Occams · drafts / survey / grid-002-seed20260920

status: CANDIDATE — generated by `python -m occams survey candidates` from survey record #9 (grid-002, seed 20260920); no standing, no alpha spent; registration is the author's `--yes` naming this cell's id (M12.5) tier: mechanism axis: regime relates_to: "TASKS-v4.md M12.5 · docs/adr/0039 · docs/surveys/grid-002-seed20260920.html#cell-0dc23760a36e1a1b"

Survey candidate 0dc23760a36e1a1b — breakout_low on sp_100

A Draft: 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 on the measurement partition. Nothing here has cost anything. It was generated from a survey cell; the screen is not a verdict, and the shrinkage from screening 38,976 cells is recorded per question (N6, M12.6).

Mechanism

A buy-limit at the lowest low of the prior 20 bars, held for 20 bars, carries positive EV in net R on an S&P 100 constituent as listed on 2026-09-12 because a fall to a new low is over-sold and bought back (reversal at a new low) — inside the up regime of the classifier frozen on SPY (F11)

The survey's cell file recorded the sentence as: A buy-limit at the lowest low of the prior 20 bars, held for 20 bars with no regime gate, carries positive EV in net R on an S&P 100 constituent as listed on 2026-09-12 because a fall to a new low is over-sold and bought back (reversal at a new low) — inside the up regime of the classifier frozen on SPY (F11) — the proposer's wording on the day it ran; the sentence above is the grid's now, and registration stamps the recorded one beside it.

Universe sp_100, regime gate up, breakout_low(lookback=20) · limit, multi_day · hold 20 · stop percent 2 · target 2R.

Both interpretations

If true: the winner cell clears the declared floor and beats random entry at the corrected alpha on the measurement partition — the mechanism the survey screened is real after costs on this universe.
If false: the winner cell is not distinguishable from random entry after costs, or does not clear the floor — the screen's margin was selection, and the shrinkage says by how much.

Falsifier

The winner cell's EV in net R at or below the null's corrected quantile, or below the floor the author declares at registration, on the measurement partition.

What the survey showed — definition partition only, zero alpha, not a verdict

tier held · 5,897 trades · EV gross +0.226, net +0.175 at bounded costs
always-long at this geometry -0.085 over 38,837 · margin +0.260
eras +0.327 / +0.152 / +0.075 · weakest name held out +0.166 · missed entries 16
same-day ρ 0.11 · N eff on the measurement partition 8,840
lowest affordable floor at k = 4: 0.075 R; at k = 9: 0.075 R

Search space

The family's sweep — hold 1, 3, 5, 10, 20, stop 2, 3, 5, target 2, 3 — 30 cells; the accountant prices it at 0.3 alpha on the regime axis (rate × 30 cells).

Registration — the author's act

python -m occams question register --from-survey archive/surveys/grid-002-seed20260920-c6i.32xlarge --grid surveys/grid-002.toml --ids 0dc23760a36e1a1b \
    --archive archive --register register/programme-2.jsonl --config configs/programme-2.toml \
    --floor-ev FLOOR --floor-frequency TRADES_PER_YEAR --by NAME --yes

The floor is declared there, by the author; this document supplies none.