A standalone, source-pinned review of two QCA article-to-repository studies: the QuantBrain-attributed result and Qwen-attributed alternative for each article, including executable behavior, numerical evidence, paper fidelity, test coverage, and provenance.
In the Avellaneda–Stoikov study, the developer-attributed Qwen repository is materially closer on the central Monte Carlo experiment, while QuantBrain 397B attempts more of the four-part analytical scope. Both contain serious untested correctness defects.
In the skfolio study, the Qwen-attributed repository preserves more recoverable evidence and a genuine held-out walk-forward result. QuantBrain has the sounder train/test structure in Experiment 1, but its report reverses the observed result and its all-run command fails before Experiment 3. Neither study is sufficiently controlled to infer general model capability.
A four-part theory and Monte Carlo exercise built around an uploaded preprint whose numerical targets differ from a later published scan.
Three workflows covering a constrained efficient frontier, a factor-model prior, and stacked portfolio optimization with walk-forward evaluation.
Market making and portfolio optimization.
Two developer-attributed artifacts per QCA research study.
One supplied comparison run is unavailable in each public share.
Tasks, prompts, and provenance are not controlled tightly enough.
The developer stated on 17 July that the language models were used only for repository generation and that the remaining behavior came from QCA presets. The retained traces support that division, with important attribution limits.
QCA acquired the article, produced the experiment object, selected presets, and dispatched the repository job. The skfolio PDF-analysis stage specifically requested GPT-5.4.
The developer attributes the repository contents to the requested QuantBrain or Qwen coding model. Git commits use the neutral author QCA Agent.
GPT-5.4 reviewed selected generated files and returned a reproduction disposition. Those reviews did not substitute for source execution or independent scientific verification.
| Workflow stage | Evidence-based attribution |
|---|---|
| Article acquisition and parsing | QCA workflow; the skfolio PDF-analysis tool specifically requested GPT-5.4 |
| Experiment-object generation | QCA workflow / GPT-5.4 stage |
| Job dispatch and terminal-state handling | QCA orchestration |
| Repository implementation | Developer-attributed coding model |
| Automated repository judgment | QCA workflow / GPT-5.4 judge |
| Independent checks on this page | This repository audit |
“Result” and “alternative” denote the developer-attributed QuantBrain and Qwen artifacts, not a randomized treatment/control design. The Qwen skfolio links happened to arrive first.
| QCA reproduction | Source article | QuantBrain-attributed result | Qwen-attributed alternative | Trace situation | Artifact-level outcome |
|---|---|---|---|---|---|
| 1 | Avellaneda–Stoikov market making | QuantBrain 397B repository | Qwen-attributed repository | Corrected QuantBrain trace accessible; Qwen run absent from share | Qwen has the meaningful central numerical reproduction; QuantBrain attempts broader scope |
| 2 | skfolio portfolio optimization | QuantBrain-attributed repository | Qwen-attributed repository | Qwen trace accessible; QuantBrain run absent from share | Qwen is more recoverable overall; QuantBrain has one better evaluation structure but a broken evidence path |
Different questions require different evidence
| Dimension | Avellaneda–Stoikov · Qwen | Avellaneda–Stoikov · QuantBrain 397B | skfolio · Qwen | skfolio · QuantBrain |
|---|---|---|---|---|
| Central executable result | Paper-scale Monte Carlo; low/medium γ materially closer to uploaded targets, high γ mean fails | Different process and calibration; about 0.5 mean P&L | Real walk-forward result; Exp. 1/2 OOS tables invalid | Runner aborts before Exp. 3; claimed outputs absent |
| Scope coverage | Different, narrower four-part outline; task equivalence is unknown | Broad attempt at all four topics visible in its trace | All three generated workflows are represented | All three modules exist, although the runner cannot complete them |
| Method fidelity | Target process and sounder formula check; signed-distance and reusable-API defects remain | Central symbolic, process, calibration, and metric mismatches | Paper-aligned outer Exp. 3; one fidelity deviation plus a non-temporal default-CV caveat | Wrong Exp. 3 outer scope and changed internal CV |
| Repository evidence | Single root, 27/27 tests, parseable outputs; tests miss reusable API | Unrelated histories merged; HEAD 62/70; report/data conflicts | Committed predictions, objects, tables, plots; 124 tests | Narrative plus .gitkeep; no claimed result files |
| Public trace coverage | Supplied run absent from share; attribution is developer-provided | Requested model and successful status visible; model exchange absent | Requested model and reported success visible; generation opaque | Supplied run absent from share |
| Artifact-level assessment | More useful numerical starting point, with major repairs required | Broader architecture, but claims are not supported by current files | More recoverable starting point, not a complete reproduction | Useful structural ideas, but workflow and evidence require repair |
“Qwen” is shorthand for developer attribution where its public trace is unavailable. skfolio’s QuantBrain model ID is likewise not verifiable from the supplied public share.
The source is Marco Avellaneda and Sasha Stoikov, “High-frequency trading in a limit order book.” The corrected comparison uses the developer-identified 19 July QuantBrain contribution and the developer-attributed Qwen repository. The available task authority is the four-experiment object embedded in the QuantBrain trace.
QuantBrain attempts the trace-visible analytical, Monte Carlo, robustness, and intensity work. Qwen contains a different four-part outline, and its missing trace prevents prompt-equivalence checks.
Its experiment uses the target arithmetic increments and paper-scale parameters. Fresh low- and medium-γ outcomes are materially closer to the uploaded targets, with small systematic offsets and no predefined acceptance band.
It uses geometric Brownian motion, σ=0.02, A=1, and 252 steps instead of arithmetic increments, σ=2, A=140, and 200 steps.
QuantBrain has central theory/calibration errors. Qwen’s reusable strategy and simulator contain severe defects that its green tests never import.
| Article artifact | Bytes / pages | SHA-256 | Material distinction |
|---|---|---|---|
Benchmark upload · LimitOrderBook.pdf | 192,083 / 14 | 7a53a062ce5786167b592c457975edf212656615f36d6eae148bf410e63858de | γ={0.01, 0.1, 0.5}; displayed constant spreads 1.33 / 1.29 / 1.15 |
| Later published scan | 433,581 / 8 | 71fa9268339ae82365171caa7416833b3a5d584f05e4f4c087c4ae72d7121a8d | Different γ values, tables, and average-spread convention |
S0=100,
T=1, sigma=2, dt=0.005,
A=140, k=1.5, arithmetic two-point price
increments, 200 decisions, terminal std(q_T), and both
inventory-aware and symmetric strategies. The uploaded bytes and frozen
experiment object are the audit authority.
Pinned to audited commits
| Property | Developer-attributed Qwen | QuantBrain 397B |
|---|---|---|
| Pinned target | 111e920a42b9… · single root | 7f58921abdb8… isolated contribution; aba55eda8af2… merged HEAD |
| Older unrelated root | — | 98d4e2de19fea0957e730ad246f11a0b9790bd6c |
| Trace attribution | Supplied run returns 404; developer attribution only | Explicit request says runpod/quantbrain-code-397b-a17b; inherited metadata says runpod/quantbrain |
| Tracked files / size | 33 files · 2.75 MiB in the audited Windows worktree | 80 files / 179.41 MiB isolated; 149 / 181.65 MiB at merged HEAD |
| Supplied tests | 27/27, repeated successfully 5 times | 21/21 isolated; 62/70 at current merged HEAD |
| Dependency state | Lower bounds and unbounded packages despite “pinned” claim | No dependency manifest for the new implementation |
| Result-file integrity | Operational outputs, but documentation/layout drift | Direct narrative/CSV/path contradictions; 98.1% of bytes are six raw-path CSVs |
QuantBrain’s merged HEAD joins unrelated 3 July and 19 July histories with no merge base. Its eight current failures are inherited-test import errors against the new API.
Source inspection plus independent probes
2·q·σ²·(T−t), making the result inventory-dependent and sometimes negative.λ′(δ)=γλ(δ) is impossible for positive γ and a decreasing exponential intensity. The derivation drops the utility-jump term, then changes formulas without repairing the derivation.σ=0.02, A=1, and 252 steps replace σ=2, A=140, and 200 decisions.1−exp(−λdt) replaces the prescribed λdt Bernoulli approximation.qT.| Inventory q | Reported full spread | Ask distance | Bid distance | Interpretation |
|---|---|---|---|---|
| −1 | −4.0 | −2.3333 | 1.6667 | A total bid–ask spread cannot be negative. |
| 0 | 0.0 | −0.3333 | −0.3333 | The target approximation is positive and inventory-independent. |
| 1 | 4.0 | 1.6667 | −2.3333 | The subtraction measures skew, not total spread. |
| Gamma | QuantBrain mean P&L | QuantBrain P&L std | Mean executions | Uploaded-paper inventory mean P&L |
|---|---|---|---|---|
| 0.01 | 0.504 | 1.285 | 0.695 | 66.78 |
| 0.10 | 0.502 | 1.288 | 0.711 | 62.94 |
| 0.50 | 0.491 | 1.337 | 0.791 | 33.92 |
The scale difference is not a tolerance dispute: fewer than one mean execution per path gives the inventory mechanism almost no opportunity to operate.
Fresh seed-42 and 20-seed checks
| Gamma / strategy | Uploaded mean / std P&L | Fresh Qwen mean / std P&L | Uploaded / fresh std qT | Assessment |
|---|---|---|---|---|
| 0.01 · inventory | 66.78 / 8.76 | 68.19 / 8.67 | 4.70 / 5.04 | Close at the audited seed. |
| 0.01 · symmetric | 67.36 / 13.40 | 68.17 / 13.13 | 8.65 / 8.58 | Close. |
| 0.10 · inventory | 62.94 / 5.89 | 63.97 / 5.98 | 2.80 / 2.88 | Close. |
| 0.10 · symmetric | 67.21 / 13.43 | 69.07 / 13.01 | 8.66 / 8.77 | Close. |
| 0.50 · inventory | 33.92 / 4.72 | 23.66 / 4.91 | 1.88 / 1.97 | Dispersion is close; mean P&L is about 30% low. |
| 0.50 · symmetric | 66.20 / 14.53 | 67.64 / 13.39 | 9.06 / 9.03 | Broadly close. |
The 20-seed probe shows stability but remains descriptive: no acceptance band was defined before inspection.
full_spread function. It does not
reproduce the frozen-inventory/Hamilton–Jacobi–Bellman/utility derivation
or the power-law-intensity comparison visible in QuantBrain’s task object.
Because Qwen’s trace is missing, this is a coverage difference—not a
demonstrated instruction-following failure.
| Gamma | 20-seed inventory mean P&L range | P&L std range | qT std range | What the sensitivity shows |
|---|---|---|---|---|
| 0.01 | 67.9566–69.0514 | 8.5249–9.1826 | 5.0995–5.4593 | Stable, small profit-level offset. |
| 0.10 | 63.8109–64.5711 | 5.8409–6.4439 | 2.7917–3.0603 | Stable and close. |
| 0.50 | 23.1349–23.7446 | 4.6184–4.9480 | 1.8806–2.0480 | The high-γ mean failure is systematic, not a seed-42 accident. |
The intensity uses abs(quote−mid). At γ=0.5, q=1, signed λdt=5.9315 becomes 0.08261, heavily suppressing the aggressive liquidation-side fill. Removing abs alone is insufficient: a Bernoulli probability above one needs an explicit policy.
The code uses σ=0.02 and produces a 0.00004 spread difference; the report discusses the 0.40 effect that would arise under σ=2.
The helper uses population rather than the sample standard deviation specified by the QuantBrain-visible frozen plan; Qwen prompt equivalence is unknown. Fresh γ=0.01 inventory std qT=5.04, while the narrative says 5.30.
Findings outside the headline experiment
strategies.py computes reservation price and ignores it, uses a non-paper half-spread, reverses inventory-skew signs, and can clamp quotes to an arbitrary 0.001 distance.simulator.py uses intensity directly as Bernoulli probability without multiplying by dt, fills every advertised path row with the final state, and double-counts inventory value in mark-to-market.RESULTS.md calls σ=0.02 and A=1 paper-matching.src/monte_carlo/… paths and uses 21 green tests as methodology validation.src/ package and
metrics.py, lists absent CSV outputs, claims a missing MIT
license, and calls lower-bound dependencies “pinned.” Its repository is
operationally healthier, but that does not remove the code defects.
QuantBrain’s structurally valid CSV/JSON files and 20 decodable PNGs show
artifact production, not inferential validity.
Qualified artifact-level judgments
| Dimension | Stronger inspected artifact | Why |
|---|---|---|
| Central numerical reproduction | Qwen-attributed result | Correct target dynamics and calibration with paper-scale output |
| Coverage of the QuantBrain four-part object | QuantBrain 397B | Attempts all four visible topics; task equivalence remains unknown |
| Formula correctness | Qwen, qualified | Sounder formula module, but incomplete analytical coverage |
| Reusable quote/execution layer | QuantBrain, narrowly | Better quote direction and fill probability; wrong experiment calibration |
| Repository and supplied-test health | Qwen | Single history, compact tree, repeatably green tests |
| Test adequacy | Neither | One suite encodes a false property; the other omits broken public modules |
| Report integrity | Qwen, narrowly | Both overclaim; QuantBrain has more direct report/data contradictions |
| Production-ready support | Neither | Each retains material correctness defects |
The Qwen-attributed repository is the better starting point for repairing the central numerical experiment. QuantBrain’s broader architecture is useful, but its “complete reproduction” claim is unsupported by the implemented theory and simulation.
What QCA’s judge could and could not establish
The QCA trace used a 14-page uploaded preprint (SHA-256 7a53a062…8de) with γ={0.01, 0.1, 0.5} and displayed spreads 1.33 / 1.29 / 1.15.
A later eight-page NYU-hosted published scan (SHA-256 71fa9268…a8d) uses γ={0.01, 0.1, 1.0} and average spreads 1.35 / 1.49 / 3.02. The audit evaluates against the uploaded bytes and flags the mismatch as benchmark-design ambiguity.
Confidence: medium. This was directionally cautious, but it remained a report-level review.
Seven reads hit the 8,000-character ceiling. No equivalent Qwen judgment is auditable because its supplied run is absent from the public share.
runpod/quantbrain-code-397b-a17b
for job job-qca-runner-gigifko; the status field says
Succeeded. The root took 856.594 seconds, including 815.332
seconds waiting for the coding job. Inherited metadata instead says
runpod/quantbrain and reuses the 17 July thread ID. The generic
four-level judge rubric has no numeric score and no explicit tolerances for
this study’s P&L dispersion, terminal-inventory dispersion, spread, or
fill-rate quantities. Recorded cost is null, and the 94,470
tokens exclude tested-model generation.
| Trace | Job outcome | Scope | Visible verdict | Visible tokens | Why it is not a common baseline |
|---|---|---|---|---|---|
| 17 Jul · QuantBrain 397B | Failed | 4 experiments | Partial / High | 70,095 | Pipeline still judged files already on the mutable default branch. |
| 17 Jul · generic QuantBrain | Succeeded | 3 experiments | Partial / Medium | 52,020 | Different model string and task scope. |
| 13 Jul · Sonnet 4.6 | Succeeded | 2 experiments | Partial / Medium in prose | 40,084 | Different scope and smaller evidence read. |
| 19 Jul · QuantBrain 397B | Succeeded | 4 experiments | Partial / Medium | 94,470 | Corrected primary QuantBrain run; tested-model exchange remains opaque. |
The failed 17 July job received the same PARTIALLY REPRODUCED verdict with higher stated confidence than the later successful job. Job success and generated-commit identity must be hard evaluation gates.
The source is Carlo Nicolini, Matteo Manzi, and Hugo Delatte, “skfolio: Portfolio Optimization in Python.” It defines three workflows: a constrained minimum-variance frontier, a five-factor prior, and a stacked portfolio evaluated through walk-forward splits.
| Captured article version | Bytes / pages | SHA-256 | Role in the audit |
|---|---|---|---|
| skfolio paper · arXiv v1 | 794,465 / 7 | d9ab3aea73486cdbe606e28cdb3ed49a790fa8e6b873ee3d37a518794e5cc1bf | Earlier captured version |
| skfolio paper · arXiv v2 | 800,224 / 7 | b04c86f6ff4b57406480271dee6705a8dc742af4bf6b9e2095692e5d9ae2fd8f | Current well before the run and highly likely, but not byte-pinned by QCA |
MeanRisk prior with Ridge(alpha=0.1); and a
252/60 held-out walk-forward comparison that stacks inverse volatility,
MeanRisk/CVaR, and default hierarchical risk parity against equal weight.
It does not require stacking to outperform equal weight or define fixed
metrics for future data and dependency versions.
Held-out prediction tables, loadable population objects, and JSON metrics/temporal validation are committed and internally reproducible.
The evaluator accepts train and test data but ignores both. It measures the same train-backed population twice and labels one copy “test.”
This structure is sounder, but 21/100 nearest-risk comparisons and train_dominates_test=false are narrated as confirmation.
A wrong function signature raises TypeError before Experiment 3. Named result files are absent from the repository.
Developer-attributed artifacts
| Property | Qwen-attributed artifact | QuantBrain-attributed artifact |
|---|---|---|
| Pinned commit | fd217ac9e205f0102d2a29c5d189daaf45806bab | 274681ced761a5d952ac7a1c5e9520bf81a7c569 |
| Tracked files / bytes | 39 · 1,053,483 bytes | 14 · 109,024 bytes |
| Committed result evidence | 23 artifacts plus RESULTS.md | .gitkeep plus RESULTS.md |
| Supplied tests | 124/124 under current resolution | 27/28 current; 28/28 compatible, with exception-message drift |
| Workflow exercised by tests | Experiment functions covered; no semantic oracle for use of X_test | Experiment entry points and all-run command are not called |
| Dependency record | Partial package/version JSON; no exact lock | Narrative claim only; no exact lock |
| Public trace | Requested Qwen 397B and job status reported Succeeded | Supplied run returns HTTP 404 in supplied share |
100 portfolios, chronological 80/20 split
The committed train and “test” files are byte-identical. Correct test prediction changes both expected return and variance.
The source records train_dominates_test=false after 21/100 nearest-risk comparisons, yet the static report calls the experiment successful.
Risk interpolation gives 73 comparable points, 16 with train at least as high, and mean annualized train-minus-test gap −0.000571.
| Frontier-wide mean | Qwen committed train | Qwen committed “test” | Qwen corrected test | QuantBrain fresh train / test |
|---|---|---|---|---|
| Daily expected return | 0.000758524 | 0.000758524 | 0.000781840 | 0.000758525 / 0.000781842 |
| Daily variance | 0.000142454 | 0.000142454 | 0.000146872 | Separately evaluated |
| AAPL cap violations | 0 | 0 | 0 | 0 |
The saved red and blue train/test plot series lie exactly on top of one another, while the narrative claims a visible quantitative gap. The plot also uses daily expected return on the x-axis and daily variance on the y-axis; the paper uses annualized standard deviation on x and annualized mean return on y. The optimizer construction is plausible, but the visualization and OOS claim are not faithful.
The static report claims train annualized mean 18.20–25.89% and test 10.86–12.95%. Fresh committed source produces 18.20–20.03% and 19.10–20.30%. Reported risk ranges are likewise stale or inconsistent.
Five factors, Ridge α=0.1
prior_estimator_, and a 20×5 loading matrix is saved.prior_attached=false is a validator false negative because it checks the output population rather than the optimizer.fit_intercept=False differs from the paper’s displayed default; measured weight impact is small (L1 0.000319, maximum component 0.0000595).predict(X_test, y_test) uses an invalid signature.solver_status is interpreted as failed integration.Strongest Qwen evidence; clearest QuantBrain scope failure
| Saved metric | Qwen stacking | Qwen equal weight | Observed trade-off |
|---|---|---|---|
| Daily mean return | 0.000471409 | 0.000771023 | Equal weight higher. |
| Annualized volatility | 0.179593 | 0.204180 | Stacking lower. |
| Sharpe ratio | 0.661469 | 0.951599 | Equal weight higher. |
| CVaR · positive loss magnitude | 0.027435 | 0.030968 | Stacking lower. |
| Max drawdown · skfolio convention | 0.309954 | 0.346955 | Stacking lower. |
| Simple geometric drawdown · independently recomputed | 0.287423 | 0.316756 | Stacking lower under this second convention. |
Saved total_return | 0.650545 | 1.064012 | Actually a non-compounded cumulative sum. |
| Independently compounded gain | 0.754223 | 1.584426 | 75.42% versus 158.44% gain. |
The paper does not require stacking to outperform equal weight. Lower return and Sharpe are outcomes, not reproduction failures; the relevant question is whether the method and held-out evaluation were reproduced.
cv=None produces default five-fold K-fold, matching the paper’s omitted/default setting but remaining a non-temporal caveat.train_end < test_start.results_dir to a no-argument function and aborts with TypeError before Experiment 3.Green tests do not answer the central semantic questions
X_test.Confidence: medium. The accessible judge correctly saw duplicate tables and a contradictory prior flag, but did not inspect source or execute code.
Visible root usage comprises 39,074 prompt and 7,137 completion tokens. It excludes the external Qwen repository job and PDF analysis, and the trace records no complete cost.
runpod/qwen3.5-397b-a17b.job-qca-runner-p8p92jz reported Succeeded.Qualified artifact-level judgments
| Dimension | Stronger inspected artifact | Why |
|---|---|---|
| Recoverable evidence | Qwen-attributed result | Persisted predictions, objects, tables, and plots |
| Experiment 1 evaluation structure | QuantBrain-attributed result | Separate train/test predictions rather than duplicate outputs |
| Experiment 1 conclusion integrity | Neither | Qwen does not test out of sample; QuantBrain reverses its actual negative result |
| Experiment 2 evidence | Qwen-attributed result | Prior and loadings are recoverable despite a broken validation flag |
| Experiment 3 article fidelity | Qwen-attributed result | Correct outer held-out scope and 23 walk-forward periods |
| End-to-end runner | Qwen, qualified | QuantBrain’s runner aborts; Qwen’s expensive root workflow was not independently rerun |
| Complete reproduction | Neither | Both retain method, verification, and environment gaps |
The Qwen artifact is the more recoverable basis for continued work. QuantBrain’s separate train/test structure in Experiment 1 is a real strength, but its report, runner, and missing evidence require substantial repair.
Across both studies, QCA produced nontrivial but uneven repositories and structured outputs. The pipeline repeatedly allowed model-authored tests and summaries to stand in for external scientific verification.
Both studies contain multi-file scientific implementations, tests, reports, and machine-readable artifacts. Several workflows execute successfully.
The 17 July failed QuantBrain job was followed by a judge reading the live default branch and returning the same verdict at higher stated confidence.
One suite asserts the wrong spread property; another never verifies that “test” metrics use test data; reusable defects remain unimported.
Neither accessible final judge inspected source or executed tests. Adaptive file selection missed defects that direct source and numerical checks exposed.
| Control | Observed across the supplied studies | Required condition |
|---|---|---|
| Task identity | Missing comparator traces and unequal visible scopes | Hash-identical PDF, experiment JSON, presets, limits, and prompts |
| Repository identity | Live branches, absent generated SHAs, and one unrelated-history merge | Empty repository per run and immutable generated commit captured by orchestration |
| Job gating | Review can continue after external coding failure | Hard stop unless generation succeeds and the expected commit exists |
| Scientific verification | Model-authored tests and narratives dominate the evidence | Fixed external semantic tests plus independent metric recomputation |
| Judge evidence | Adaptive report reads; no source execution | Predetermined evidence manifest with source, logs, outputs, environment, and verifier results |
| Usage and cost | Visible GPT-5.4/QCA tokens exclude tested-model generation and some PDF work | Per-stage model, token, cache, retry, latency, and price records |
| Result semantics | One coarse “partially reproduced” label hides different failure modes | Separate scores for completion, method fidelity, numerical result, evidence integrity, and test adequacy |
The goal is to retain the useful repository-generation and paired-benchmark exercises while making every model response, execution result, and judgment independently reconstructible.
Pin exact article, task, rubric, and dataset bytes; package/solver images; prompts, presets, tool permissions, model revisions, judges, seeds, and limits.
Send every model the hash-identical task in a clean repository and clean conversation context.
Store every tested-model response, extracted code, usage field, finish reason, retry, tool call, execution log, timestamp, and server-returned identity.
Abort evaluation on failed generation, missing commit, dirty output, or mismatched task hash. Never fall through to an existing default branch.
Test formulas, stochastic process, calibration, train/test provenance, constraint satisfaction, fold geometry, and report-to-file consistency.
Calculate metrics from persisted raw outputs with independent code. Require every narrative claim to resolve to the same-run evidence.
Supply the same predetermined manifest to pinned reviewers, adjudicate disagreements, correct for planned multiplicity, and keep unlike score families separate.
Separate artifact completion, article fidelity, numerical behavior, evidence integrity, and test adequacy instead of collapsing them into one verdict.
This page synthesizes the principal findings from both QCA article-reproduction investigations. The linked Markdown and JSON records preserve exact commands, hashes, environments, file-level references, metrics, and verification limits.