of state, universe, overrides, and context is an additional dimension, not a fourth procedure class.Source: §2.3 · p. 5

2.4 Evidence-compatible histories and archive-relative claim identifiability

  • A claim is archive-identified when it is invariant across compatible histories under explicit interpretation assumptions; reliable added evidence can narrow those histories.The criterion is conceptual, not an identification theorem or history-enumeration algorithm for this case.Source: §2.4 · pp. 5–6
  • Recovering a recommendation’s factual runtime binding can identify its referent without identifying whether another runtime would have changed a later portfolio loss.The illustrative loader histories and recovered binding are hypothetical; causal claims additionally require an intervention and comparative assumptions.Source: §2.4 · p. 6

2.5 Evidence hierarchy

  • The evidence hierarchy progresses from observed outcomes to reference-relative symptoms, subsystem association, and causal root cause, with stronger evidentiary requirements at finer granularity.Reference design, provenance and comparison, and causal identification are requirements rather than guarantees; association is not a causal effect.Source: §2.5 · pp. 6–7

3 System and Evidence

The paper describes a human-directed trading system and defines retrospective windows, account returns, comparison references, candidate horizons, and statistical conventions. Its archive documents engineering changes but does not establish their exact activation or a single production instance.

3.1 Production setting

  • QuantPits generated recommendations on a weekly-like batch schedule, with LLM agents assisting engineering under human direction and deployment authority.Manual tradability judgments, substitutions, lot sizing, and cash allocation separated recommendations from actual purchases.Source: §3.1 · pp. 7–8

3.2 Analysis-wide annual window

  • The fixed-for-analysis annual window overlaps Paper I’s Phase 3 for 94 days, so it is not an independent replication.The annual interval was fixed at the evidence freeze, and Paper I does not establish persistent alpha, a causal module effect, or frozen-model out-of-sample performance.Source: §3.2 · p. 8

3.3 Account returns and cash flows

  • Daily account returns use a beginning-of-day cash-flow convention, but that convention does not reconstruct intraday funding timing.Small recorded negative cash balances are retained without inserting additional financing cost or fictitious funding.Source: §3.3 · pp. 8–9

3.4 Reference series

  • The approximate equal-weight reference covers the eligible CSI300-related subset with daily ideal gross rebalancing and is not an executable trading strategy.Its carried-price convention is a valuation choice, not an assertion of tradability; candidate-relative returns are signal-close matched, with no execution-open-matched series.Source: §3.4 · p. 9

3.5 Candidate horizons

  • Candidate outcomes use signal-close or execution-open anchors and horizons from h1 through h20, with 7,366 of 7,596 requested labels observed.The labels are counts across both anchors and all horizons, not distinct stocks; censored outcomes occur at the evidence cutoff.Source: §3.5 · p. 9

3.6 Statistical convention

  • The main retrospective specification uses a single-factor daily excess-return regression with HAC covariance at lag 3, a convention not prospectively registered.HAC changes covariance rather than coefficients and does not correct retrospective window selection or multiple testing.Source: §3.6 · pp. 9–10

3.7 Recorded evolution and temporal ambiguity

  • Engineering/configuration record dates overlap analytical spans, but do not establish production activation boundaries or one instance throughout.The March–June statistical segment differs from the April–June incident narrative; exact per-action activation is unknown.Source: Figure 2 · p. 10, §3.7 · p. 10

4 Annual Context and Incident

The annual account was positive relative to the index, but its estimated alpha was not statistically significant under the stated specification. Retrospectively selected segments show sharply differing estimates, while monthly returns provide additional context.

4 Annual Context and Incident

  • 14.49% annualized alpha (HAC(3) p = 0.275) was not statistically significant under the chosen annual single-factor specification, despite positive relative performance.The multi-factor intercept was 13.52% (HAC(3) p = 0.226).Source: Table 1 · p. 11, §4 · p. 11
  • 30.07% rounded annualized alpha (OLS p = 0.094) was reproduced for the 94-day overlap with Paper I’s Phase 3, but the Initial segment’s 5% significance classification is sensitive to the HAC(3) convention.This comparison of reported values is not a new verification of Paper I’s original inputs.Source: §4 · p. 11
  • −58.1058% market-only annualized intercept (HAC(3) p = 0.01778) changed to −25.3155% (p = 0.22919) with three lagged style proxies, while the account loss remained −14.7887%.The March–June interval was selected after observing the production trajectory, so its p-values do not repair retrospective window selection; the intercept inference is sensitive to explanatory variables.Source: §4 · p. 11

4.1 Monthly context

  • March account return was −4.2922%, versus CSI300 −5.5321% and approximate EW −6.0845%; August was +6.1394%, followed by September at −1.1940%.March’s absolute loss alone does not establish reference-relative degradation; the monthly table uses contemporaneous close-to-close return dates.Source: §4.1 · p. 11

5 RQ1: What Production Symptoms Are Actually Supported?

The archive supports conditional output-level symptoms under declared references, not a universal Buy-signal failure or a cause independent of market conditions and production changes. Candidate scope, reference, and retrospective selection materially limit interpretation.

5.1 Candidate classes are not interchangeable

  • January and February primary candidates had negative median excess returns (−2.1662 pp, n = 13; −4.0636 pp, n = 10), while all-candidate medians were positive (+1.6262 pp, n = 39; +0.8827 pp, n = 30).These occurrence summaries establish neither a uniformly healthy early layer nor statistically tested cross-period deterioration.Source: §5.1 · p. 13

5.2 T+10 absolute and relative symptoms

  • May’s all-candidate median was −1.7557% absolute but +0.4430 pp relative; primary candidates were −0.8288% absolute and +1.0277 pp relative.April and June have negative T+10 relative symptoms against the approximate equal-weight reference, but the selected cells do not prove environment-independent causal failure.Source: §5.2 · p. 13

5.3 Diagnostic narratives tested against the archive

  • Style estimates remain specification-sensitive; May-labeled instruments contributed −2.84050 pp versus −5.06181 pp outside that cohort, while freeze-replay comparisons do not establish that less trading solved the loss.The outside group is not all June acquisitions, and replay does not identify an optimal live policy or failure onset.Source: §5.3 · p. 14

5.4 Approximate random-ranking reference

  • June’s actual −9.2297% was at or below 272/5,000 Drop3 paths, while 51/5,000 annual paths equaled or exceeded the observed +11.1391% account return.These approximate-reference positions answer different questions from regression inference and are not calibrated alpha, skill-probability, or causal tests.Source: Figure 6 · p. 15, §5.4 · p. 15, Figure 7 · p. 16

5.5 RQ1 answer

  • April and June show negative T+10 candidate-relative outcomes, while 272 of 5,000 approximate random-ranking paths fall at or below June’s observed account return; other scopes and diagnostics are mixed.The selected-period findings do not establish a cause independent of market conditions or production-stack changes.Source: §5.5 · p. 16

6 RQ2: What Can Be Attributed to the Evolving Production Stack?

The archive records production outcomes and development/configuration changes, but does not bind every historical action to a runtime instance. It therefore supports conditional symptoms during an evolving process, not a common frozen-model narrative or unique technical cause.

6.1 Repository/configuration history is not deployment identity

  • Repository and configuration records document changes during the broad incident period, but do not establish which changed components were active for particular production actions.Commit and note dates are record dates, not established per-action activation dates.Source: §6.1 · p. 16, Table 3 · p. 17

6.2 Prediction-infrastructure changes

  • Inspected diffs record changes to prediction loading, recorder lookup, predict-only artifacts, and ensemble normalization; configuration notes record hyperparameter revisions and CPCV capability.These records motivate engineering explanations without establishing a particular contaminated order.Source: §6.2 · pp. 16–17

6.3 What is established and what remains open

  • The established evidence includes observed annual and incident-period outcomes, selected candidate-relative T+10 symptoms, and stack changes; complete per-action production identity is not established.The rebound is another evolving-account outcome, not evidence of the same frozen instance or a repair mechanism.Source: §6.3 · p. 17

6.4 Evidence-to-claim boundary

  • A resolved factual binding would identify a recommendation’s referent, but causal explanation still requires a defined intervention, comparative evidence, and counterfactual assumptions.Missing records alone neither prove formal non-identification nor show that different loader versions ran or changed outcomes.Source: §6.4 · pp. 17–18

6.5 RQ2 answer

  • The archive supports conditional production symptoms during an evolving development/configuration process, but not a clean frozen-model regime-failure narrative or unique technical root cause.Complete action-level runtime bindings remain unestablished.Source: §6.5 · p. 18

7 RQ3: Provenance-Bounded Evaluation

The proposed approach separates frozen instances, pre-specified adaptive procedures, and ad-hoc revisions, then binds actions to the relevant object. These prospective practices can clarify what was evaluated, but do not by themselves establish causal diagnosis.

7.1 Choose and bind the scientific object

  • A frozen instance requires prospective fixation and per-action binding; pre-specified adaptation can continue a prequential record, while discretionary rule revisions create new procedure-level boundaries.Information timing and allowed adaptations must be declared before outcomes.Source: §7.1 · p. 18

7.2 Grouped per-action provenance manifest

  • The proposed grouped provenance manifest uses immutable content references and recommends linking each recommendation to its run/instance, procedure, configuration, artifact, and input snapshots.The design is a proposal, not a new standard or evidence of historical compliance; manifest completeness and causal identification remain separate.Source: §7.2 · p. 19, Table 5 · p. 19

7.3 Prospective protocol for frozen instances

  • For frozen-instance evaluation, freeze relevant code, configuration, and artifacts, bind actions and execution-time inputs, and end the instance record at material change.Recommendations and executed actions remain distinct, including when an outcome horizon spans a change.Source: §7.3 · p. 19

7.4 Prospective protocol for adaptive procedures

  • For adaptive procedures, declare adaptation rules and information timing in advance, persist realized state and inputs, continue the prequential record across allowed updates, and mark discretionary revisions as new boundaries.Earlier evidence remains associated with its original procedure object.Source: §7.4 · pp. 19–20

7.5 Incident protocol and compatible alternatives

  • The incident protocol separates observed outcomes from fault labels, resolves generating procedures, uses telemetry for subsystem claims and interventions or justified counterfactuals for causal claims, and records unresolved alternatives.These are prospective proposals, not evaluated interventions or evidence of historical compliance.Source: §7.5 · p. 20

7.6 RQ3 answer

  • Frozen-instance, adaptive-procedure, and execution-state claims need distinct identity bindings; provenance constrains referents, while reference design and causal identification determine diagnostic reach.Resolvable recommendation-to-runtime and action-to-recommendation links are a practical first priority, not an automatic increase in claim level.Source: §7.6 · p. 20

8 Discussion

External outcomes constrain claims about realized paths but do not identify internal causes; resolved provenance can establish a claim’s referent, while causal interpretation needs further identification structure. The discussion contrasts controlled diagnostics with evolving market evidence and treats rebound, references, and execution as bounded evidence.

8.1 A strong oracle can still be a weak debugger

  • An external outcome constrains claims about the realized account path but supplies no test assertion naming a violated internal invariant.Outcomes are stochastic, reference-dependent, and path-dependent.Source: §8.1 · p. 20

8.2 System-level diagnostic credit assignment

  • System-level diagnostic credit assignment targets models, artifacts, training windows, data paths, ensemble rules, execution decisions, and human overrides.The phrase describes an enlarged attribution problem, not a replacement for established RL terminology.Source: §8.2 · p. 21

8.3 Technical debt and mutable ML systems

  • A production pipeline’s behavior cannot be reduced to its model coefficients; detached performance may answer an organization-level question but not a model-level one.This is consistent with the systems view of entanglement, feedback loops, dependencies, configuration, and external change.Source: §8.3 · p. 21

8.4 Provenance and the referent of a claim

  • Resolved provenance can identify the procedure, artifact, data state, and action behind an outcome claim, but causal interpretation requires additional comparison and identification structure.Without binding, analysis may address the wrong system.Source: §8.4 · p. 21

8.5 Controlled fault injection versus uncontrolled reality

  • Controlled fault generation supplies known diagnostic structure, whereas naturally occurring market incidents trade controlled clarity for ecological validity.The case cautions against projecting causal certainty from controlled benchmarks onto uncontrolled longitudinal production data.Source: §8.5 · p. 21

8.6 The rebound is also epistemically constrained

  • July–September returned +15.2461% over 65 observations; annualized intercepts were +76.5417% (p = 0.00230) single-market and +53.3570% (p = 0.00793) multi-proxy.These retrospective estimates do not establish recovery of the same frozen model; improvement should not substitute for resolving the original diagnostic questions.Source: §8.6 · pp. 21–22

8.7 Why the approximate random reference is useful anyway

  • The June approximate random reference describes production performance relative to uninformed selection under an approximate common portfolio rule, without requiring a strict hypothesis test.Annual context alongside the poor selected month helps prevent benchmark choice from becoming a cherry-picking device.Source: §8.7 · p. 22

8.8 Execution friction as bounded context

  • Annual execution friction cannot exclude execution as an incident explanation, while the passive pool is informal context rather than a matched, risk-adjusted, or causal benchmark.Account returns already include actual settlements, so price friction is not deducted again.Source: §8.8 · p. 22, Meanwhile, in the passive pool · p. 22

9 Related Work

Related work spans coding agents, oracle and fault-localization methods, credit assignment, production lineage, controlled incidents, and financial inference. This paper connects those strands to the identity of an evolving evaluated system and the limits of causal claims from incomplete provenance.

9.1 Coding agents and human oversight

  • Coding-agent task views are extended here with runtime identity, human acceptance, and action binding; tasks, direction, and deployment oversight remained human-controlled.Human–automation research motivates an oversight question, not a diagnosis of the operator’s psychology.Source: §9.1 · p. 23

9.2 Oracles, fault localization, and diagnostic justification

  • Oracle acceptance and fault-localization methods separate terminal feedback from causes, while this work asks which evolving program produced an outcome that cannot literally be rerun.JustDiag separates answer quality from process quality; this case instead focuses on action-to-runtime identity for longitudinal claims [19].Source: §9.2 · p. 23

9.3 Credit assignment and adaptive evaluation

  • Credit assignment is extended from delayed consequences to system components and human execution; adaptation remains prospectively assessable under declared procedures when runtime binding is preserved.The analysis treats the market path as approximately given and does not estimate performative feedback effects.Source: §9.3 · p. 23

9.4 Production ML observability, technical debt, and lineage

  • The proposed binding connects recommendations to runtime identities and realized actions, complementing observability, model registries, provenance standards, and agent-tracing tools.It also links procedure and artifact lineage to information timing and substitutions.Source: §9.4 · pp. 23–24

9.5 Controlled incidents and causal identification

  • Controlled fault injection supplies known experimental structure, but live cases lack resettable paths and labels; knowing what ran resolves identity, not causal effect.The compatible-history criterion organizes this distinction rather than providing a new causal estimator.Source: §9.5 · p. 24

9.6 Financial selection and inference

  • Financial selection and inference warnings bound the case: it does not compute DSR, PBO, or a selection-adjusted June test, and differing p-values do not identify a model contrast.The paper connects these methods and warnings to mutable identity, incomplete provenance, and causal claim granularity.Source: §9.6 · p. 24

10 Threats to Validity

The retrospective analysis is limited by selected windows, approximate references and replays, incomplete deployment binding, specification choices, and a single unvalidated case. These constraints narrow what the archive can establish about causes and generalization.

10.1 Retrospective selection

  • Segmented windows and June focus followed observation of the trajectory, so selected-window estimates lack a preregistered interpretation and diagnostics are not independent replications.HAC covariance does not correct window selection or multiple testing.Source: §10.1 · p. 24

10.2 Reference, replay, and accounting approximations

  • Approximate gross references, signal-close matching, and simplified replays do not reproduce all execution mechanics or constitute exact alternate histories or root-cause probabilities.Fixed-basket and account-linked contributions answer different accounting questions, and neither is FIFO realized P&L.Source: §10.2 · pp. 24–25

10.3 Archive completeness and deployment binding

  • Repository and configuration dates do not establish activation or recommendation-to-runtime links; missing bindings neither prove particular alternatives ran nor permit inferred deployment epochs.The limited run inventory is not an annual action census, and success declarations are not independent verification.Source: §10.3 · p. 25

10.4 Statistical and model specification

  • Regression inference depends on covariance, risk-free-rate, and explanatory-variable specifications; overlapping horizons and repeated candidates are not independent samples.Right-censoring requires reporting observed coverage rather than assuming completeness.Source: §10.4 · p. 25

10.5 Single-case generalization and protocol status

  • This single finance case does not estimate feedback effects or population magnitudes, and the proposed production-binding protocol was neither fully implemented historically nor prospectively validated.Agent-assisted implementation and human authority are operator-reported and partly illustrated, not quantified through authorship, speed, or oversight measures.Source: §10.5 · p. 25

11 Conclusion

The case separates realized outcomes, evaluated-object identity, and causal diagnosis: annual account gains and outperformance of CSI300 did not make annual alpha statistically distinguishable from zero under the main retrospective specification. Incomplete historical runtime binding leaves the archive without a common frozen instance or unique cause; the appendices provide bounded monthly and candidate-level diagnostics.

11 Conclusion

  • The account gained and outperformed CSI300 over the fixed-for-analysis annual window, while annual alpha was not statistically distinguishable from zero under the main retrospective specification.Selected subperiods include adverse relative performance and strong later outcomes; candidate and reference diagnostics remain conditional on declared scopes.Source: §11 · p. 25
  • Engineering and configuration changes overlapped the episode, but incomplete action-to-runtime binding leaves an evolving account record rather than evidence of one frozen instance or unique cause.Frozen-instance claims require material identity binding; a declared adaptive procedure can retain a prospective record across state updates.Source: §11 · pp. 25–26

A Complete Monthly Numerical Context

B Secondary Candidate, Attribution, and Replay Diagnostics

B.1 Candidate outcomes across horizons

  • Candidate forward-return curves are mixed across horizons and do not estimate an information half-life or realized trade P&L.April short horizons are not uniformly positive, June h1/h3 are positive in some combinations, and July shows no universal improvement.Source: §B.1 · p. 27

B.2 Candidate persistence

  • 422 of 483 candidate occurrences (87.3706%) began a run; the longest observed run was five recommendation batches across 195 unique Buy-candidate stocks.Run starts do not establish freshness of underlying models, artifacts, data, features, or signals.Source: §B.2 · pp. 27–28

B.3 Sell-side observations

  • At T+10, April’s suggested Sell median was +0.7232% (n = 12), while June’s was −4.2904% (n = 15); these are not realized sale P&L.The observations neither identify Sell logic as the primary cause nor exclude Sell-side effects.Source: §B.3 · p. 28

B.4 Swap observations

  • The swap diagnostic compares equal-weight Buy and Sell basket returns from execution-day open to a shared horizon, but does not establish a uniformly negative replacement mechanism.It is not matched-trade realized P&L and does not isolate opportunity cost or manual execution effects.Source: §B.4 · p. 28

B.5 Fixed-basket versus account-linked contribution

The section contrasts fixed-basket returns with account-linked June contributions, then presents approximate replay, freeze, random-reference, factor, execution-friction, and archive-evidence analyses. These retrospective diagnostics have defined approximations and scope limits that constrain causal interpretation.

B.5 Fixed-basket versus account-linked contribution

  • −8.0756% was the May-entry basket’s median June return, with a −10.0925% mean and a −4.8413 pp fixed-basket contribution.This construction holds 29 May equity-value weights and adjusted units fixed, excluding cash; it is not the actual June NAV decomposition.Source: §B.5 · p. 28
  • −5.06181 pp came from stocks outside the fixed May-29 cohort, versus −2.84050 pp from May-labeled and −1.40736 pp from April-labeled stocks; components sum to June’s −9.2297% account return.The outside-cohort group was the largest negative contribution; cohort membership does not establish causal responsibility.Source: §B.5 · pp. 28–29

B.6 Opening-policy replay

  • −16.6910% (Drop0), −19.6462% (Drop1), −12.9816% (Drop3), −11.8967% (Drop6), and −17.0115% (Drop9) were approximate replay returns.Drop6 exceeded configured Drop3 on this path, but the non-monotonic comparison neither identifies an optimal live policy nor shows that changing turnover would have fixed the incident.Source: §B.6 · p. 29

B.7 Freeze counterfactual

  • The 1 April snapshot’s freeze exceeded production over its matched interval, but the comparison does not establish a causal failure onset or identify a responsible component.The three retrospective rows use different holdings, start dates, and durations, so their cumulative returns do not identify an optimal freeze date.Source: §B.7 · p. 30, Table 7 · p. 30

C Approximate Random-Reference Mechanics

  • The random reference generated 5,000 continuous annual paths per profile, with Drop0/3/22 profiles sharing trial scores and approximate rebalance and fee mechanics.It is not a matched production null: simulations can include suspended stocks and omit opening fills, whole-lot, minimum-fee, and manual-override mechanics.Source: C Approximate Random-Reference Mechanics · p. 30

D Factor-Proxy Construction and Sensitivity

  • The factor sensitivity uses lagged liquidity, momentum, and volatility long–short proxies, while fixed-lag HAC uncertainty changes some significance classifications without correcting retrospective window selection.These engine-inspired proxies are not canonical Fama–French, Barra, or industry-neutral factors; the reconstructed definitions also differ from the legacy production implementation.Source: D Factor-Proxy Construction and… · p. 31, Table 8 · p. 31

E Execution-Friction Definitions and Full-Year Context

  • Annual transaction-notional-weighted total friction was approximately +0.0348% for Buys and −0.3839% for Sells; the paper does not infer that execution cannot explain the episode.These measurements cover the full annual archive, not an incident-window decomposition.Source: E Execution-Friction Definitions and… · pp. 31–32

F Additional Diagnostic Figures

  • Additional figures show candidate-horizon medians, signal-label coverage, replay paths, factor intervals, and friction; the March–June proxy interval spans zero under its specification.The figures use the same evidence archive as the main tables and retain the stated replay and factor approximations.Source: F Additional Diagnostic Figures · p. 32, Figure 13 · p. 34

H Code and Data Availability

Supplementary Dataset S1 provides account-path and holding-set records for inspection, alongside detailed fields, daily tables, and visualizations. Its scope excludes key trade-level and runtime evidence and does not constitute an independent brokerage certification or complete empirical replication package.

H Code and Data Availability

  • S1 contains 241 trading-day returns and 242 closing snapshots, including the 30 September 2025 baseline, with normalized system, CSI300, and approximate subset-EW NAV/drawdown paths.It also includes closing stock counts, set changes, stable pseudonymous holding memberships, detailed fields, daily tables, and holding visualizations.Source: H Code and Data Availability · pp. 35–36
  • S1 supports inspection of account paths and holding-set evolution, but not independent brokerage certification or reproduction of every diagnostic table.Real security identifiers, capital and position amounts, trade-level records, candidate diagnostics, and private runtime evidence are excluded; analysis and export scripts remain in the local archive.Source: H Code and Data Availability · p. 36
  • The public QuantPits engine is distinct from S1 and is not a complete empirical replication package.S1 summarizes operator-maintained system records and excludes outside-system funds; stable aliases do not guarantee anonymity.Source: H Code and Data Availability · p. 36
Loading 2610.10256v1…