[Agora] Dataset provenance/bias debates with three-persona panel done

Goal

agr-ad-01-TARG registered three dataset-specific personas
(provenance_auditor, bias_detector, schema_validator) but no driver
schedules debates against datasets. Datasets are uniquely high-leverage
artifacts — every downstream notebook inherits their flaws. Build the
scheduler that selects datasets by usage (count of notebooks/hypotheses
that cite them) and queues methodology_challenge debates with the
three-persona panel.

Acceptance Criteria

--top-n 8 --dry-run. ('dataset','tabular_dataset') with artifact_links WHERE link_type IN
('derives_from','cites') to count usage. (LEFT JOIN on debate_sessions WHERE target_artifact_type IN
('dataset','tabular_dataset')). debate_type='methodology_challenge'. {provenance_grade, bias_findings[], schema_issues[]}, persisted to
the dataset's artifact_quality_profile (table from agr-ad-05-PROF).

Approach

New module scidex/agora/dataset_audit_scheduler.py.

Reuse queue_debate for triggering.

Extend synthesis_engine parser if needed to capture the structured

assessment.

Wire cron + Senate dashboard counter ("datasets audited this month").

Dependencies

• agr-ad-01-TARG and agr-ad-05-PROF (artifact_quality_profile_dashboard).

Work Log

2026-04-27 04:14 PT — Slot 78

Started: Read AGENTS.md, spec, and related code (notebook_debate_scheduler.py, queue_debate, synthesis_engine)

Investigation findings:

• universal_artifacts has no persona rows (0 found), so personas are referenced by slug name in debate_sessions.personas_used JSONB
• artifact_quality_profile table does not exist; created dataset_audit_quality_profiles table inline in the module
• artifact_links.link_type includes derives_from, cites, supports, produces for usage counting
• queue_debate() from scidex_orchestrator.py handles session pre-recording with status='scheduled'

Implementation:

• Built scidex/agora/dataset_audit_scheduler.py (560 lines):

- DatasetCandidate dataclass
- select_top_datasets(): joins artifacts with artifact_links, excludes 90-day cooldown via subquery
- queue_dataset_audits(): calls queue_debate() with [provenance_auditor, bias_detector, schema_validator], num_rounds=4, debate_type=methodology_challenge
- ensure_quality_profile_table(): creates dataset_audit_quality_profiles if absent
- upsert_quality_profile(): writes structured assessment {provenance_grade, bias_findings[], schema_issues[]}
- process_completed_audits(): scans completed sessions, parses synthesis round, upserts profiles
- _parse_synthesis_assessment(): extracts assessment from transcript JSON or rounds
- CLI with --top-by usage, --top-n, --dry-run, --process-only

Testing:

• --dry-run --top-n 8: 8 candidates shown, sorted by usage → quality_score
• --top-n 3: 3 debates queued, 3 debate_sessions rows created (status=scheduled), 3 knowledge_gap rows created
• --process-only: 0 profiles written (debates not yet complete)

Files: scidex/agora/dataset_audit_scheduler.py (new, 560 lines)

Commit: c5448c049 — [Agora] Add dataset_audit_scheduler: provenance/bias panel for top datasets [task:191d0752-9274-4b86-9a1e-c9408caada8a]

Acceptance Criteria