FedLock (Gate 7)

Gate 7 is an external consistency check: do two independently constructed text scores agree on meeting-day hawkishness? It is not a FedLock replication, and it is not the TrueSkill replica documented in results/fedlock_replica/FINDINGS.md.

This note is self-contained. A reader who has not read the main analysis should still be able to rebuild the Gate 7 match and the rank correlations from the frozen artifacts. Numbers are taken from results/gates.json (7_fedlock_consistency). None were invented for this write-up.

1. Terms, taught once

Federal Open Market Committee (FOMC). The Federal Reserve committee that sets the federal funds target. The documents on this side of Gate 7 are Chair press-conference openings (jsort-style), not necessarily the full press-conference transcript.

Behavioral label versus textual hawkishness. The main bench’s behavioral label is d_same, the same-day change in the federal funds target (hike, hold, or cut). On a hold, d_same is identically zero, so it cannot encode hawkish-hold versus dovish-hold wording. Gate 7 does not correlate against d_same. It correlates two text scores with each other.

FedLock. An independent published scoring project (methodology; snapshot data/raw/fedlock/data.json). FedLock V3 asks Llama 3.3 70B which of two anonymized speeches is more hawkish given contemporaneous macro conditions, then aggregates ~60,000 comparisons on ~4,000 speeches with TrueSkill (defined below). Published press_conference fields used here:

The primary Gate 7 contrast is raw m. Era-adjusted ma is a sensitivity, not the headline.

TrueSkill versus Bradley–Terry. Gate 7’s left-hand scores come from this repository’s main protocol: dual-order Choice under more hawkish about inflation, aggregated by Bradley–Terry (a pairwise strength model on a fixed gold-pair graph), plus a secondary direct Score pass (score_jev). TrueSkill is Microsoft’s Bayesian skill-rating system (prior mean 50, prior uncertainty about 8.33, target uncertainty σ < 2). Published FedLock uses TrueSkill. Gate 7 does not re-run TrueSkill. It only reads FedLock’s published means and asks how they rank against Bradley–Terry / Score.

Macro conditioning and anonymization (what Gate 7 does and does not share).

• Shared with FedLock, in a limited sense: the object of measurement is pairwise textual hawkishness; this repository strips names and dates on its own Jev Choice calls (scripts/strip_meta.py); FedLock scores are used as an external reference.
• Not used on the Gate 7 left-hand side: macro-conditioned judge prompts (core personal consumption expenditures inflation, unemployment, real gross domestic product growth, and the CBOE Volatility Index at speech time); TrueSkill aggregation; Swiss or uncertainty-targeted pairing; Llama 3.3 70B; FedLock’s full speech corpus.

Gate 7 versus the FedLock-faithful replica.

Jev versus Haiku in the main bench (not this table’s third column). The Gate 7 numbers below are Jev (Bradley–Terry and Score) versus published FedLock. A same-protocol Haiku Choice arm exists for the main gates (cost / inversion); it is not the Gate 7 left-hand series. The replica’s Haiku arm is a different run (claude-haiku-4-5-20251001 under the TrueSkill protocol).

Spearman’s rank correlation (ρ). Do the two series produce a similar ordering of meetings? Reported with a bootstrap standard error (s.e.) — the standard deviation of 1,000 meeting resamples — and a percentile 95 percent confidence interval. This note writes “standard error” or “s.e.” It does not use the label STE.

2. Procedure a reader can re-run (from cached outputs)

1. Load meeting-level scores from results/statement_scores.csv (Bradley–Terry score, score_jev) and the meeting calendar / exclusions from data/labels/meetings.parquet.
2. Load FedLock press_conference rows from data/raw/fedlock/data.json.
3. Match on title-embedded meeting date when a 20YY-MM-DD string appears in the FedLock title; otherwise use the FedLock d field with calendar offsets 0, +1, −1, +2 days.
4. Restrict to the main-analysis sample (scheduled meetings; drop the pre-registered crisis dates 2020-03-03 and 2020-03-15; flag 2023-03-22 without dropping).
5. Compute Spearman’s rank correlation of Bradley–Terry versus m and versus ma (n = 46 meetings with a Bradley–Terry score) and of score_jev versus m and versus ma (n = 90). Bootstrap 1,000 meeting resamples for the standard error and the percentile interval.
6. Do not treat the result as a TrueSkill replication. If the question is protocol fidelity, stop and open the replica note instead.

3. Matching (observed)

• Policy: prefer title-embedded meeting date; else FedLock d with offsets 0, +1, −1, +2
• Matched meetings (all calendar): 92
• Main-analysis matched rows: 90
• Same-calendar-day on the FedLock d field: 1
• Offset (fedlock_d − meeting) on the main-analysis match: {'0': 1, '1': 89}
• Match route on the main-analysis match: {'title_date': 90}
• Mean FedLock s on the matched main set: 1.7762 (TrueSkill uncertainty; lower means more settled)

When the match falls back to the d field, the offset is almost always +1 day. That is a dating convention, not a second sample.

4. Results

Agreement with raw m is stronger than with era-adjusted ma, especially for the Score pass. The Score-versus-m contrast is near a rank ceiling (ρ = +0.944, s.e. = 0.016, n = 90). Bradley–Terry versus m is lower (ρ = +0.679, s.e. = 0.112, n = 46) on a smaller, sparser graph.

5. What this does and does not imply

Gate 7 ρ asks: do two independent text-scoring systems agree on meeting-day hawkishness?

It does not claim: we replicated FedLock’s tournament, judge, corpus, or aggregator.

Corpus mismatch. Our documents are chair openings. FedLock press_conference may be fuller presser text. Agreement is still informative; it is not a same-document comparison.

Length design choice (jsort --max-chars 8000). jsort / jgrep default --max-chars 8000 truncates what Jev sees in jsort tournaments. On our opening corpus (data/clean/statements.jsonl, n=95): mean ≈ 7,917 characters; p50 ≈ 7,547; p95 ≈ 11,706; max = 13,794; 39 / 95 (41.1%) exceed 8,000 characters. The frozen main run_id=fedjev-2026-09-20 Score/Choice pass did not apply that 8k client cap. Audit: results/khaled_sensitivity/char_length_audit.json. Passage-filter Score pilot: ANALYSIS §13 · results/khaled_sensitivity/filter_pilot.json.

Separate experiment. A FedLock-V3-faithful protocol (macro-conditioned relative hawkishness, anonymization, TrueSkill with Swiss / uncertainty pairing) was run on the 95 openings, comparing Jev, Haiku, and published m / ma. That run has its own run_id (fedjev-fedlock-replica-2026-09-20). See results/fedlock_replica/FINDINGS.md and ANALYSIS §12. It is not a 4,000-speech / ~60,000-comparison scale replication, and it does not replace Gate 7.