Every file here reports the recognition-verdict metric over the 36-model
panel — the single metric the paper uses. The aggregate tables are regenerated
from the record-level recognition run by
code/build_release_tables.py, which
cross-checks its output against the paper's own figure numbers
(paper/figures/computed_numbers.json)
and refuses to write a table that drifts.
Record-level source of truth:
experiments/t6_v2_protocol/outputs/recognition_final.jsonl
(one binary recognized verdict per entity × model, from the open-book
anti-confabulation judge).
| File | Contents |
|---|---|
namerank_per_entity.csv |
Per-entity recognition score (namerank), panel size, refusal rate, embedding similarity. |
namerank_matrix.json |
Per-entity → per-model recognized (0/1) matrix, main run (4,730 × 36). |
per_model_summary.csv |
Per-model recognition and refusal rates; vendor and Western/Chinese class. |
cohort_summary.csv |
Per-cohort recognition distribution (mean/median/sd/percentiles, frac_recognized, refusal). |
credential_ladder.csv |
The credential-treadmill table (IMO/IOI/ICPC/Putnam/CMO/NOI/CPhO/Rhodes/MSRA/DeepSeek/GPT-5 authors). |
cs_faculty_by_country.csv |
CS-faculty recognition by institution country (the country gradient). |
cross_language_per_entity.csv |
EN vs ZH recognition on the 240-entity Chinese-prompt sub-run, with Western/Chinese model splits. |
attribution_pairs_v2.csv |
Creator vs artifact recognition for the verified inversion pairs. |
model_cutoffs.json |
Per-model knowledge-cutoff metadata (metric-independent). |
Cohort-level tables (cohort_summary, credential_ladder, cs_faculty_by_country)
report the paper's cohort-mean population: real (non-synthetic) entities with a
v2 gold answer, judged by the full panel — matching
paper/figures/_data.cohort_table. namerank_per_entity.csv lists every probed
entity without that filter.
To regenerate: python3 code/build_release_tables.py.