The frozen decision contract
How a result becomes official here — and why an interesting signal found afterwards does not change one.
STOP-v1
STOP-v1 is a decision contract that was written and locked before any results were seen. It defines which axes may be tested, how they must replicate, what threshold counts as a pass, and what happens when an axis cannot be tested at all.
- Calibration
- Calibration scans every axis and every level. Its job is to find a structurally usable difficulty candidate, d*. It selects; it does not decide.
- Confirmation
- Confirmation runs only on axes where calibration found a usable d*. It uses two disjoint item forms, A and B, against the same checkpoint.
- Frozen thresholds
- The thresholds were locked before the results were seen, including MIN_DIFFICULTY_DROP = 0.10 applied to the locked d* neighbours inside confirmation_A.
- Multiplicity
- The within-axis rule is max(p_A, p_B). Across the eight locked axes the correction is Holm-Bonferroni.
- Fail-closed behaviour
- If an axis has no usable d*, no confirmation is produced. That is the protocol failing closed, not missing data. The alternative would be choosing a level after seeing results, or forcing an unsuitable test.
Failing closed is not missing data
If an axis has no usable d*, no confirmation is produced. That is the protocol failing closed, not missing data. The alternative would be choosing a level after seeing results, or forcing an unsuitable test.
- In Model-0 only M, I and A proceeded to confirmation.
- In the Pilot-0.2 final chain only R, I and A proceeded to confirmation.
- Axes that could not proceed were not deleted from the chain; they received a failing AxisReport representing all eight axes.
- T had no d* because it is not ordinal.
An axis that could not be tested still produced a report. Eight axes went in and eight axes came out; some of them came out as failures of measurability rather than failures of the model.
A frozen decision stays frozen
An interesting post-hoc signal does not modify a frozen decision after the fact.
After the Model-0 results were in, three axes showed replicated above-chance signal. That is a genuinely interesting finding, and it is published in full on the Model-0 page.
It did not change the verdict. The frozen difficulty-drop condition was not met, so the axes were not declared measurable, and the decision remained STOP with 0 of 8 measurable axes. The thresholds were not adjusted afterwards to accommodate what had been seen.
This is the entire point of writing the contract down in advance. A protocol that can be relaxed once the results are visible is not a protocol.
Source: Scientific Evidence Inventory v1.0 (evidence cutoff 30 August 2026), §5.1, §8.1.