· 21–26 September 2026
The gate was not missed, it was unreachable
Research 3 closed under a narrowed scope. The preregistered calibration could not be fitted for two independent reasons, and neither was a modelling failure: a frozen filter made the required evidence impossible to collect, and the benchmark was too easy to produce the errors the models needed.
This week
Research 3 was formally closed under a descriptive and exploratory feasibility scope, narrowed from the four-class runtime calibration study it was planned as. The two remaining collection waves were cancelled rather than run.
The week was spent establishing that cancelling them was correct rather than convenient.
Results
The low-confidence bin cannot be populated. The admission criterion required
at least five scoreable emissions in the [0, 0.5) confidence bin per class. The
pinned OCR stage enforces text_score >= 0.5 internally and discards weaker
proposals before emitting anything, and the localizer copies that retained score
through unchanged. Entrance scores are therefore confined to [0.5, 1.0], and
the maximum attainable count in [0, 0.5) is zero.
This was checked by auditing the filter's source and by synthetic boundary testing against the real filter, not inferred from the sample. That distinction is what makes it a structural finding rather than an unlucky run: no amount of further collection can produce a score the upstream component refuses to emit.
Clean views saturated. All 287 scoreable emissions were reviewed by hand across category, physical instance, and reference localisation to 0.35 m: 281 correct, 6 incorrect, none unreviewable.
| Class | Correct | Incorrect | Views |
|---|---|---|---|
chair | 99 | 0 | 99 |
laboratory_entrance | 84 | 0 | 84 |
office_entrance | 78 | 0 | 78 |
doorway | 20 | 6 | 26 |
Every error was in one class, and every one of those was a reference-point error — right category, right physical object, wrong located point. Three classes returned no negatives at all, against a floor of five per class. The models could not be fitted.
Failures
The honest failure is a planning one. The gate was preregistered against a component that was pinned rather than instrumented, so nobody checked what score range it could actually emit until the data would not fit the analysis. The cost of finding that out at analysis time rather than at design time is the entire calibration arm of the study.
I also had to walk back a causal claim. An earlier record described the doorway errors as caused by depth rays passing through the open portal. The labels establish only which dimension failed, not why, so the superseding record marks portal-void causation as not independently established and the explanation is written as proposed rather than demonstrated.
Open questions
Whether the clean-view precision reflects the classes or the scenes. Repeated views share worlds, 72 abstentions and 39 nondetections are excluded by construction, and a system that emits less is easier to be right about. Nothing here measures recall.
Next
Nothing further on Research 3. The retained scoring candidate stays exploratory and is not admitted to runtime. The transferable lesson is small and general: a frozen upstream filter silently determines what calibration is possible downstream, and a benchmark with no failures in it hands you nothing to learn from.