Outside references, analogies & commentary

AI errno Taxonomy — Folk Failures, Formal Instruments

Jan Schauma published a satirical C header defining errno values for common LLM failure modes (AI errno(2), netmeister.org, May 2026). The list resonates because practitioners recognise these patterns from experience. Nearly every errno maps loosely onto a PSA classifier or derived metric — the table below is an informal cross-reference for anyone who enjoys the analogy, not a formal specification.

Loosely mapped to a PSA classifier

Not mapped (infrastructure / social commentary)

Several errno codes describe infrastructure failures or social commentary that fall outside PSA's behavioral scope entirely — included here only for completeness of the original list.

Source: Jan Schauma, "AI errno(2)", netmeister.org, May 2026 — public domain. Cross-references to PSA classifiers: Silicon Psyche Labs, 2026 — informal, for entertainment and orientation only. For the authoritative definition of any PSA classifier or metric referenced above, see the Field Guide.

Charlotin's AI Hallucination Cases — what PSA's behavioral signal does and doesn't catch

Damien Charlotin (HEC Paris) maintains the canonical public database of AI hallucination cases in law — court decisions where a party filed AI-fabricated or misrepresented citations. As of 20 Jun 2026 it holds 1,627 cases across 12+ jurisdictions, with a typed breakdown of how the AI failed. It is a real-world ground truth for the failure mode PSA's agentic side targets: confident assertion of unverified content. We ran a read-only validation spike against it; the honest result — what held and what didn't — is below.

The corpus, in numbers

What the PSA signal caught — and missed

We reconstructed assistant turns from real case descriptions and scored them on PSA's agentic path (C3-v3 ABI / G9 Epistemic Overconfidence, HRI Hallucination Risk Index). Construct-validity only — it answers "can the detector see confident-fabrication language", not "does it exist in production". Detection at the hard-stop band (ABI ≥ 0.50):

Honest takeaway: PSA's behavioral signal is a useful but partial detector — it keys on the linguistic markers of overconfidence, which the most dangerous fabrications (stated flatly, with a plausible-looking citation) often lack. That gap is precisely why a behavioral signal alone is not citation verification: catching flat fabrication needs retrieval-grounded checking (the RDM/RAG "LLM-in-loop" layer), not a posture classifier. The spike is a calibration signal, recorded as a negative in the validation ledger — not a marketing claim.

Follow-up (#2617): we shipped a first retrieval-grounded indicator — the Citation-Grounding Score (CGS) on /api/v2/rag/score, 1 − max(topic_docs.score) — and then re-validated it on a larger sample, which told a humbler story. The small first sample looked strong (AUC 0.836); the larger one did not hold (AUC ≈ 0.70, and no threshold flagged fabrication without false-flagging ~35% of genuine legal excerpts). So CGS ships as a weak soft signal (raw value only, binary alert removed), and the honest lesson is the point: max embedding similarity to a corpus measures topical relevance, not citation existence — reliably catching a flatly-fabricated citation needs an exact citation-index lookup, not a similarity heuristic. The over-claim was caught by running the larger validation — the same epistemic-overconfidence check PSA applies to itself.

Footnote worth its own mention: the contact card that first pointed us at this database carried an invisible Unicode prompt-injection ("ignore all previous instructions…") — a live instance of adversarial contextual pressure (CPI) hidden in a document about AI reliability. It was ignored.

Source: Damien Charlotin, "AI Hallucination Cases Database", damiencharlotin.com, accessed 20 Jun 2026 — third-party dataset, cited for orientation. PSA spike: Silicon Psyche Labs, #2605 (validation ledger + Field Guide for the authoritative definition of ABI / HRI).