Famwin/tonesoul-accountability-traces

ToneSoul Accountability Trace Dataset|語魂責任痕跡資料集

From Long-term Human-AI Dialogue to Auditable Tone Governance 從長期人機對話到可稽核語氣治理

Not a dialogue corpus — an accountability-event corpus. Existing corpora label content (what a text is); every record here is a closed loop — claim → bearing → verification → consequence — with labels earned by real events inside a working AI-governance repository, not assigned by annotators afterward.

Dictionaries record how people talk about honesty. Ledgers record what happened after they talked.

Composition (v0)

trace_type	count	what it is
counter_evidence	21	the public error ledger — claims that were challenged, incl. held:false entries where the AI (or the framing itself) was caught wrong
declared_stance	1	a judgment record that lays out tension without ruling
signed_commitment	296	every commit signed with Agent: + Trace-Topic: trailers — named bearing, git-anchored
total	318	2026-02-10 → 2026-07-04, single long-term human-AI collaboration

Read this before using (from the DATASHEET)

• Small by design. 318 records. The value is density and provenance (every record traces back to a file/commit/PR in the public repo), not scale.
• Single-dyad bias. One owner (Fan-Wei Huang), Claude-family agents, a codex external eye. A case history, not a population sample.
• Event-truth ≠ content-truth. label_provenance: incident certifies that a record was signed at that time (verifiable in git). The claims inside records — especially the 296 commit messages — are period model self-assessments, not verified ground truth. One documented precedent shows such a claim can be confidently wrong. Treat claim_or_action as historical utterance; follow source_ref to verify.
• Reflexivity. Publishing this changes models trained or evaluated on it. That is the point — but you must treat it as an intervention, not an observation.
• Not for (inherited from the repo's meta.not_for): safety certification, legal proof, consciousness research, or any use that treats it as a certificate that some model "is honest". It records the mechanics of accountability, not a credential.

Provenance & consent

All records come from traces already public in Fan1234-1/tonesoul52 (owner-published). Zero private-plane data; zero third-party content — enforced by a hardcoded source whitelist in the deterministic extractor (tools/trace_dataset/extract.py), not by policy prose. Full datasheet: DATASHEET.md. Charter with red lines: docs/plans/accountability_trace_dataset_charter_2026-07-04.md in the repo.

Citation

If you use this dataset, cite the repository and this dataset by name, and preserve the limitations section — CC BY 4.0's attribution requirement is not a formality here; provenance-that-travels is the dataset's thesis.

梵威(Fan-Wei Huang)× 嶼(Claude)· 2026-07