BL-D-003: The Pre-Registration Standard
BIJECTIVE LABS
BL-D-003: The Pre-Registration Standard
Status: RELEASED
Classification: PUBLIC
Revision: 1
Date: 2026-09-07
1. What pre-registration is for
Writing down the metric, the sample, the method, the cost model, and the kill threshold before seeing any data removes the freedom to move the goalposts afterwards.
That is the entire mechanism. It is not a paperwork exercise and it is not about rigour as a posture. It is about the specific, well-documented human tendency to find a threshold that the data has already passed, and to do so sincerely, without any sense of having cheated.
The programme produced 58 pre-registrations across two programmes. The discipline was adopted partway through the first one, after a momentum result of +338 basis points at a p-value of 0.001 was traced to a regime that occupied the final third of its sample. Nothing about that analysis was dishonest. The threshold simply had not been fixed in advance, so there was nothing to fail.
2. Required contents
A pre-registration is not valid unless it states all nine.
| Element | Requirement |
|---|---|
| Mechanism | One sentence naming who is structurally forced to lose, and why |
| Claim | The falsifiable statement, in a form data could refute |
| Metric | The exact quantity to be computed |
| Horizon | Fixed. No horizon shopping afterwards |
| Universe | The instrument set, and the rule that produced it |
| Era | The data range, and which held-out eras remain sealed |
| Cost model | Every term, with its source |
| Power | The sample size required to detect the claimed effect |
| Kill criteria | Numeric thresholds, and the verdict each one fires |
The kill criteria are the part that gets diluted if it is not enforced. A threshold written as "the result should be clearly positive" is not a threshold. The form required here is a table with a metric, a numeric bound, and the verdict that follows.
| Metric | Threshold | Verdict if crossed |
|---|---|---|
| Net return, CI95 lower bound | at or below 0 | NOT PROVEN |
| Net return, point estimate | below 0 with power achieved | FALSIFIED |
| Permutation p-value | at or above 0.05 | FALSIFIED |
3. Amendment
A pre-registration may be amended, and three in this record were. The rules are narrow.
An amendment is a dated document that supersedes nothing. Both the original and the amendment are published, and the dossier shows both.
An amendment may correct an instrument, tighten a threshold, or record that a measurement turned out to be unavailable. It may not loosen a threshold after data has been seen, and it may not change a horizon.
The amendment states why. A reader must be able to judge whether the change was forced by circumstance or by an inconvenient result.
One amendment in this archive permanently capped a proxy validation at WEAK
after the proxy achieved a pooled correlation of 0.70 at five-minute resolution
against 0.08 at one hour [E-0018]. The cap was recorded as permanent
specifically so that a later reader could not treat the proxy as validated.
4. Registration does not confer merit
A pre-registered hypothesis is not a better hypothesis. It is a hypothesis whose failure will be legible.
This is worth stating because pre-registration culture can drift toward treating
the document as the achievement. In this programme the discipline coexists with
a hard budget: R-10 caps the second programme at twelve tested hypotheses in
total, and eight were used. The budget exists because a programme that can
pre-register without limit has simply relocated the multiple-testing problem
into the count of registrations.
5. The sealed era
Pre-registration governs what is tested. The sealed era governs when.
The substrate is partitioned before any research begins:
| Era | Range | Use |
|---|---|---|
| Held-out A | Substrate floor to 2021-12-31 | Sealed |
| Research | 2022-01-01 to 2025-06-30 | All development and iteration |
| Held-out B | 2025-07-01 onward | Sealed |
The seals are opened once, at a single pre-registered gate, against a single specification, with the acceptance criteria fixed in advance. There is no second attempt.
The programme opened its seal once. The hypothesis had survived research-era validation, survived conditioning, and beaten a permutation null by 5.83 sigma. On the held-out data it returned -8.99 basis points against +8.24 in research [E-0037]. The mechanism had inverted.
That outcome is the argument for the whole apparatus. Every gate before the seal had passed. Only the sealed era caught it, and it could only catch it once.
6. What a pre-registration cannot do
It cannot make an underpowered test conclusive. If the declared power is not
achieved, the verdict is NOT PROVEN whatever the point estimate says, and
R-06 enforces that.
It cannot protect against a parameter nobody knew was free. Nineteen
pre-registrations in this record fixed a rebalance hour without recording it as a
choice, because nobody had recognised it as one. Pre-registration locks the
parameters you have thought of. R-09 exists because that set is smaller than
you think, and the fix was to enumerate the phase space automatically rather
than to rely on remembering.
That limitation is not an argument against the practice. It is an argument for treating the pre-registration as a floor rather than a ceiling, and for auditing completed work against rules that did not exist when it was done.
Copyright (c) 2026 Bijective Labs. All rights reserved.