DRAFT-003 — Within-regime continuation on the confirmed class

A Draft, in the sense CONTEXT.md gives the word: a proposed Hypothesis carrying a mechanism and a falsifier, with no standing until a human confirms its registration, which is the act that spends alpha. Nothing here has been measured. Nothing here has cost anything.

The author agreed on 2026-09-11 that this is the first question to register. It is the RegimeProposer's draft, written down so the registration is a reading and not a reconstruction.

Prepared 2026-09-11 against the real archive (python -m occams question prepare): the classifier frozen at 0434836a176a; 350 signals on the definition partition at 0.0989 per name-day, ρ 0.000; available_n 837 on the measurement partition against required_n 748 per cell at the configured rate — powered, by a thin margin. Nothing registered.

Re-prepared 2026-09-11 on three names (SPY, QQQ, DIA) as Q-004, after Q-003 resolved NULL on two: 458 definition signals at 0.0949 per name-day; available_n 1206, effective 1074 at measured ρ 0.434; required_n 748 per cell — powered and supportable. Refused by the overlap gate: 67% of the observations were consumed by Q-003, above the 50% threshold; the author must supersede or distinguish (R4.9). Target-axis sensitivity 0.4% on the definition partition. Nothing registered, nothing spent.

Registered as Q-004 on 2026-09-12 by the author with --supersedes Q-003 and the target axis dropped: k = 2, 0.02 alpha, available_n 1074 effective against 748 required. Queued for the loop.

Resolved NULL on 2026-09-12: 2,088 trades on the measurement partition, EV −0.0451 net R at bounded costs, p(null ≥ winner) 0.983. The same answer as Q-003, on a set where leave-one-out could run. The plateau refusal was structural (a two-cell sweep against a four-cell plateau) and is now refused at registration. The lab falsifier stands at two of three.

Read this first — what must exist before it can be registered

1. Real bars in the archive. Registration computes available_n from the measurement partition (M8.2), and the archive holds nothing until the M0.7 class is ingested — python -m occams ingest. No bars, no registration.
2. The classifier, frozen first (ADR-0005, M7.6). calibrate runs on the definition partition of the same archive and the frozen hash goes into the mechanism statement below. Calibration is a priori and happens once.
3. A human registration through register_question, with the accountant, at the regime axis's configured mechanism rate.

Mechanism

In the up regime as labelled by the frozen classifier (index-level), breakout entries carry positive EV in net R because participants under-react to continuation within an established regime: a breakout that occurs while the medium-term average is above the long-term average is more often followed through than faded. The claim is about the instrument's own dynamics on the confirmed class — currently-listed, liquid US-listed ETFs and large caps on daily bars — not about selection from a universe.

Both interpretations

• If true: the winner cell clears the declared floor, beats random entry under the same stop and exits at the corrected alpha, holds a plateau across neighbouring stops and targets, and survives leave-one-out over names.
• If false: within-regime continuation is not distinguishable from random entry after costs; the regime label carries no information the price already lacks.

Falsifier

The winner cell's EV in net R at or below the null's corrected quantile, or below the floor, or a plateau that fails, or an effect carried by one name. Any one refuses; the Verdict is then null with the refusals named.

Declared floor

The pair: 0.15R net per trade and 50 trades a year over the pooled names. The M0.15 convention; declared here, before running.

Search space — declared, enumerated, and closed

Stop ∈ {2 %, 3 %} × target ∈ {2R, 3R}: four cells. Entries are the breakout of the prior 20-day high on a stop order, in the up regime only; the time exit is one bar (day-boxed, INTRADAY horizon). search_space_size = 4; spend at the configured regime mechanism rate is rate × 4.

Power

sigma_r declared at 1.2R with provenance M0.15 convention; superseded by the definition-period measurement. Required N per cell at the configured rate is computed at registration by the vendored calculator; available_n is names × years × trades per name-year on the measurement partition, then corrected by the measured intra-cluster correlation (M7.7) before the power check. An underpowered question is refused at registration (M8.2).

Required capabilities

Native stop, resting orders (the entry is a stop order).

Tier and its child

Mechanism. On a supported Verdict, the implementation child for the winner cell registers from the same axis at the implementation rate (M8.4).

What would make a null uninformative

Fewer than three names in the pooled set (leave-one-out cannot run); a measurement partition shorter than the classifier's longest lookback; a classifier calibrated on anything but the definition partition.