reed-sample-500 — start here

500 verified behavioural records of real code changes, from 141 open-source Python and JavaScript repositories. Free sample. Every number on this page is quoted from a file in this folder.

What this is

Each row is one merged pull request, executed. We rebuilt the environment the change originally ran in (the interpreter and the dependencies as they were in the month it merged), ran small generated test programs ("probes") before and after the change, and recorded exactly what changed in the output. Then we found the smallest set of the PR's hunks that causes that change (the causal core; the rest of the PR is labelled inert on these probes) and tested whether removing only that core from today's version of the repository removes the behaviour cleanly.

Most pipelines that turn Git history into training data keep only what runs under today's libraries and discard the rest. This one rebuilds the past instead, and measures what each change actually did.

Key numbers

		Source
Blind re-execution on fresh machines: per-probe verdicts identical	99.85 % (10,408 / 10,424)	evidence/E77-RESULT.md
Blind re-execution: recorded delta outputs identical	100 % (5,960 / 5,960)	evidence/E77-RESULT.md
Cross-machine (Linux vs macOS) verdict agreement	99.76 % over 4,501 probes	evidence/E59-RESULT.md
Changes executable in their own era's environment vs today's (interim population count)	78.5 % vs 8.6 %	evidence/E82-PREREG.md
Removing a merged feature cleanly at today's HEAD: causal core vs git revert (301 features, 2 repositories)	54.8 % vs 28.6 %	evidence/PAPER-A-UNMERGE.md

What's in each row

The repository, PR and both commit SHAs · the environment specification (interpreter, declared requirements, package index cutoff, exact pins where they re-resolve) · the probes · raw and normalized outputs before and after · the causal core as a patch · the removal verdict at today's HEAD beside the git revert baseline · a reproducibility receipt where one exists · every source file the row was read from.

Verify any row yourself with only git and Python — no code of ours needed. See reproduce.md.

Read this before you compute anything

• Every row is a behaviour change (every probe of a layer moved, by construction). The full population also records the probes that did not move for each PR; ask if you need no-change examples.
• Causal cores are hunk-level, not line-level.
• "Inert" means inert on this row's probes (median 5 per row). Code the probes never execute is unmeasured.
• Some deltas are documentation or version edits: 14 rows' probes read only docstrings or version strings, 16 rows mix these with runtime probes. Filter with the delta_kind field.
• The removal verdicts are skewed on purpose: 235 of 500 rows (47 %) are not clean at HEAD, because clean rows ran out in some age buckets under the five-per-repository cap. Don't read the sample's clean rate as a population rate.
• Python 493 rows, JavaScript 7. Ages: under 1 year to over 8 years (only 17 rows over 8 years).
• Limits and prior art are stated plainly in README.md §3.

Folder map

rows.jsonl the 500 rows · SCHEMA.md every field · reproduce.md verify a row · stimuli/ probes · outputs/ before/after outputs · cores/ causal-core patches · environments/ environment specifications · evidence/ the result files quoted above · full documentation below (also README.md) · LICENSES.md licensing (dataset CC-BY-4.0; GPL-family repositories held out) · PROVENANCE.md, BUILD-LOG.md how the set was built, corrections kept.

Beyond the sample

The population behind this sample is about 20,000 executed layers across 241 projects. We also run the pipeline on repositories you choose:

1. Reachability audit — send up to 200 repositories; in two weeks, learn what share of their history is executable, under which environment, why the rest isn't, and receive 100 sample rows from your list.
2. Pilot — 1,000 verified rows on your repositories in six weeks.
3. Ongoing supply — verified rows every month on your target languages, repositories and era range.

Contact: Advaith Gunna · Unorganized Machines Inc. · advaith@unorganizedmachines.com

---

Full documentation

An evaluation set of 500 executed behavioural records of merged pull requests, drawn by seed from Reed's population runs. Built 2026-09-22T16:42:38Z at Reed commit dbc7291, while the population was still closing (278 of 419 drawn repositories complete, 33 in flight, the uniform re-ladder not finished). Every field in every row was read from a recorded artifact; nothing is fabricated. Selection seed 500; capture date 2026-09-22. Documents revised 2026-09-22/23 for a reader who has this folder and nothing else.

1. What a row is and how it was produced

A row is one layer: a merged feature of one pull request, defined by the generated probes whose behaviour that PR changed.

For each pull request we checked out the merge's parent (the before-world) and the merge itself (the after-world), and rebuilt the environment of the change's era: dependencies resolved against the package index as of the merge month, and an interpreter found by execution from a ladder (CPython 3.13 down to 3.5; Node 22/20/18 for JavaScript), era winning ties. A model-generated battery of small programs — the stimuli — ran in both worlds. A versioned comparator normalized both outputs (timestamps, paths, addresses, line numbers cancelled; version recorded per row) and judged each probe delta / no_delta / junk / flaky. Delta debugging over the PR's hunks found the minimal set of hunks whose removal extinguishes each probe's delta — the causal core — composed per layer. That core was then reverse-applied at the repository's current HEAD and the battery re-run, recording whether the feature could be removed cleanly, beside the same test for git revert -- '*.py'.

The row ships the stimuli, both raw and normalized outputs per probe, the environment specification (interpreter, declared requirements, index cutoff, and exact pins where the declared set re-resolves — see §5), the core patch, and the removal verdict at HEAD beside the git-revert baseline.

Selection (selection.json): 3189 eligible layers in 193 repositories after the eligibility rule (all of: before/after sha, merge date, manifest, stimuli, both outputs, layer core, HEAD verdict) and the licence hold-out; 500 drawn with seed 500, at most 5 rows per repository (141 repositories); age buckets, target ~100 each: <1y 100, 1-3y 100, 3-5y 148, 5-8y 135, >8y 17 — the >8y bucket had only 17 eligible rows under the five-per-repository cap, so its deficit was filled from 3-5y (48) and 5-8y (35); the honest count of rows older than eight years is 17; rows whose core is not clean at HEAD: 235 (47 % — the build asked for ~20 %; clean rows at 3–8 years run out under the five-per-repository cap, so the excess is failures, counted per bucket in selection.json); environment that won: era 438, today 55, unrecorded 7 (see §3 on what "won" means); today's environment evaluated at all on 213 rows; 12 rows from E77's blind re-execution projects; 7 JavaScript rows (koa, ajv); sampling rule of the HEAD verdict: stable 456, seeded 44; 413 unique environment manifests.

2. How to verify a row without our code

Everything needed is in this folder. reproduce.md walks one row: clone the public repository, check out before_sha, build the environment from environments/<manifest_hash>.json (interpreter, declared requirements, index cutoff; exact pins where they re-resolve), run the row's probes with run_stimuli.py (each probe in a fresh interpreter, as we ran them), repeat at after_sha, and compare your raw outputs to outputs/<row_id>.before.jsonl and .after.jsonl. Then reverse-apply cores/<row_id>.patch at HEAD and re-run to check the removal verdict. No Reed code is required. How the set itself was built is in PROVENANCE.md and BUILD-LOG.md.

3. What this is NOT

• Execute-and-compare is prior art. ChangeGuard (Gröninger, Souza, Pradel, FSE 2025, doi:10.1145/3715760) executes the pre- and post-change versions of a modified function under learning-guided execution and compares their runtime behaviour, reporting a precision of 77.1% and a recall of 69.5% for semantics-changing changes. Testora (Pradel, ICSE 2026, doi:10.1145/3744916.3764527) generates tests for a pull request with an LLM, executes both versions, compares outputs by exact string equality, and classifies the differences against the PR's stated intent with an LLM (classifier precision 55%, recall 67%). Kessel (preprint 2604.16933 on arXiv, April 2026; not our work) proposes a paradigm that "couples the Git history with a Behavioral Archive". What this set adds is the record at population scale with the environment inside it and the removal verdict — not the idea.
• The verifier's recall is below a mature test suite. E49 (evidence/E49-RESULT.md): the two-sided declared-delta certificate rejected 56/92 = 60.9% of undeclared changes while the project's own suite alone rejected the same changes at 65/92 = 70.7% — a NULL result by its pre-registration.
• Reach is the probes' execution path. E43 (evidence/E43-RESULT.md): of same-file mutants the project's suite misses, the emitted guards killed 6/89 = 6.7% [3.1–13.9], and on code the feature's probes never execute 0 / 185 = 0.0% [0.0–2.0]. A row's outputs describe behaviour on its recorded stimuli and nowhere else.
• env_chosen is which environment WON, not which could run the change. The executor tries today's environment only when the era's fails to run the whole battery (E82, evidence/E82-PREREG.md); today_evaluated says whether it was tried.

4. Trust numbers

• Blind re-execution (evidence/E77-RESULT.md): on fresh nodes, per-(PR, probe) verdict identity 10,408 / 10,424 = 99.85 %, delta outputs identical 5,960 / 5,960 = 100 %, exact removal cores 4,584 of 4,585; the ladder's per-trial verdicts follow the sibling sample (V5 88.0 % pooled, FAIL as pre-registered). Rows from the eight re-executed projects carry the project's V1 identity in reproducibility; the rest are null with the reason (receipts.jsonl).
• Cross-machine agreement (evidence/E59-RESULT.md): the Linux executor returned the macOS box's verdicts at 99.76% pooled over 4,501 probes of pygments and cattrs; the platform accounted for 4 of 4,501, all flaky-class.
• The population's ladder rows were deduplicated by layer id before the draw (shards had re-scored finished layers; the duplicates were content-identical).

5. Known limitation of the environment record, found while writing the recipe

Each manifest carries the dependency listing the population run recorded (resolved_deps). That listing was captured from the shard's most recently built prefix, not necessarily the prefix that executed the row (defect D77, 2026-09-22): 217 of the 406 era manifests list a package version newer than the manifest's own index cutoff. The listing is kept for the record and must not be installed from. The reproducible specification is the interpreter, the declared requirements and the cutoff — the three inputs the executor installed from. Re-resolving them under the cutoff with uv gives exact, installable pins for 171 of the 413 manifests (reresolution.pins); for 200 the declared requirements are unsatisfiable under the cutoff as one set, which is why the executor installed each requirement on its own and skipped the ones that failed — the recipe does the same; 13 need an interpreter version the compiler cannot resolve for; 7 have no declared requirements; 22 failed for other reasons. Counts and the tool version: environments/RESOLUTION-SUMMARY.json.

6. The rest of the folder

• LICENSES.md — frame licence per repository, SPDX mapping, the 69 GPL/AGPL/LGPL repositories held out pending a decision, the two weak-copyleft repositories included and flagged, the derived-data statement, and the licence of the dataset's own part (CC-BY-4.0).
• SCHEMA.md — every field of rows.jsonl and of the manifest files; its source column is provenance, not a dependency.
• PROVENANCE.md — how we produced the rows (our pipeline's own commands). BUILD-LOG.md — the build's log, with the wrong-then-corrected entries kept. field-sources.tsv, completeness-receipt.json — the build's own audit.
• evidence/ — the result files cited above and the two paper drafts they belong to.