gavel-gates

Honest labels: real sandbox, generated speech. Gates G0 to G4 ran against the real PayPal sandbox, never against live money. G5 is a cached evaluation of recorded provider outputs. G3 passed: the multi-day probe is finished and docs/GATES.md records PASS (final probe 12 Nov 2026).

What it is

A small machine-readable index of the six go/no-go experiments ("gates", G0 to G5) that decided which payment and voice features the project turns on: for each gate its status, date, method and honest label, the workflow runs and evidence files behind it, with a sha256 and a link for every evidence file; plus the per-spec results of the nightly sandbox runs.

It is an index, not a copy: the narrative and the evidence files live in the repository.

How it was made

• gates is parsed by scripts/hf/build-datasets.mjs from the status lines of docs/GATES.md; nothing is overridden in code. A gate whose status line says NOT RUN while its run line starts with IN PROGRESS (a probe still running) reads IN PROGRESS, and status_note quotes the run line; once GATES.md records a result, the dataset shows that result.
• evidence_files has one row for every file under docs/evidence (all extensions, all folders), hashed at build time. Where docs/GATES.md writes a sha256 next to a file name, matches_claim says whether it equals the computed hash; otherwise it is null.
• nightly_results reshapes the committed nightly reports (docs/evidence/nightly/NIGHTLY-<run id>.json).

Provenance

Built from commit 65be30bfccc88c8fb836d48f11432423460933f5 of https://github.com/harshbhx/gavel-auction by https://github.com/harshbhx/gavel-auction/actions/runs/43196464384. The same values are in MANIFEST.json.

MANIFEST.json lists the sha256 and size of every file and the row count of every table. python verify.py recomputes the hashes and fails on any difference.

• Every row of evidence_files gives the repository path, the sha256 and a link to the file at the build commit. Download the file from that link, run sha256sum, compare. matches_claim says whether the hash equals the one written in docs/GATES.md.
• Assurance rows carry a query_url; opening it in GitHub (with repository access) shows the same count as the row.
• Rows with workflow_run_urls point to GitHub Actions runs; open them (repository access may be needed and logs expire) to confirm date and result.
• The evidence files themselves are not copied here; they are committed redacted in the repository.

Fields

gates.jsonl / .csv

One row per gate (6 rows).

Field	Type	Meaning
gate	string	G0 to G5
title	string	Short title
status	string	PASS, FAIL, PARTIAL, NOT RUN or IN PROGRESS
status_note	string or null	Extra note when the status is derived (IN PROGRESS)
decision	string or null	Recorded decision keyword
date_utc	string or null	First date or timestamp on the gate's run line in docs/GATES.md (G3: start of the probe)
method	string	How the gate was run
label	string	Honest label: real sandbox, generated speech
workflow_run_urls	string	GitHub Actions run URLs, separated by semicolons
evidence_files	string	Evidence paths in the repository, separated by semicolons
gates_doc_url	string	Link to docs/GATES.md at the build commit

evidence_files.jsonl / .csv

One row per evidence file.

Field	Type	Meaning
gate	string	Gate (G0 to G5, G3-early belongs to G3) or the evidence folder (nightly, latency, measure, hosted, ...)
path	string	Path in the repository
bytes, sha256	integer, string	Size and sha256 at the build commit
sha256_claimed_in_gates_md	string or null	Hash written in docs/GATES.md, if any
matches_claim	boolean or null	Computed hash equals the claimed one
url_at_commit	string	Link to the file at the build commit

assurance.jsonl / .csv

Aggregate counts of the red-team and audit findings of the repository, from the GitHub issues API. Counts only: no issue title, text or author is stored. The query is pinned to issues (not pull requests) carrying at least one of the labels red-team, audit-r1, audit-r2 or redteam-r4; severity rows are for the red-team label and include an unlabelled row (red-team findings with no severity label), so they add up to the red-team count; area labels overlap (an issue can carry several), so area rows can add up to more than the population, and the unlabelled area row counts findings with no area label.

Field	Type	Meaning
dimension	string	overall, label, severity or area
key	string	Label, severity or area the row counts
found, closed, open	integer	Findings, how many are closed and how many are still open
closed_rate	float	closed divided by found
median_hours_to_close	float or null	Median hours from opening to closing of the closed issues; a proxy for time to fix, not a measured fix time
query_url	string	GitHub issue search that reproduces the count (repository access is needed to open it)
as_of_utc	string	Date the counts were taken

nightly_results.jsonl / .csv

One row per spec per nightly report.

Field	Type	Meaning
run_id, run_url	string	Nightly workflow run
generated_at, finished_at	string	Timestamps (UTC)
built_from_sha	string	Short commit the nightly ran on
spec, gate	string	Spec id and gate
status	string	pass, fail, inconclusive or skipped
observed	string	One-line observation

Intended use

Letting a reader check, in a few minutes, which claims about the payment and voice features rest on real sandbox runs, which on simulation, and what is still open.

Limitations

• Status lines are the maintainer's own record. The index proves that a file existed with a given hash at a commit, not that the file is true; for that, follow the run links.
• Evidence files without a sha256 in docs/GATES.md have matches_claim null: the hash is computed but there is nothing to compare it with. G5 also relies on recorded provider outputs and the fixtures of the other dataset.
• Sandbox only. Nothing here says how production PayPal behaves.
• G3 passed: the multi-day probe is finished and docs/GATES.md records PASS (final probe 12 Nov 2026).

License

CC-BY-4.0 (Creative Commons Attribution 4.0 International). You may copy, adapt and redistribute the data, including commercially, if you credit the project (see Citation). The licence covers the data files in this dataset only; it grants no right to the marks or services of PayPal, Deepgram, Groq or any other third party named in it.

Simulated or real

Labels: real sandbox, generated speech. Read the label note at the top of this card before using any number. Nothing here is production data, and no real person or real money is involved.

Citation

@dataset{gavel_gavel_gates_2026,
  title  = {gavel-gates},
  author = {Bhanushali, Harsh},
  year   = {2026},
  url    = {https://huggingface.co/datasets/harshbhx/gavel-gates},
  note   = {Generated from https://github.com/harshbhx/gavel-auction}
}

Links

• Source code, the build in scripts/hf and the evidence: https://github.com/harshbhx/gavel-auction
• Project site: https://gavel.dualmindlab.tech
• Live demo (start a room in the browser, sandbox payments): https://gavel.dualmindlab.tech/demo
• Documentation: https://docs.gavel.dualmindlab.tech
• Documentation for this data: https://docs.gavel.dualmindlab.tech/docs/judges/open-data
• Evidence files on main: https://github.com/harshbhx/gavel-auction/tree/main/docs/evidence
• Gate verdicts on main: https://github.com/harshbhx/gavel-auction/blob/main/docs/GATES.md
• Evidence Explorer (charts from these datasets): https://huggingface.co/spaces/harshbhx/gavel-evidence-explorer

Contact

Open an issue at https://github.com/harshbhx/gavel-auction/issues. The code that builds this dataset is in the scripts/hf folder of that repository.