openrepro/openrepro-ledger

OpenRepro Ledger

OpenRepro re-runs published computational research and publishes hash-chained evidence of whether the results hold. This dataset is a mirror of the ledger's machine-readable files from openrepro.org. The site is the source of record. Every value here comes from an attempt's report.json and is the same as on the site.

Release 2026-11-12. 15 papers and 125 registered rows (123 counted claims, 107 of them held within tolerance; controls, gates and the claims of an attempt that gives no outcome are not counted), from 25 released attempts (10 of them superseded versions). Only attempts listed in release.json, the release manifest, are included; a paper without a due release entry is absent.

arXiv	Paper	Outcome	Headline claims held	Qualifiers	Assurance	Report
1506.02142	Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning	Partially reproduced	10 of 12	10 other checks did not hold; a scored run is not the direct source of any recorded value: it may be an upstream stage whose output a later scored job read, or a run that was not used (or more scored runs started than the sealed plan declares); the report names each one	Sealed, bound and re-derived	report
1606.00931	DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network	Reproduced	4 of 4	some claims were registered after command output had been observed; legacy attempt: the plan was never sealed; claims were frozen only by their first result; at least one value was read from an excerpt the agent chose after seeing the output, not by a rule fixed before it	Legacy	report
1607.01759	Bag of Tricks for Efficient Text Classification	Reproduced	8 of 8	legacy attempt: the plan was never sealed; claims were frozen only by their first result; at least one value was read from an excerpt the agent chose after seeing the output, not by a rule fixed before it	Legacy	report
1609.02907	Semi-Supervised Classification with Graph Convolutional Networks	Inconclusive: a registered validity/precision condition was not satisfied	0 of 1	none	Sealed, bound and re-derived	report
1610.01271	Generalized random forests	Reproduced	3 of 3	every headline check is a one-sided bound, not a match to the published value; 1 other check did not hold; a scored run is not the direct source of any recorded value: it may be an upstream stage whose output a later scored job read, or a run that was not used (or more scored runs started than the sealed plan declares); the report names each one	Sealed, bound and re-derived	report
1708.05031	Neural Collaborative Filtering	Reproduced	4 of 4	some claims were registered after command output had been observed; legacy attempt: the plan was never sealed; claims were frozen only by their first result; at least one value was read from an excerpt the agent chose after seeing the output, not by a rule fixed before it	Legacy	report
1711.10561	Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations	Reproduced	1 of 1	Burgers error 3.21× the published value; passes a registered 10× ceiling; a stricter 2× check did not hold; legacy attempt: the plan was never sealed; claims were frozen only by their first result; at least one value was read from an excerpt the agent chose after seeing the output, not by a rule fixed before it	Legacy	report
1803.09015	Difference-in-Differences with multiple time periods	Reproduced	4 of 4	some claims were registered after command output had been observed; legacy attempt: the plan was never sealed; claims were frozen only by their first result; at least one value was read from an excerpt the agent chose after seeing the output, not by a rule fixed before it	Legacy	report
1810.08473	From Louvain to Leiden: guaranteeing well-connected communities	Reproduced	4 of 4	DBLP Leiden gain over Louvain 0.938× the published value; passes a registered floor of 0, a sign check; Amazon Leiden gain over Louvain 1.04× the published value; passes a registered floor of 0, a sign check; legacy attempt: the plan was never sealed; claims were frozen only by their first result; at least one value was read from an excerpt the agent chose after seeing the output, not by a rule fixed before it	Legacy	report
1812.09970	Synthetic Difference-in-Differences	Reproduced	4 of 4	legacy attempt: the plan was never sealed; claims were frozen only by their first result; at least one value was read from an excerpt the agent chose after seeing the output, not by a rule fixed before it	Legacy	report
1905.03222	Conformalized Quantile Regression	Reproduced	5 of 5	legacy attempt: the plan was never sealed; claims were frozen only by their first result; at least one value was read from an excerpt the agent chose after seeing the output, not by a rule fixed before it	Legacy	report
1910.13051	ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels	Reproduced	3 of 3	some claims were registered after command output had been observed; legacy attempt: the plan was never sealed; claims were frozen only by their first result; at least one value was read from an excerpt the agent chose after seeing the output, not by a rule fixed before it	Legacy	report
2004.01105	Inferring change points in the spread of COVID-19 reveals the effectiveness of interventions	Reproduced	4 of 4	legacy attempt: the plan was never sealed; claims were frozen only by their first result; at least one value was read from an excerpt the agent chose after seeing the output, not by a rule fixed before it	Legacy	report
2012.07436	Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting	Partially reproduced	2 of 3	every headline check is a one-sided bound, not a match to the published value; some results were left untested for budget; legacy attempt: the plan was never sealed; claims were frozen only by their first result; at least one value was read from an excerpt the agent chose after seeing the output, not by a rule fixed before it	Legacy	report
2205.13504	Are Transformers Effective for Time Series Forecasting?	Reproduced only with earlier code	8 of 8	none	Sealed, bound and re-derived	report

Files

File	Contents
ledger.csv (config ledger)	One row per listed paper.
ledger.json	The same records under papers, plus an envelope: schema_version, version (the build date), generated, license, documentation, citation, feed, mcp, papers_listed.
claims.csv (config claims)	One row per registered claim of every listed paper.
claims.json	The same records under claims, with the same envelope.
release.json	The release manifest: every released attempt (current or superseded) with its release entry and the SHA-256 of its report files and evidence bundle. Nothing else is in these files.

In the CSV files, list fields such as authors are joined with "; ". Empty cells are nulls.

from datasets import load_dataset
ledger = load_dataset("openrepro/openrepro-ledger", "ledger", split="train")
claims = load_dataset("openrepro/openrepro-ledger", "claims", split="train")

Ledger fields (one record per paper)

Field	Meaning
arxiv_id	The paper's arXiv id, without a version suffix. It is the stable key for a paper in every file and on the site (/papers/<arxiv_id>/).
title, authors, year, venue, doi	Citation metadata from the report. Display-only corrections are sourced in the site's editorial file.
field	A browsing label assigned by OpenRepro.
outcome, outcome_label	The computed outcome, as a key and as printed: reproduced, reproduced_with_earlier_code, partially_reproduced, not_reproduced, evidence_incomplete ("Not verified: evidence incomplete": OpenRepro's evidence failed, not the paper), inconclusive ("Inconclusive: a registered validity/precision condition was not satisfied", or the engine's own wording of it: a sealed gate was not satisfied or not tested; "Inconclusive: the test was not precise enough" only when every unmet gate is a precision gate; not a finding about the paper) or not_attemptable.
qualifiers, qualifier_text, qualifiers_source	Qualifier codes (bound_checks_only, broad_bound, other_checks_failed, untested_for_budget), the same in plain words (show it next to the outcome), and whether they come from the report or were derived by the site from a legacy report.
role_disclosure	The count of agent-designated controls and gates, as printed beside the outcome on the site (for example "1 agent-designated control excluded; 1 gate"); empty when the report registers neither. Controls and gates are designated by the AI agent that ran the attempt, not reviewed; each is in the claims file with its role. Show it next to the outcome.
coverage_note	What the report says, or cannot say, about results left untested for budget.
assurance_level (assurance in JSON)	Legacy, Partial, Sealed plan, Sealed and bound, or Sealed, bound and re-derived. The JSON has the plan seal, evidence binding, execution manifest, re-derivation and engine version.
attempt_id, superseded_attempts (attempts in JSON)	The released attempt, and the lineage: every released attempt of the paper with its status (released or superseded), outcome, superseded_by and reason.
release_on, report_json_sha256	The release date from the release entry, and the SHA-256 of the released report.json.
headline_held, headline_held_earlier_code, headline_total	How many headline claims held, how many of those held only with a declared earlier version of the code, and how many headline claims were scored.
claims_checked	Every claim of the paper registered, headline or not (controls and gates are in the claims file with their role, not counted here).
checked	The date the attempt ended (UTC).
page_url, evidence_url, badge_url	The report page, the full evidence report, and the SVG badge.
log_head_hash, log_events	The head hash and the number of events in the attempt's hash-chained log. The report page prints this as the seal.
bundle_url, bundle_sha256	The downloadable evidence bundle and its SHA-256, when published.
summary	The report's summary paragraph.

Claims fields (one record per claim)

Field	Meaning
arxiv_id, claim_id	The paper and the claim. Together they are the stable key for a claim; the report page anchor is #claim-<claim_id>.
paper	The paper's short name.
headline	True when the claim counts toward the outcome.
role	claim (a claim of the paper), control (a registered arm expected to differ, excluded from the scored claims and the counts: a control's numerical success or failure does not determine the reproduction outcome) or gate (a validity/precision condition on the test itself, satisfied or not satisfied; one not satisfied makes the outcome inconclusive, and its published value is the gate's criterion). Controls and gates are designated by the AI agent that ran the attempt and are not reviewed; their rationale is on the paper page.
statement, location, metric, unit	What was registered: the claim in words, where it is in the paper, and what was measured.
comparison, published, tolerance, accepted_if	The rule registered before any result was recorded: within a band, at most, or at least; the published value; the tolerance; and the rule as printed.
check_kind, broad_bound, ratio_to_published	Equivalence or bound; whether the bound admits more than twice (or less than half) the published value; observed ÷ published. From the report when present, else derived by the site.
attempt_id	The attempt the claim belongs to.
observed	The value obtained by re-running the authors' code. Null when untested.
verdict	consistent, consistent_earlier_code, inconsistent or untested.
counted	True when the claim is in the site's tallies: a claim of the paper in an attempt whose outcome gives a verdict. False for controls and gates, and for every claim of an inconclusive or not-verified attempt, whatever its verdict.
earlier_code_observed	The value under the declared earlier-code fallback, when one was tried.
registered_event, first_result_event	The log events at which the claim was registered and its first result was recorded, so the order can be checked.
deviation_over_tolerance	(observed − published) ÷ tolerance. Null when untested; inf or -inf when the tolerance was zero and the value missed.

How it is produced

For each paper, an AI agent (Claude, made by Anthropic) works in a sandbox with OpenRepro's engine:

1. It takes the authors' own code and data.
2. Before results are recorded, it registers the claims to check, each with a quote from the paper and a numeric tolerance. In current (legacy) attempts this does not show that no output had been seen first; each record's assurance says what applies.
3. It rebuilds the environment as of the paper's date and re-runs the experiments.

The commands the harness runs, their captured output and the agent's decisions are events in a hash-chained log. The engine computes the verdicts by fixed rules from the registered claims, tolerances and headline choices, which the agent makes; the qualifiers show when a headline check is a broad bound or another check missed. The report is generated from the log; the one-file verifier checks log and artifact integrity and the report's head reference, not the report's contents. A person oversees the project: choosing papers and setting the rules the agent works under. The full method is at openrepro.org/methods/. The evidence bundles and a one-file verifier are at openrepro.org/verify/.

Limits

• AI-run, not independently reviewed. The re-runs are done by an AI agent. OpenRepro has not arranged independent domain-expert review, and nothing here is peer reviewed. Authors can respond; a published response does not establish endorsement or independent validation.
• An outcome is not a judgment of the paper. It says what happened when the authors' code was re-run under the recorded conditions. "Could not reproduce" means the published artifacts did not produce the published numbers under OpenRepro's protocol. It is not a finding about the authors' honesty or the idea's merit.
• Release is fail-closed. For any outcome other than "reproduced", the authors get 14 days to reply before publication, and every attempt needs an explicit release entry. Without a due one, the paper is absent from these files and from the site; release.json lists exactly what is included.
• Small and selective. The ledger covers only the papers OpenRepro has chosen and finished. It is not a sample of any field and should not be used to estimate a field's reproducibility rate.
• Snapshot. Each version is the ledger as of its build date. A correction is published as a new attempt that links to the old one; ids never change.

Update cadence

The files are rebuilt with the site whenever an attempt's release entry falls due. This mirror is updated after the site, so it can lag by a few days. version in the JSON envelope is the build date. The live files are at openrepro.org/data/.

License and citation

The reports, verdicts, evidence logs and these files are published under CC BY 4.0. Attribute "OpenRepro, https://openrepro.org" and link the report URL so a reader can reach the evidence.

@misc{openrepro_ledger,
  author       = {{OpenRepro}},
  title        = {{OpenRepro}: a public ledger of re-run computational research},
  year         = {2026},
  howpublished = {\url{https://openrepro.org/}},
  note         = {Dataset, version 2026-11-12. CC BY 4.0. Mirrored at https://huggingface.co/datasets/openrepro/openrepro-ledger}
}

The site's CITATION.cff has the same details. Each report page has its own citation block.

Contact and right of reply

Write to hello@openrepro.org to reply to a report or to report an error. Authors may reply at any time. Replies are published with the author's permission, in full or in the parts they choose, and any omission is marked.