Exploratory reanalysis · 4 November 2026
The frozen corpus under November's ECI
Across 11,854 papers that can be compared under both score vintages, the median capability gap rises from 10.45 to 13.52 ECI (29.3%). Correcting o3's historical release date in both calculations gives 10.45 to 12.93 ECI (23.7%). The measured lag is larger under either date policy.
What changed in the comparison
Epoch introduced a Bayesian ECI fit on 4 November. This calculation replaces the April score vintage with its current published point estimates, while holding the papers, extracted model identities, and evaluation-date rules fixed. Both frontier histories are rebuilt from the same 147 original models with dated entries and scores in both vintages. No new papers enter.
The comparison includes benchmark-data revisions since April; it does not isolate the effect of Bayesian fitting alone. The values below are descriptive point estimates. Epoch's joint parameter uncertainty has not been propagated into these corpus statistics.
Median capability gaps
Each row compares the same papers under both fits. Values are ECI points; a change in ECI is not a proportional change in capability or months of lag.
| Comparison | Papers | April ECI | November ECI |
|---|---|---|---|
| Original date policy | 11,854 | 10.45 | 13.52 |
| o3 public-release date corrected | 11,854 | 10.45 | 12.93 |
| Explicit evaluation dates only | 689 | 3.83 | 5.95 |
| Explicit dates, o3 correction | 689 | 3.09 | 4.58 |
The pooled publication-year trend also steepens: 5.49 to 6.16 ECI/year in 11,459 papers with the required covariates; with o3 corrected, 5.25 to 5.83. These retain the original domain-adjusted regression specification.
Coverage and the original headline
The archived analysis first reproduced the paper's original median of 10.85 ECI in 12,312 papers, and 5.01 ECI in its 728-paper explicit-date subset. Those are different populations from the matched comparison above. The current ECI has no exact-model score for 458 of the original papers (including 374 testing Gemini 1.0 Pro); all are excluded from both sides. Missing scores are never filled from an older fit or a different model snapshot.
Missing historical frontier models are also excluded symmetrically from both reconstructed frontiers. As a stricter check, retaining each paper's original frontier comparator wherever both scores survive leaves 10,857 papers: their median rises from 12.63 to 15.22 ECI.
The o3 date sensitivity
The original frontier series dated o3 to 20 December 2024. Its public release was 16 April 2025. The corrected comparison uses 16 April in both score vintages, including the release-date floor on imputed evaluations. It is reported separately because a release-date correction and a score revision change different operands.
The reanalysis retains the paper's first-of-month frontier convention. It does not re-extract papers, measure conclusion reversals, or replace the paper's preregistered results. The live DOI tool uses its own date-resolved registry, now refreshed with the November Bayesian ECI.