pub-10 data
Data license: Not yet licensed — internal review only · Data source: Ahmed ElBatanony, Innopolis University
Queries
- 25 random final statements
- Everything that happened to one application's statements
- Corrected statements: original vs final wording
- Rejected for a reason other than duplication
- Run A vs run B, per application
- Blind audit: rows the auditor would not accept
- Blind audit: share passing each check, by stratum
- Token usage by agent role
- Search statements and excerpts
Tables
apps
app_id, frame, frame_key, name, vendor_url, pcf_id, pcf_category, archive_status, sources, source_words, features, both_runs, run_a_only, run_b_only, added, archive_note
25 rows
audit
feature_id, supported_by_excerpt, supported_in_context, is_capability, single_capability, no_marketing, actor_ok, trivial, correct, note, stratum
150 rows
candidates
candidate_id, app_id, run, statement, excerpt, source_url, source_file, script_check, verifier_verdict, verifier_reason, verifier_correction, adjudicator, adjudicator_reason, in_final_set
3,819 rows
features
feature_id, app_id, statement, form, excerpt, source_url, source_file, source_sha256, found_by, found_by_runs, verification, pipeline_id, audited, audit_correct, audit_note
2,021 rows
pilot_features
feature_id, app_id, statement, excerpt, source_url, extracted_by, verification
322 rows
provenance
agent, role, run, apps, models, effort, max_output_tokens, claude_code, started, finished, input_tokens, output_tokens, cache_read_tokens, cache_write_tokens
79 rows
screening
frame, candidate, screener_verdict, screener_code, screener_reason, adjudicated, adjudication_reason, in_sample, pcf_id, set
576 rows
... and 5 hidden tables
Download SQLite DB: pub10.db 3.5 MB