fedjev-bench

# fedjev-bench

Pre-registered evaluation of TypeSafe/Jev pairwise rankings of Federal Open Market Committee (FOMC) chair press-conference openings under the criterion `more hawkish about inflation`. The behavioral comparison is the same-day federal funds target change (`d_same` — hike, hold, or cut; zero on holds by construction). The external *text* comparison is FedLock, an independent published scoring project ([methodology](https://jnathan9.github.io/fedlock/)): Gate 7 is rank agreement with those published press-conference scores, not a TrueSkill replication.

| | |
|--|--|
| **run_id** | `fedjev-2026-09-20` |
| **model** | jev-1.13.0 |
| **criterion** | `more hawkish about inflation` |
| **Pages** | https://maybern-tripp-smith.github.io/fedjev-bench/ |
| **Gates** | 1 PASS · 3 PASS · 4 PASS · 6 PASS (2 / 5 / 7 report) |
| **Main cost** | ≈ $0.031 · ablation +$0.011 · grand ≈ $0.042 |

**Abstract.** Same-day funds-rate changes are an incomplete label for textual hawkishness when the target is unchanged. The evaluation asks whether Jev recovers easy hawk/dove orderings, agrees in rank with `d_same` on scheduled action days, and still separates holds from cuts on the text axis. Gate 7 reports Spearman’s rank correlation with FedLock’s published press-conference scores (raw TrueSkill mean `m`; era-adjusted `ma` as a sensitivity). It reads those scores; it does not re-run FedLock’s tournament. A separate FedLock-faithful TrueSkill replica on the 95 openings is documented in [`results/fedlock_replica/FINDINGS.md`](results/fedlock_replica/FINDINGS.md).

How to read the scoreboard: [`HOW_TO_READ.md`](HOW_TO_READ.md) ([Pages](https://maybern-tripp-smith.github.io/fedjev-bench/how-to-read.html)). Long-form: [`ANALYSIS.md`](ANALYSIS.md). Gate tables: [`REPORT.md`](REPORT.md