For each central-bank speaker, build a vocabulary-feature corpus of their public communication, fit a walk-forward model that maps the features at presser t to the cross-asset reaction at meeting t+1, and report the directional accuracy + position-sized P&L Sharpe against a same-direction-as-last-meeting momentum baseline.

• Source. Official transcripts only. Fed FOMC press conferences (2018–present), ECB monetary-policy press conferences (Draghi 2017–19, Lagarde 2020–present), BoE inflation report / Monetary Policy Report press conferences (Carney 2015–19, Bailey 2020–present), BoJ kaiken summaries (Kuroda + Ueda).
• Tokenization. Standard whitespace + word-character regex. Per-1000-token normalization applied to every term frequency to make different transcripts comparable.
• Lexicon. 165 hand-curated terms grouped into hawkish, dovish, patient, confident, concerned, hedge, and topic-naming categories. Same lexicon applied to all chairs — the architecture transfers if and only if a Fed-built lexicon picks up signal at non-Fed central banks.
• Features. Per event: lexicon counts per 1000 tokens, plus aggregated category sums, plus a net_hawk_dove spread. The model sees only prior_* (event t-1) and d_* (event t-1 − event t-2) for predicting event t. No future leakage by construction.

• Walk-forward only. For each test row at index i, the model is re-fit on rows [0, i) and produces a single prediction for row i. Standard time-series cross-validation. No re-fitting on test data, no peek-ahead features.
• Two estimators per target. Linear (Ridge, alpha=1.0) on 22 verbal features for SPX / FX / equity targets. Gradient-boosted (HistGradientBoostingRegressor) on 76 features including macro regime variables (CPI, NFP, unemployment, prior-meeting actions) for US treasury yield targets. Separate CV-tuned per target.
• Metrics reported. Direction accuracy vs momentum + majority baselines; predicted-vs-actual Pearson correlation; cumulative position-sized P&L (clip-and-cap at 75th-percentile absolute prediction); annualized Sharpe; bootstrap 95% CI on Sharpe (5,000 samples).
• Decision posterior. For Powell only (so far): multinomial logistic regression on prior-presser features, blended with a 12-meeting trailing base-rate prior (alpha=0.15). Calibration scored via Brier and log loss against the same base-rate baseline.

• Powell. n=43 walk-forward. Decision argmax accuracy 74.4%. Brier 0.456 vs 0.623 base-rate (−27%). Cross-asset edges on rates (Sharpe 1.32–1.62, CIs tight), SPX (Sharpe 0.66, CI [−0.18, 1.42]), credit, and oil. Sector rotation (XLE/XLK) is null — an honest disclosed limit.
• Lagarde. n=46 walk-forward. EuroStoxx +14.2pp / Sharpe 0.74. DAX +12.2pp / Sharpe 0.56. The architecture transfers cleanly to a non-Fed central bank without retraining the lexicon.
• Bailey + Carney. Early evidence (Bailey n=17, Carney n=10). Bailey GBP/USD +39.3pp / Sharpe 0.92, FTSE 250 +27.2pp / Sharpe 0.31. Carney GBP/USD +36.7pp / Sharpe 1.06 / corr +0.56. CIs are wide given the small samples. We treat these as suggestive until n=30.
• Draghi. Walk-forward n=41 (2014–2019, including QE-launch era). DAX edge +16.0pp / Sharpe 0.20 / corr +0.02 / direction accuracy 61.0% vs momentum 45.0%. The signal lives in the 2014–2016 QE-launch period; the original 2017–2019 sub-sample (n=18) showed null. The point estimate is positive at scale and the direction edge is large, but the Sharpe CI still crosses zero — we treat as early-evidence, not validated. Two ECB chairs now both show positive DAX edges, which strengthens the role-level architecture claim.
• BoJ. Honestly weak per-chair. Pooled (Kuroda+Ueda, n=130) shows a modest +5.8pp edge on Nikkei but the per-chair signals are negative-to-noisy. The FED-built lexicon does not transfer cleanly to BoJ. Either Japan-specific recalibration or an honest "vocmarkets does not work on BoJ" disclosure.
• Pooling fails. 520 Fed-governor speeches pooled into a single signal: negative edges across all targets. Same with Carney + Bailey pooled. Speakers are individuals; the right architecture is per-speaker, not role-level. This is a structural finding, not a hyperparameter problem.

• Cut-tail under-confidence. Powell decision posterior missed the September / October 2025 cut cycle (predicted hold both times, outcome cut). Per-class avg P(correct) on cuts is 0.36 vs 0.68 on holds and 0.77 on hikes. The model needs more cut-cycle history to calibrate the cut tail; we expect this gap to close as 2025–26 cuts accumulate.
• Decision posterior — extended to ECB. The first attempt on Lagarde-only at n=26 walk-forward showed Brier improvement only +1.6% vs base rate. After backfilling Draghi 2014–2019 (+27 events), the combined Draghi+Lagarde decision posterior is trained on the full ECB rate-decision history. Specific argmax-accuracy and Brier numbers from this combined run are pending pipeline verification and will be published here once they trace cleanly to a saved ecb_posterior_combined.json artifact (single-source-of-truth discipline applies). BoE Bailey decision posterior remains next-session work (Bailey n=22 is below the n≥30 validation threshold we apply).
• Sample size on non-Fed chairs. ECB n=46 (Lagarde) / 41 (Draghi). BoE n=17 (Bailey) / 10 (Carney). BoJ n=18 (Kuroda) / 21 (Ueda). Most non-Fed chair Sharpe CIs still cross zero. Time fixes this; we are not claiming validation at n<30 with CI crossing zero.
• Live track record is single-digit days old. Three live pre-registered calls (Bailey May 8, Lagarde June 5, Powell June 17). The walk-forward backtest 2-hits / 2-misses on the four ledger entries is honest and they are not pre-registered live calls; that distinction is preserved on the public ledger. The broader walk-forward record across all n=43 FOMC meetings is 74.4% argmax-correct.
• No real-money validation. All P&L is paper, position-sized 1x notional with no slippage / costs / financing. Realistic desk costs would compress most of the borderline Sharpes. We seek a desk partner to run a live shadow book in 2026 H2.

• For rates desks. A magnitude forecast: predicted bp move on the meeting day, MAE 18% better than naive zero baseline on TNX and FVX. Sized directly off the model output.
• For equity / FX / credit desks. A direction signal + recommended position-size cap. Magnitudes are correctly-signed but ~2x too aggressive on the un-shrunk model output; apply 0.5x calibration factor or use the sign only.
• For event traders. A pre-registered probability call deployed to /fomc before each meeting, resolving publicly with the meeting outcome. Calibration and Brier improvements are tracked across all resolutions.

• May–July 2026. Resolve Bailey May 8, Lagarde June 5, Powell June 17. Each becomes a public data point on the live track record.
• Q3 2026. Williams + regional Fed presidents (each its own per-speaker corpus). ECB and BoE decision posteriors (cut/hold/hike probabilities, not just cross-asset).
• H2 2026. Live desk shadow run in partnership with a friendly proprietary desk. Anthropic-API-backed live thesis extraction (currently the page's thesis field is hand-written from model features).