Occams — a falsification lab for retail trading hypotheses

Three programmes, six questions, five verdicts — four null, one supported whose entry added nothing — and one refusal at measurement. Every page here is rendered from a hash-chained Register; none carries a script, a network reference or a money figure. The lab's product is its refusals.

Pages

Conclusions

Surveys

• Survey grid-001-seed20260912
• Survey grid-002-seed20260920
• Readiness under the fifth check — grid-001 seed 20260912
• Readiness under the fifth check — grid-002 seed 20260920

Documents

• A strategy that passed four checks and turned out to be the market
• Extending the lab — a new entry kind, a new grid, your own bars
• How this lab prevents p-hacking
• M0 answers
• PUBLICATION
• The re-score of October 2026
• Occams — runbook
• Setup — clone, configure, run, and burst
• The inference review of 2026-10-03, verified by rerun
• occams-test-lab — FIXES-AND-TESTS

Decisions (ADRs)

• Hypothesis and Strategy are separate aggregates
• The spec that measured is the spec that trades
• Every Strategy declares a stop, and R is planned risk
• Unadjusted prices, with corporate actions as a separate series
• The regime classifier is frozen before any strategy question
• History is partitioned, and the reserve permits one look per spec hash
• The spec hash is identity, not measurement context
• Two event stores, split on publishability
• Approval is a signature, and the proposer never runs on the execution host
• The retirement rule is declared at approval, not decided during drawdown
• Alpha exhaustion is terminal until new data, and replenishment is mechanical
• A declared instrument universe is pooled and costs one search
• Second looks are detected by observation overlap, not prevented
• Forward testing falsifies; it does not confirm
• Approval requires clearing a conservative cost bound
• The proposer emits drafts and cannot spend alpha
• No alpha is allocated to an axis that cannot run
• Required capabilities are identity; the venue is context
• The portfolio envelope binds, and correlation is measured at approval
• Time is an instant, never a date
• The archive stores the bars, not a reference to them
• Signed approvals are the only thing that crosses to execution
• The forward window is declared in trades and in time
• Halts are durable, and clear on their own terms
• The universe is a point-in-time rule, not a list
• FX drift is an exposure, not an edge
• Hypotheses come in two tiers: mechanism and implementation
• The data budget chooses the instrument class
• The apparatus must prove it can accept, not only that it can refuse
• The execution host is a durable process, and it is priced before M1
• The forward period runs before approval, not after
• The forward window executes for real, at minimum size
• The lab has its own falsifier
• Assumptions are discharged in configuration, not in a proof assistant
• Autoresearch: the shape is adopted, the keep rule is refused
• The Verdict freezes the winner cell's hash; the template names the family
• The entry enum is extended by decision, one mechanism at a time
• The calendar is frozen: partitions are cut from a recorded span, never the archive's live one
• The second programme runs beside the first, in the same repository, on its own Register
• A resting limit that the bar opens through fills at the open
• The fill audit is exhaustive, and an unobtainable fill is a missed trade
• Flat bars are named in the record, not cut out of the calendar
• Beating always-long at the same geometry and gate is the fifth check
• No alerting subscription and no execution host inside the cap: FORWARD is a record
• The surface the sweep optimises is the margin over always-long
• The third programme asks whether a margin the screen finds transfers
• Fixes bind forward, and a re-score is a diagnostic, never a verdict
• The null is drawn at the winner's count, by calendar date, with p counted plus one
• The fifth check is side-matched and dependence-aware
• The floor is cleared by a lower confidence bound, and power is planned against a declared alternative
• A plateau is judged in standard errors, without the winner in its own median
• The Register's heads are pinned, and attested outside the repository
• The engine is identified by content, and a dirty tree cannot measure

Drafts — no standing

• DRAFT-001 — Forced selling at index deletion
• DRAFT-002 — High-conviction disclosure drift
• DRAFT-003 — Within-regime continuation on the confirmed class
• DRAFT-004 — Short-term reversal on daily bars
• DRAFT-005 — Reversal after a streak
• Survey candidate 3fff965a58639eae — down_run on dow_30
• Survey candidate 4d1eee50721ff72a — pullback_in_trend on sector_etfs
• Survey candidate 4f0326e3644df864 — down_run on dow_30
• Survey candidate 5193d48bac341ff2 — down_run on index_etfs
• Survey candidate 64d656ca7e638722 — pullback_in_trend on sector_etfs
• Survey candidate 6fe5444b518075b9 — down_run on dow_30
• Survey candidate 719063aad2318c00 — pullback_in_trend on sector_etfs
• Survey candidate 78e13cce5e21fb83 — down_run on index_etfs
• Survey candidate 7ac7bd72aaa1e5f4 — down_run on sp_100
• Survey candidate 7d8f04aa2066452a — down_run on dow_30
• Survey candidate 8835b76f08d63847 — return_below on index_etfs
• Survey candidate 89200ec086be73e6 — return_below on index_etfs
• Survey candidate abd742abc807e288 — return_below on index_etfs
• Survey candidate bb3394ed0820ae72 — down_run on sp_100
• Survey candidate c00c0c7e6fe3d5fb — down_run on sp_100
• Survey candidate e02384ae41677547 — pullback_in_trend on sector_etfs
• Survey candidate edcf23b4b5e9b4f6 — down_run on dow_30
• Survey candidate f2199b5b297d85bb — down_run on sp_100
• Survey candidate f8500a69db544a3c — down_run on sp_100
• Survey candidate f9d09f4299ed5aa7 — down_run on sp_100
• Survey candidate 0dc23760a36e1a1b — breakout_low on sp_100
• Survey candidate 151ea6728ab85559 — breakout_low on index_etfs
• Survey candidate 2f89cd3fc3abbd83 — breakout_low on sp_100
• Survey candidate 387611fc36db0b29 — breakout_low on dow_30
• Survey candidate 42f89ac43ea32bac — breakout_low on dow_30
• Survey candidate 4d1eee50721ff72a — pullback_in_trend on sector_etfs
• Survey candidate 5193d48bac341ff2 — down_run on index_etfs
• Survey candidate 5834abf51248fbb5 — breakout_low on index_etfs
• Survey candidate 5d78cf45a8e7aa9a — breakout_low on sp_100
• Survey candidate 5da80c610ab6da80 — breakout_low on dow_30
• Survey candidate 69ecf3439041f4eb — breakout_low on sp_100
• Survey candidate 71a0d8161a682950 — breakout_low on sp_100
• Survey candidate 8373d84db5ffd36b — breakout_low on dow_30
• Survey candidate 83acca56f082d646 — breakout_low on dow_30
• Survey candidate 8e17b913b3c624f2 — breakout_low on index_etfs
• Survey candidate 9c62b6cd8da2bbb3 — breakout_low on sp_100
• Survey candidate 9ed03d49b0d204e4 — breakout_low on sp_100
• Survey candidate e02384ae41677547 — pullback_in_trend on sector_etfs
• Survey candidate f2199b5b297d85bb — down_run on sp_100
• Survey candidate f9d09f4299ed5aa7 — down_run on sp_100