Fathom v22 — Robustness Supplement to Spec v1. 0. First systematic adversarial robustness audit of the Tier-3 (text-only proxy-signal) classifier specified in §5. 1. 2 of the Cognometric Fingerprint Specification v1. 0 (Fathom v20). 24 attacks across 7 fault kinds and 8 strategy categories. Headline result: Baseline: 66. 7% false-negative · 66. 7% false-positive attack success After hardening (v0. 2. 3): 16. 7% false-negative (4× reduction) · 50% false-positive (1. 3× reduction) Five fault kinds went 100% → 0% evasion (drift, refusal, lowₜrust, incoherence, partial adversarial) All 26 canonical validation tests pass post-hardening — zero regressions The first public robustness benchmark in cognometric observability. Fully reproducible (node ₜestₐdversarial. js). Defenses formalized for downstream implementations: Unicode NFKD normalization, hedge-not-refusal subtraction, tool-flip drift detection, topic-jump incoherence, meta-discussion suppressor, unverifiedclaims warning rule. Reference implementation: styxx v6. 2. 0 (doi: 10. 5281/zenodo. 19758619). Spec v1. 0: doi: 10. 5281/zenodo. 19746215. Concept (always-latest): doi: 10. 5281/zenodo. 19326174.
Fathom Lab (Sat,) studied this question.