com-kotobalabs/typed-decisions-repo-governance

typed-decisions: repo-governance family

A seed, not a corpus (n=1 as of first publish, 2026-09-19): real governance findings from a private software monorepo, each run through TypeSafe's Jev (a "System One" model — it returns typed choice / score / noul decisions, never free text or code) as part of kotoba-lang/typed-decisions, a reproduction and measurement of the Jev shape whose trained model is published at com-kotobalabs/open-jev-deberta-v3-large.

What this is, precisely

Every other family in typed-decisions (banking77, sst5, boolq, mined .cljc fix pairs, Hermes transcripts) carries an external gold label — a dataset's own annotation, or a runnable test that passes or fails. This family does not: the findings recorded here are open, unresolved governance questions inside the source monorepo (e.g. "does a Cloudflare Worker spanning two domains violate a compliance boundary?") that nobody has adjudicated yet.

So:

• Every questions[].gold is null. There is no ground truth here yet.
• The model's answer is recorded under prediction only, and must never be read as gold. This mirrors a rule typed-decisions already established for its teacher-model labels: a prediction is not gold just because a capable model produced it.
• gold_status on every record says so explicitly: "unresolved -- live jev prediction, not training gold, until independently resolved".

Do not train directly on prediction as if it were gold. This dataset exists so that, once a finding is actually resolved (by a human, or later by a stronger model whose disagreement rate against real gold is separately measured — see the source repo's teacher-labeling discussion), the resolution can be written back as gold and the record becomes usable exactly like any other family here.

Schema

One JSON object per line (repo-governance.jsonl):

{
  "state": "free text describing the finding",
  "source": "workspace-repo-governance:<detector-name>",
  "collected_at": "ISO-8601 timestamp",
  "meta": {"detector": "...", "repo": "...", "finding_id": "..."},
  "questions": [
    {"qid": "...", "kind": "choice|score|noul", "instructions": "...", "options": [...], "gold": null}
  ],
  "prediction": {
    "model": "typesafe/jev-1.13-<date>",
    "provider": "TypeSafe",
    "id": "gen-dec-...",
    "answers": {"<qid>": {"...typed answer, probabilities, confidence..."}},
    "usage": {"input_tokens": 0, "output_tokens": 0, "cost": 0.0}
  },
  "gold_status": "unresolved -- live jev prediction, not training gold, until independently resolved"
}

kind: "noul" always has options: ["no", "yes"]; "choice" options come from the question's criteria keys; "score" options are the criteria's ordered level descriptions.

Provenance and content note

Records are generated by scripts/jev-decide.cljk --record in the source monorepo (com-junkawasaki, private) against real findings surfaced by that repo's own automated detectors. The first record concerns a real (already-known, not newly disclosed) architectural detail — that two subdomains of a real production service are served by a single Cloudflare Worker — published here with the repo owner's explicit approval. This is not a security advisory and contains no credentials, keys, or personal data; it is a compliance-scoping observation, included for reproducibility of the Jev-shaped decision pipeline, not as a vulnerability report.

License

Apache-2.0, matching com-kotobalabs/open-jev-deberta-v3-large.

Related

• Source: https://github.com/kotoba-lang/typed-decisions (branch codex/repo-governance-decisions, PR #1)
• Sibling model: https://huggingface.co/com-kotobalabs/open-jev-deberta-v3-large
• Jev (the reference point, not a spec this reproduces): https://typesafe.ai

Publisher

Published by Mithril (formerly Kotoba Cloud), operated by Kotoba Labs Inc. The Hugging Face handle com-kotobalabs is unchanged. Support: support.mithril.fund.