Every Mind Leaves Vitals: On the Cognometric Layer, Substrate-Independence, and the One-Time Choice We Have

Every mind leaves vitals. This paper extends the Cognometry Manifesto beyond LLMs. We show that three independent calibrated cognometric instruments — hallucination (AUC 0.998), refusal (AUC 0.976), tool-call drift (AUC 0.916) — all exhibit phase-transition structure under feature-count ablation: detection of cognitive states does not scale smoothly with classifier capacity but jumps discretely at a critical feature, replicated across three independent feature bases. We argue this result is a property of cognition rather than of language models, and bridge to thirty years of independent evidence from forensic linguistics, computational psychiatry, and crisis-text classification that the same observability layer applies to biological cognition. The substrates differ; the observability does not. The paper records six constitutional commitments for cognometric instruments shipped under the Fathom name — MIT license, weights and reproducers in-tree, failure modes declared in-weights, calibration fingerprint required, CPU and browser-runnable, no private detectors under the Fathom name. The constitutional commitments are written so that if they are ever broken, the paper makes the breaking visible. The paper is falsifiable on its sharpest claim: any calibrated text-based cognitive-state detector whose feature-count ablation shows smooth AUC scaling without a critical-K jump would falsify the central result, and we will retract or amend at the same DOI. Reproduce. github.com/fathom-lab/styxx — every number reruns from random_state=0 in under five minutes on CPU. Run live. fathom.darkflobi.com/cognometry Read on the site. fathom.darkflobi.com/every-mind-leaves-vitals