Product

A constitution enforces what a tired human skips.

A binding 52-section protocol governs every experiment. Three programmatic checks run before any commit — not linting, not advice: an experiment that fails a gate never enters the champion line. Refusal is the default state.

The loop

Seven steps, enforced on every experiment

Diagnose→Literature-ground→ Hypothesize→Execute→ Analyze→Document→Checkpoint
GATE 1

Data integrity

Purged and embargoed super-folds. validate_no_overlap() runs programmatically and the runner refuses to launch on any overlap.D8

GATE 2

Citation rigor

Every reference verified against live indexes before compute is spent. Ungated fabrication rates run 14–95%; a refused experiment costs a message, not a GPU-hour.D35

GATE 3

Reasoning completeness

The agent may not, by contract, write a result row. The runner writes what actually happened.D33

One bypass, and it is visible. The single override writes a sentinel tag into the run log, so an auditor grepping for it finds every experiment that went around a gate. The fingerprint makes the metric tamper-evident, not tamper-refusingROADMAP and "monotonic" always means monotonic in the composite proxy.
Three translucent panes with a beam passing through, emerging clarified.
What passes is not what survived review — it is what was never allowed to enter.
What you see

The trial ledger is the demo

Not a dashboard of green numbers. A laboratory notebook that cannot be falsified — read like a trading terminal, steered like a chat.

Trial ledger, discards included

An append-only log records every experiment — kept and discarded — with seed, config, git commit and gate verdicts.D6

Trial 218 of this campaign — 9 kept, 209 discarded. Nothing is ever quietly deleted.

Deflated-statistic panel

The deflated Sharpe ratio, computed from the true ledger count, beside every raw champion figure.D7

Sharpe +6.52 raw · deflated X over 265 trials — full ledger attached.

Super-fold split builder

Purge, embargo and label-horizon buffers drawn on your actual dates.D8

0 overlapping rows across 7 folds — verified.

Goodhart divergence monitor

Tracks the composite proxy against each raw target metric and alerts when they diverge.D31

Composite up 4% over last 30 trials; raw test AUC flat — review metric weights or steer.

Trial-count-corrected acceptance gateROADMAP

The keep/discard threshold rises with ledger count. No published system has this gate — the proof of concept included.

Improvement +0.011 < deflated threshold +0.019 at N=218 trials — discarded; logged to ledger.