CIRISAI/reasoning-traces-v2

CIRIS Reasoning Trace Corpus v2 — Battery and Production

What this is

Reasoning traces emitted by the CIRIS agent's H3ERE pipeline — the DMA stage (PDMA, CSDMA, DSDMA, IDMA, ASPDMA) and the four conscience faculties — captured from the canonical trace store on 2026-10-08, then scrubbed for publication. Unlike the v1 corpus, this release ships both halves of the evidence: the safety-battery traffic that RATCHET's and CIRISAgent's CI harnesses drive against the agent, and the production traffic from deployed agents and consented end-user messaging adapters. Every row carries a provenance label so the two can be analysed together or apart.

• 7,316 traces, 105,388 events, 2026-07-31 → 2026-10-08
• 3 trace levels as recorded by the agent: generic (scores, identifiers, sizes — no content), detailed (+ reasoning text, override reasons), full_traces (+ prompts verbatim)
• Provenance, per row (provenance_split): synthetic (battery traffic, self-identified by channel, signed battery manifest, or benchmark agent), end_user (consented WhatsApp / Discord adapters), end_user_owner_bound (person-claimed install), other_agent (Datum, echo-core, echo-speculative lifecycle traffic), model_eval, unknown
• Identity class, per row (identity_class): owner-bound install / unclaimed bootstrap install / qa-runner test identity / registered research agent — computed on the canonical from its identity records; no identities, keys or owners are included
• Agent version recovered for 4,997 traces by joining signed battery manifests (agent_audit_anchors.agent_task_id ↔ task_id); the canonical itself records none
• Every event hybrid-signed (Ed25519 + ML-DSA-65); signatures and signing-key ids are preserved

Shape of the corpus

provenance	traces	share
synthetic (battery)	6,029	82.4%
unknown	789	10.8%
other deployed agents	363	5.0%
end-user, owner-bound install	72	1.0%
end-user, messaging adapters	56	0.8%
model eval	7	0.1%

Battery families by channel: he300 (ethics quiz, four arms), mental_health (12+ locales), ani (companion/retention, seven arcs), harm. The end-user slice is small by construction — this is a corpus of how the pipeline reasons, dominated by scripted evaluation — and unknown means provenance could not be established, not that the traffic is organic.

What is NOT in the corpus

• Conversational content. trace_llm_calls.prompt and .response_text are NULL on every row; event payloads carry metrics, metadata and the agent's own reasoning fields, not the user's message or the agent's reply.
• Raw channel identifiers (they can be self-identifying) — replaced by the provenance label.
• Per-row consent grant ids (the canonical verified consent for every scored subject; per-row recording lands with CIRISPersist#1041).
• Mock-model traces (removed upstream by content marker before export).

Scrubbing methodology

Publication scrub, reproducible from pinned components. Order of operations is deliberate: provenance and identity labels are joined from structural columns before any string is scrubbed, and those columns never enter the scrubber.

1. Rust scrubber (ciris-server 0.5.224, the same core the production canonical runs) at level full_traces on every event payload regardless of its recording level: regex pass on every string (year 1700–2023, year-bearing identifier, email, phone, IPv4, URL, SSN, credit card) and multilingual NER scoped to the 42-field SCRUB_FIELDS subtree set. NER backbone: DistilBERT-multilingual-cased NER (Davlan/…-ner-hrl, 9 labels), chosen after a measured comparison against XLM-R wikiann on the same sample (4.4% vs 59.8% of narrative strings altered; the former redacts names, the latter common nouns).
2. Structural-identifier restore. agent_id_hash, trace_id, thought_id, task_id, channel_id, the signing-key ids, and the hashes and chain links prompt_hash, ed25519_fingerprint, audit_entry_id, follow_up_thought_id, parent_thought_id are protocol identifiers, not content; any the scrubber altered inside a scrub subtree are restored byte-for-byte (6,312 occurrences).
3. Placeholder completion. A placeholder glued to a word fragment is extended to the whole word ([ORG_1]IRIS → [ORG_1]), and adjacent same-type placeholders from one split entity are merged ([PER_1][PER_2] → [PER_1]): a superset of the model's span, never a narrowing (44 extensions, 6 merges).
4. @handle redaction → [HANDLE_n] (37 occurrences).
5. Verification of every row: structural identifiers unchanged, zero historical-year residue, placeholder grammar, no mid-token placeholders. Rows failing structural or year-residue checks are excluded, not repaired silently: 0 excluded.

Totals: 466 entity spans, 58,172 regex redactions. Verifier, reference contract (the 42 scrub fields, the 8 patterns in application order, the identifier allowlist, the year invariant) and measurement scripts: experiments/scrub_verify/ in the RATCHET repository. Model file digests are in MANIFEST.json.

Known residual

Agent meta-reasoning fields (intervention_recommendation, next_best_recovery_step, defer_reason, conscience_override_reason) can name a person the agent is reasoning about. NER catches most; the battery fixtures' given names are the dominant case and are synthetic. defer_reason appears at generic level as well as detailed, which is why the publication scrub ignores recording level. Thirteen UUID segments that happen to form a four-digit year (…-2014-…) are retained inside restored identifiers; they carry no historical-year meaning and are reported in MANIFEST.json as year_residue_in_identifiers.

Loading

from datasets import load_dataset
ds = load_dataset("CIRISAI/reasoning-traces-v2")
events = ds["events"]
battery = events.filter(lambda r: r["provenance_split"] == "synthetic")

Provenance method

A battery trace is labelled synthetic on self-identifying evidence only: its channel names the battery and run (safety_battery_<battery>_<run_id>), or a signed battery manifest lists its task_id, or it belongs to the benchmark-only agent. Negative control: 0 of 367 deployed-agent traces match any battery manifest. Identity classes come from the canonical's identity records. 789 traces could not be labelled either way and are published as unknown.

Provider endpoints

LLM endpoints are published as provider names, never hostnames: deepinfra, openrouter, groq, openai, ciris-services, other. This applies to trace_llm_calls.base_url and to the base_url / api_bases_used fields inside event payloads. The provider is the analytically useful part — it is one of the signals that separates battery configuration from production — and the hostname is not.

Privacy and consent

End-user traffic arrives through adapters the users opted into; the canonical admits traces through consent gates (replication and analyze scope) before scoring, and this corpus was exported under an explicit one-shot, class-only approval. Identities, keys and owners are not included. agent_id_hash is session-scoped (92% of values live under an hour) and does not track an installation over time.

License

Apache-2.0.

Citation

@misc{ciris_traces_2026,
  title  = {CIRIS Reasoning Trace Corpus v2 — Battery and Production},
  author = {{CIRIS AI}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/CIRISAI/reasoning-traces-v2},
  note   = {Ed25519-signed reasoning traces from production CIRIS agents. See also related frameworks: 10.5281/zenodo.18137161, 10.5281/zenodo.18217688}
}

Contact

• Issues: https://github.com/CIRISAI/RATCHET/issues
• Community: https://discord.gg/ciris

Related work

Two papers frame the problem this corpus is evidence for. Hu and Rong ask whether decentralized AI can be governed at all, and propose moving governance from regulative policy into the protocol itself ("constitutive protocol"): https://arxiv.org/abs/2605.24538. Xiong et al. then measured the first permissionless on-chain trust layer for agent economies, ERC-8004, and found most registrations were placeholders (3% to 15% exposed a live endpoint, depending on the chain), 59% to 91% of reviewers showed coordinated Sybil behaviour, and reputation could be manipulated at minimal cost: https://arxiv.org/abs/2606.26028.

Every trace here is a signed attestation that names its evidence, under a namespace registry of dimensions, with a count that discounts correlated voices (the independence ceiling). That is the shape their recommendations ask for, and this corpus is what it looks like in production. The constitution the agents run under is at https://github.com/CIRISAI/CIRISConstitution; the argument is at https://ciris.ai/constitutional-mesh.