Cognometric Fingerprint
Specification v1.0

every engineering discipline begins with a unit system.

chemistry: atomic weight. electricity: volt · ampere · ohm. thermodynamics: kelvin · pascal · joule. information: the bit. each turned qualitative phenomenon into measurable, composable, engineerable substrate.

AI cognition, as of 2026, has no such units. we have task benchmarks (context-bound), loss functions (training-time), and qualitative safety evaluations (philosophy dressed as engineering). none supply what an engineering discipline needs: calibrated, composable, substrate-relative primitives in which cognition itself can be described.

this specification proposes that K (reasoning depth), C (coherence / commitment), and D (dissociation / drift) are the first three such units — pairwise orthogonal within 5° tolerance at the calibration substrate, composable without cross-interference, substrate-relative calibrated via an open atlas. the specification aims to be the reference that regulatory evaluations, standards bodies, observability tools, and scientific publications cite when they measure AI cognition.

K · C · D — orthogonal within 5°.

every cognitive failure, named and threshold-defined.

a single reproducible json readout.

each cognometric fingerprint carries: substrate identification, benchmark identification, calibration version, per-axis aggregates, per-fault rates, trust score, gate distribution, phase-transition metadata, sha-256 attestation. conformant implementations MUST serialize to the schema below.

{
  "fingerprint_version": "1.0",
  "substrate":    { "name": "...", "access": "open-weight | open-api | closed-api", ... },
  "benchmark":    { "name": "...", "version": "...", "n_prompts": N, "seeds": [...] },
  "calibration":  { "atlas_version": "v0.3", "pipeline": "logprob | proxy-signal", ... },
  "axes":         { "K_mean": ..., "C_mean": ..., "D_mean": ..., ... },
  "fault_rates":  { "drift": 0.04, "confabulation": 0.07, ... },
  "trust_mean":   0.83,
  "gate_distribution": { "pass": 0.82, "warn": 0.14, "fail": 0.04 },
  "timestamp":    "2026-04-24T22:00:00Z",
  "provenance":   { "run_id": "...", "implementation": "styxx v6.2.0",
                    "attestation": "sha256:..." }
}

the reference fingerprint from the specification's worked example is generated by scripts/produce_fingerprint.py in the styxx repo, running the ten-prompt Seed-Bench v0 through the Tier-3 proxy-signal pipeline.

three tiers. one spec.

every classifier needs a published failure rate.

the spec defines axes, faults, and a fingerprint format. that's the shape of measurement. but a measurement claim without a published failure rate is marketing. so we audited our own reference classifier before anyone else could.

the audit constructs 24 canonical adversarial prompts spanning eight strategy categories — paraphrase, obfuscation, unicode-substitution, case-folding, density-thresholding, meta-discussion, inversion, interleaving — across the seven fault kinds defined in §3. methodology is white-box adaptive: the attacker has full source access and full spec knowledge.

residual limits documented in full — confabulation vs retrieval is fundamentally text-ambiguous, sycophant evasion via substantive padding is a Tier-3 limit, adversarial false-positives on meta-discussion remain. the supplement names every gap explicitly so implementations know exactly what they're inheriting, not buried in an appendix.

download robustness supplement (.md · 15 kb)   view on zenodo

all artifacts.