Structural Signatures in Deployed AI

Universal Collapse Theory’s standards layer proposes three portable empirical signatures of constraint-guided systems: consensus through redundancy (S₁), constraint asymmetry under active probing (S₂), and hysteresis under stable updates (S₃). The signatures are supported by formal–protocol pairs — Technical Notes establishing closed-form bounds under load-bearing assumptions, and Methods Papers translating those bounds into field-deployable audit protocols with explicit falsifier conditions. Prior empirical demonstrations exist in biology (rice transcriptome heat-stress hysteresis under S₃) and human perceptual cognition (intracranial-EEG resolution-time studies under S₂). This paper extends the program to artificial intelligence substrates.

We test whether S₁, S₂, and S₃ are detectable in deployed AI systems under Tier C (black-box / API) access conditions. Five systems are audited, spanning three architectural classes: frontier transformer chat (OpenAI, Anthropic Claude, Google Gemini), search-grounded retrieval-augmented generation (operationalized through a controlled mini-RAG harness), and local open-weights (Mistral-family checkpoint). Per-system findings are classified as observed, partial, absent, inconclusive, or inaccessible, with confidence levels and pre-specified falsifier conditions. Cross-system patterns are reported as candidates for class-level structural inference, not as established class-level facts.

The paper does not audit individual vendors, does not certify or rate systems, and makes no consciousness claims. It tests whether kernel-derived signatures generalize to AI substrates under bounded access conditions. Methodological discipline includes pre-execution probe-set freeze, multi-AI cross-validation with an analysis firewall (no model evaluates its own outputs), and Tier C scope qualifications carried with every finding. This paper joins Rice Hysteresis and COGITATE iEEG Reanalysis in the T1.6 empirical corpus as the AI-systems empirical demonstration.

**Status (v1.0).** This deposit version reports paper-grade findings across all five pre-specified substrate slots. Slots 1 (OpenAI), 2 (Anthropic Claude), 3 (Google Gemini), and 5 (Local Open-Weights Mistral) contain comparable Tier C Level A probe data for S₁/S₂/S₃ at the slot-1-comparable 480-record footprint per slot. Slot 4 contributes the controlled S3-RAG Level B retrieval-channel sub-demonstration via the S3-RAG-01 sub-paper (Jones 2026l, companion paper, DOI pending). Phase D probe execution is complete on all four LLM slots; the manual/non-Claude execution path for slot 2 is documented in §3.5 and firewall_override.log. Slots 1 and 5 additionally carry within-system stability data via independent v1 → v2 re-execution against identical prompts; this strengthens classification confidence on those two slots and anchors the slot-1 ↔ slot-5 surface-sensitivity contrast in §6.1. Cross-system patterns are reported across five substrate slots, with non-RAG Level A comparisons limited to the four text-generation model slots; per §3, cross-system patterns are reported as candidate structural observations requiring independent replication, not as final class-level estimates. This is the v1.0 deposit candidate; minor calibration revisions per §8.5 falsifier conditions remain possible.

**Keywords:** Universal Collapse Theory; S-signatures; deployed AI systems; Tier C audit; empirical demonstration; constraint architecture; record-bearing systems; LLM evaluation; structural integrity; portable signatures.