trdrbot - the Coach

Autonomous subsystem improvement. Generated 2026-08-31T21:42:25+00:00. 99 snapshot(s), 1 experiment(s), 0 promotion(s).

What changed

Nothing yet. The Coach opens its first experiment once the muse has run and the mutation cooldown has elapsed.

Open experiments

muse.prompt: v1 challenging v0
P(challenger better) 0.379 after 9 paired run(s) - challenger 39/45, incumbent 40/45, 3 voided
needs 0 more run(s), 0 more candidate(s), and P ≥ 0.9 to promote

Gauges

North star: calibration.brier with its sample count - the one number improvement is for, because a lucky week cannot move it and a well-calibrated view can. Guardrails (must not degrade, both directions): sizing.refused_rate and exit.uncorroborated_decisives (rule compliance - a seam losing the conditional payoff, or a stop firing on quote noise), the trade/decline balance behind attribution.attributable_rate (always-trading and never-trading are both drift, and this book has measured both), and coach.cost_usd_today with model.cal_age_days (cost and staleness). Everything else here is diagnostic.

book.beta_delta_dollars

-376,578.26

min -379,003.07 / max -330,183.51 / n=3

book.vega_dollars

174.13

min 171.26 / max 211.53 / n=3

calibration.brier

0.1712

min 0.1712 / max 0.1712 / n=29

calibration.n

min 27 / max 87 / n=99

calibration.n_eff

11.2

min 11.2 / max 11.2 / n=29

calibration.resolved

min 3 / max 36 / n=99

coach.cost_usd_today

0.5119

min 0 / max 2.033 / n=99

coach.open_experiments

min 0 / max 1 / n=99

coach.promotions_total

min 0 / max 0 / n=99

discovery.gauntlet_survival

0.4286

min 0.4286 / max 0.4286 / n=86

exit.uncorroborated_decisives

min 0 / max 0 / n=60

model.cal_age_days

min 0 / max 2 / n=91

model.inflation_5d

1.3

min 1.3 / max 1.3 / n=91

muse.candidates_per_run

min 3 / max 5 / n=99

muse.funnel_overlap_rate

0.36

min 0.34 / max 0.38 / n=54

muse.runs_total

min 10 / max 22 / n=99

muse.seed_entropy

min 12 / max 29 / n=99

muse.survival_rate

0.92

min 0.4333 / max 0.92 / n=99

research.opportunities_per_run

1.143

min 1.143 / max 2 / n=86

Levers and how to steer

lever	incumbent	since	state
`muse.prompt`	`v0` 7809d229	seed	running

Variants live in data/state/levers/*.json. Set "paused": true to stop experimenting on a lever - it closes any open experiment and freezes the incumbent, which is how you hold behaviour still for a demo. Editing the incumbent text by hand is supported; the fingerprint is recomputed from the text on load. Nothing here can reach code, gate thresholds, sizing, or the constitution.