teranode-ai/council-decisions-benchmark

Teranode Council Decisions Benchmark

Anonymous structural telemetry + head-to-head eval comparisons from the Teranode Council — a multi-agent reasoning system for regulated financial advisory.

This dataset is the public-facing complement to the live system at teranode.ai. It captures every model-pin head-to-head comparison Teranode has run during model selection, plus anonymized structural telemetry from every Council deliberation. The dataset is regenerated nightly from the production database and intended as a citable benchmark for AI-in-financial-advisory research.

What this dataset is for

• Reproducibility — researchers benchmarking AI in financial advisory can cite specific eval comparisons by ID and verify the model-pin claims Teranode makes publicly.
• Transparency — CCOs and regulatory observers can see what the system has actually done, not just what marketing material claims.
• Cross-vendor model evaluation — every comparison runs the same scenario through two candidate models for one role, holding all other roles constant. Useful for vendor-neutral assessment of frontier-model performance on regulated-finance reasoning tasks.

What's in it

Two configs, both released as JSONL.

head_to_head_evals — model-pin A/B comparisons

One row per head-to-head comparison run via /admin/eval. Each row varies one role's model (A vs. B), holds the other four roles + chairman prompt constant. Founder grades each comparison: a_wins, b_wins, tie, or skip.

Columns:

• id — eval row primary key
• created_at — ISO timestamp
• judged_at — ISO timestamp the founder graded the comparison; null if pending
• varied_role — one of critic, builder, strategist, contrarian, chairman
• model_a, model_b — OpenRouter model identifiers (e.g., openai/gpt-5.4, anthropic/claude-opus-4.7)
• chairman_prompt_a, chairman_prompt_b — chairman prompt versions when prompt-A/B (null on model-A/B runs)
• judgment — a_wins | b_wins | tie | skip | null (null = pending)
• policy_version — model-policy.json version active when the run executed

Excluded by design (request NDA detail at founders@teranode.ai):

• Scenario text
• Model outputs (output_a, output_b)
• Judgment notes (founder-typed rationale)

council_telemetry — anonymous structural telemetry

One row per Council deliberation, regardless of opt-in. Captures structural metadata about each computation; never captures content.

Columns:

• created_at — ISO timestamp
• source — api_analyze, api_stream, eval_admin, eval_tier_comparison, eval_examiner, eval_sufficiency, other
• tier — quick | standard | battle_tested | null (null = no tier specified)
• mode — fast | deep
• domain — retirement | portfolio | tax | estate | risk | general | null (null = no structured input supplied)
• duration_ms — total wall-clock for the run
• role_failure_count — count of role calls that failed (0-4 typically)
• cross_exam_eligible — boolean, whether chairman returned a strongest_objection_source
• cross_exam_fired — boolean, whether the cross-examination round actually ran
• cross_exam_decision_revised — boolean, whether chairman changed the decision after cross-exam (null when cross-exam didn't fire)
• examiner_ran — boolean
• examiner_verdict — pass | gap | null
• chairman_prompt_version — v1 | v2 | v3 | v4
• model_policy_version — version stamp from model-policy.json

Explicitly NOT captured:

• Scenario text, transcript, or any user input
• Output text of any kind
• Confidence values, dissent text, recommendations, risks
• IP address, name, email, cookies, cross-site identifiers
• Anything that could be linked back to a specific user

How the data is generated

The system runs at teranode.ai. Every Council deliberation goes through one of these entry points:

• Homepage demo (/api/meeting/analyze or /api/meeting/stream)
• Reliability-graph scenario stars (/api/meeting/stream)
• Admin eval scaffold (/api/admin/eval/run)
• Script-based evals (tier-comparison, examiner-validation, sufficiency-loop)

Each entry point writes one telemetry row (council_telemetry). Admin eval comparisons additionally write to eval_runs with the founder's manual judgment.

The dataset is regenerated nightly from the production Postgres database via the dataset/build.ts script (in the source repo). The static JSONL files in this repository are the most recent regenerated snapshot.

Architectural context

The Teranode Council is Mixture-of-Agents (MoA) with a per-role-pinned chairman synthesis layer plus two optional add-on rounds:

• Phase 0 (always on, default behavior): v3+ chairman includes role-attributed dissent (strongest_objection_source).
• Phase 1 (battle-tested tier or env flag on): targeted cross-examination round. Chairman v4 judges per-run whether the dissent is material; when flagged, the dissenting role responds and the chairman re-synthesizes.
• Phase 3 (battle-tested tier or env flag on): FINRA-/SEC-examiner-mode adversarial review of the chairman synthesis for Marketing Rule 206(4)-1 / IAA §206 / FINRA Rule 2111 / Reg BI / SEC 2026 Exam Priorities failure modes.

See teranode.ai/methodology for the full architecture writeup, or methodology.md in this dataset for an offline-readable copy.

Citation

If you use this dataset in research:

@dataset{teranode_council_decisions_benchmark,
  title  = {Teranode Council Decisions Benchmark},
  author = {Zimon, Daniel and contributors},
  year   = {2026},
  url    = {https://huggingface.co/datasets/teranode-ai/council-decisions-benchmark},
  note   = {Multi-agent reasoning telemetry + head-to-head eval comparisons for regulated financial advisory.}
}

License

CC BY-SA 4.0. Free to use, redistribute, and build upon for any purpose, with attribution + share-alike.

Contact

founders@teranode.ai — questions, NDA-gated detail (scenario text + outputs + judgment notes), or to flag a row that needs correction.

Limitations & honest framing

• Founder-graded judgments. The judgment column on the head_to_head_evals config currently reflects single-rater (founder) judgment. Independent judges and inter-rater agreement are on the roadmap as the company scales.
• Sample size. Most subsets have N < 100 at the time of first publication. The dataset compounds with every production run; subsequent versions will have more depth.
• Domain coverage. Currently weighted toward retirement and portfolio scenarios. Estate, tax, risk, and general scenarios are present but lower-volume.
• Cross-vendor representativeness. Models pinned in the production policy are evaluated against candidates also in the OpenRouter catalog. Models outside OpenRouter (private deployments, on-device models) are not represented.