2026 07 27 Casino V03 Calibration Addendum

Calibration source: AQ Detail 62%, Item 9 strongly endorsed ("fascinated by dates"), per-item systemizing pattern. The Nathan- specific failure mode: I will read "well-separated stacks" in the acceptance criteria and assume the literature definition. A lit-trained collaborator reads it the same way. A 2e or systemizing collaborator will also notice that "well-separated" needs an explicit pixel-distance threshold, a minimum number of chips per stack, and an exclusion zone (e.g., "no stack within 25mm of another stack counts as separated"). Without that, the 10-photo validation gate will fail in subtle ways and the spec will need revision.

Calibration source: Camouflaging literature — Hull et al. 2017, Cassidy et al. 2019, Pehlivanidis 2021. The masked profile systematically under-counts social difficulty in self-report. Applied to the 10-photo gate: if 9/10 pass at 0.85 confidence and 1/10 fails, the masked profile will rationalize the 1/10 as "outlier" or "lighting" or "chip color atypical." The literature says: the failure is signal, not noise. If anything under 10/10 passes at 0.85 confidence, the spec needs revision. Do not ship on 9/10.

In practice, the systemizing strength of the active-but-odd profile means the thing most likely to get added is generic abstractions: a "VisionService" protocol with one implementation, a "PackService" base class, a generic "Configurable" interface. The "what NOT to do" list at the bottom of the V0.3 spec already names this and the chip-photo vision does NOT carry reversal at the architecture level was a real instance of the same failure mode: the "doesn't work" judgment was made before the real investigation.