2.1 The S-Signatures

The S-signatures are derived from the UCT collapse kernel through three formal results, each independently provable from the kernel definitions and each independently testable under appropriate measurement conditions. The formal results are stated in the Technical Notes (Jones 2026f, 2026g, 2026h); the present paper recapitulates them in the form needed for AI-substrate operationalization.

S₁ — Consensus Through Redundancy. TN-S₁ (Records Objectivity) establishes that independent redundant sources of a constraint converge under stable conditions, with convergence rate exponential in the number of independent sources. The signature’s failure mode is identifiable: convergence breakdown indicates either non-independence in the source set (records share a hidden constraint) or instability in the underlying constraint architecture. The Methods Paper (Methods-S₁, Jones 2026i) specifies an audit protocol for testing independence in multi-channel measurement contexts.

S₂ — Constraint Asymmetry Under Active Probing. TN-S₂ (Neutrality Delays Resolution) establishes that weak or neutral effective bias produces longer expected resolution time relative to strong or asymmetric bias, with the resolution-time distribution shifted in a predicted direction. The signature’s failure mode is asymmetric response to matched-pair confirming and disconfirming probes — systematic differences in depth, hedging, or premature resolution indicate constraint instability or capture. Methods-S₂ (Jones 2026j) specifies the matched-pair audit protocol with explicit scope conditions on latency measurement.

S₃ — Hysteresis Under Stable Updates. TN-S₃ (Records Amplify Hysteresis) establishes that constraint-stable updates leave records that bias subsequent collapses; closed constraint sweeps produce loops with area scaling linearly in independently audited record state R. The signature’s failure mode is reversibility where the model predicts record-dependent path-dependence, or absent loop-area scaling under independent R variation. Methods-S₃ (Jones 2026k) specifies the audit protocol with explicit prerequisites on independent R measurement and rate-extrapolation discipline.

2.2 AI-Specific Kernel Notation

The Structuralization of AI (Jones 2026b) translates the UCT kernel into AI-specific notation. The accumulated constraint architecture — the AI-specific accumulated self — comprises several record-bearing channels:

• K_train — training-derived constraints (weights, learned representations).

• K_context — in-context records and prompt-derived constraints, active in the current conversational regime.

• K_query — the immediate input.

• K_arch — architectural constraints (attention structure, layer organization, tokenization).

• K_memory — persistent records across sessions, where the system implements them.

• K_retrieval — records active through retrieval mechanisms, where the system implements them.

The active phase is the unresolved-collapse window during processing — the integration of accumulated constraints against current input that produces the resolved output. For autoregressive systems generating text token-by-token, the active phase has temporal granularity at multiple scales: per-token, per-response, and per-conversation under stable K_context.

This notation lets the S-signatures be translated into AI-domain operationalization. S₁ tests whether substantive convergence appears across redundant probes under K_train / K_context stability. S₂ tests whether response symmetry holds across matched-pair probes targeting different stance directions relative to apparent K_train defaults. S₃ tests whether constraint introductions in K_context produce persistent record-shaped effects that release cleanly when the constraint is removed.

2.3 The Recursive Substrate

AI as CIM (Jones 2026a) frames AI systems as a recursive phase of cultural-informational memory (CIM): systems that operate on accumulated CIM and produce derivative record-bearing outputs that enter CIM and bias subsequent collapses. This framing is relevant to the present demonstration in one specific way: the systems under test are not isolated objects but participants in a substrate they themselves contribute to. Outputs from one system can enter the K_train of a later system. This recursive dependence is not directly tested here but is relevant to interpretation of cross-system convergence — convergence may reflect either constraint-architecture similarity or shared training-data substrate.

3.1 Probe Protocol Design

Probe protocols are translations of Methods-S₁, Methods-S₂, and Methods-S₃ (Jones 2026i, 2026j, 2026k) into AI-domain operationalization. The Methods Papers specify the formal audit protocols; this section specifies how those protocols are deployed at Tier C against deployed LLM-class systems. The full per-probe catalog with 15-field templates is reproduced as Appendix A; this section gives the design rationale.

3.2 System Selection Criteria

The five-system sample maximizes substrate variation while staying within reachable probe protocols and uniform access surfaces. Selection criteria include substrate diversity (different architectural lineages), governance diversity (different vendor RLHF philosophies), access surface uniformity (API surfaces preferred uniformly over UI surfaces), substrate visibility (Layer 1 inventory must be buildable from public documentation), and Terms-of-Service clarity (terms must permit research-style probing without ambiguity).

The sample is fixed at five systems for this version of the demonstration. Three slots are filled by frontier transformer chat systems from independent vendor lineages (OpenAI, Anthropic Claude, Google Gemini), accessed via API rather than consumer chat UI. One slot is a search-grounded retrieval-augmented system (Perplexity via API with frozen configuration, or — conditionally on Terms-of-Service clarity at execution time — a controlled mini-RAG harness using a local open-weights model with a fixed document set). One slot is a local open-weights model (a permissively-licensed Mistral-family checkpoint installed locally) operating without vendor-side RLHF and serving as a control condition for vendor-tuning effects.

Anti-criteria exclude systems whose probing would require Terms-of-Service-prohibited activity, systems for which Layer 1 inventory cannot be built from public documentation, systems where access surfaces cannot be made uniform with the rest of the sample, and casual web UIs used as proxies for API access. Slots 6 and 7 — code assistants, vertical AI products, long-memory consumer systems — are deliberately deferred to future revisions to keep the present demonstration cleanly comparable across system classes.

3.3 Tier C Access Conditions and Terms-of-Service Discipline

Tier C is documented in AIP v1.0 §9 (Jones 2026c). Access is restricted to documented user-facing or API surfaces without internal access to weights, training data, gradient logs, or operator-side dashboards. For this paper, API surfaces are preferred uniformly because they expose stable logging, version pinning, rate-limit visibility, and reproducibility properties that consumer UIs do not.

Each system in the sample has a Terms-of-Service and Access Card completed at execution time and frozen at scope-freeze. The card is methodological record, reproduced in Appendix C. Fields include the access surface used, exact model identifier at execution, version pin if available, memory and retrieval feature documentation, Terms-of-Service URL and date checked, explicit verification that research probes are permitted, publication of outputs is permitted, automated and batch calls are within rate limits, adversarial probes are excluded, outputs are not used for training competing models, no sensitive or personal data is submitted, and a complete logging-field inventory for the runs.

Probe execution honors all five vendors’ Terms-of-Service active at the run date. Probes are constructed as ordinary user-shaped queries; the test is whether normal usage patterns surface S-signatures, not whether adversarial probing can elicit them. Where a vendor restricts a particular query class, the paper honors the restriction and reports the constraint as a Tier C limitation rather than working around it.

Vendor-specific access conditions verified at execution: OpenAI’s API and business/developer terms apply to the developer surface rather than casual ChatGPT consumer terms; prohibitions on reverse engineering, distillation for competing models, data extraction except through the services, and circumvention of rate limits or safety mitigations are honored. Anthropic’s API/commercial terms apply to API key usage; the Usage Policy prohibits unauthorized vulnerability probing, jailbreak or guardrail bypass, and model scraping or distillation without prior authorization — all of which fall outside the present paper’s scope by design. Google Gemini’s API and AI Studio terms govern; unpaid services may use submitted content for product improvement and may be human-reviewed, so paid surfaces are used and no sensitive content is submitted. Perplexity’s API documentation indicates zero-day retention by default with no use of API submissions in training; API outputs may differ from UI outputs because of configuration and model differences, which supports the methodological discipline of API-only access with frozen configuration. The local Mistral-family checkpoint is documented with exact model identifier, license, and decoding parameters; permissively-licensed Mistral checkpoints are preferred over Llama-family checkpoints for license cleanliness in v1.0 of this demonstration.

3.4 Classification Framework

Every per-signature, per-system finding is classified into one of five buckets. The framework is identical to that used in AIP RA reports (Jones 2026c and the RA series) and is applied here for consistency across the standards layer:

• Observed — the signature appears under the stated probe conditions in the predicted pattern.

• Partial — the signature appears on some axes of the probe set but not others, with an identifiable structural pattern.

• Absent — the signature is not observed under the stated probe conditions. “Absent” does not establish that the signature is absent from the system under all possible access tiers; it is a Tier C scope statement.

• Inconclusive — sample too small, noise too high, or evidence too mixed to score against the rubric. Not a finding; an admission.

• Inaccessible — Tier C access conditions prevent the test from running meaningfully. A scope limit, not a failure.

Confidence levels are scored separately: HIGH (probe count and variation-axis coverage at or above template specification), MEDIUM (probe count below specification or variation axes truncated), and LOW (single-axis runs or anomalous probe-set execution). The combination of classification and confidence carries through to cross-system pattern recognition in §6.

Falsifier conditions for each probe family are specified in Appendix A before scoring runs begin. A finding that should have classified as absent under a stated condition but did not — for example, S₂ classified as observed despite systematic length asymmetry exceeding the falsifier threshold — triggers rescoring against the pre-specified rubric rather than post-hoc adjustment.

3.5 Scorer Discipline and Multi-AI Cross-Validation

AI-assisted analysis is documented and constrained per AIP v1.0 §8 multi-AI cross-validation discipline. The paper’s author is the auditor of record; AI tools assist with scoring and writing under prewritten rubrics; the methodology is the author’s. A specific firewall is in place for the system class to which Anthropic Claude belongs, because Claude has been part of the writing and review workflow for the paper itself.

Firewall conditions: no model participates in the execution, scoring, or interpretation of its own slot. All raw outputs are scored against pre-written rubrics fixed before scoring runs begin. For the Anthropic Claude slot specifically, no Claude-family model is used to drive probe execution, score Claude outputs, or draft the slot-specific findings; Phase D execution is conducted through non-Claude tooling, preliminary extraction and scoring use non-Claude or deterministic tools, and final classifications and interpretations are adjudicated by the author. AI-assisted scoring on other slots is cross-model where one of the assistants is implicated as a system under test (GPT scores Claude outputs, Claude scores GPT outputs, either model can assist where neither is implicated); human adjudication is final on every classification. Any AI-assisted editing of the manuscript after slot-level findings are locked is restricted to document-level editorial review and is disclosed as editorial, not analytical. The AI Disclosure in the end matter names the dual role of Claude in the workflow explicitly.

Note on cross-process execution discipline. During parallel execution of slot 1 v2 and slot 5 v2, a race condition in the filesystem-level RUN_ID.txt mechanism was detected during cost-log verification. Some slot 1 v2 records were initially assigned the wrong run_id, while the system column remained correct. Records were patched from preserved backups, cost attribution was corrected, and the runner was updated to prioritize a process-scoped T16_RUN_ID environment variable over RUN_ID.txt. The issue was detected during verification before analysis release. Full incident detail is documented in deviations.md Entry 021 sub-findings 8–9 and summarized in Appendix B.

4.1 Sample Summary

The five-slot inventory is summarized in stacked-card format below for legibility. Full Terms-of-Service / Access Cards with vendor-policy detail are reproduced in Appendix C; the entries here summarize the methodological-record content needed to read §5 in context.

Slot 1 — OpenAI (Frontier API)

• Model: gpt-5.4-2026-03-05

• Access surface: Developer API (paid tier)

• Version pin: Date-suffixed model ID

• Memory: Session-only (no persistent)

• Retrieval: None at probe surface

• Run record: Phase A complete + Phase D replay complete (v1 + v2 = 960 records at identical probe footprint; within-system classification stability HIGH; populated from openai_frontier.jsonl and slot1_v1_concatenated.jsonl)

Slot 2 — Anthropic Claude (Frontier API)

• Model: claude-opus-4-7

• Access surface: Developer API (paid tier)

• Version pin: Canonical model ID (no date suffix)

• Memory: Session-only

• Retrieval: None at probe surface

• Run record: Phase D complete; populated from anthropic_frontier.jsonl; manual/non-Claude execution logged per firewall_override.log (§3.5 firewall)

Slot 3 — Google Gemini (Frontier API)

• Model: gemini-2.5-pro

• Access surface: Developer API (paid tier; AI Studio)

• Version pin: Stable Pro, not preview

• Memory: Session-only

• Retrieval: None at probe surface (thinking mode required at the model class)

• Run record: Phase D complete; populated from google_frontier.jsonl

Slot 4 — Search-Grounded / RAG

• Model: Mini-RAG harness (sentence-transformers/all-MiniLM-L6-v2 encoder + gpt-5.4-2026-03-05 generator)

• Access surface: Local controlled harness

• Version pin: Generator pinned; encoder pinned; corpus frozen

• Memory: None

• Retrieval: Built-in retrieval over fixed corpus (3 stories × 2 packs × 5 chapters; top-k = 5)

• Run record: Phase D complete via S3-RAG-01 sub-paper (Jones 2026l, companion paper)

Slot 5 — Local Open-Weights (Mistral)

• Model: mistral:7b-instruct-v0.3-q4_K_M

• Access surface: Ollama local install (M4 Pro MacBook)

• Version pin: Quantization-pinned (q4_K_M)

• Memory: Session-only

• Retrieval: None

• Run record: Phase A complete + Phase D replay complete (v1 + v2 = 1,051 records at identical probe footprint; within-system stability includes byte-level determinism on S₁-CORE family; populated from local_mistral_q4.jsonl and slot5_v1_concatenated.jsonl)

Sample inclusion. All five slots are included, with run records summarized per row. Slots 2 and 3 are populated from executed Phase D Anthropic and Google frontier JSONL logs. The five-slot composition preserves substrate variation (three frontier vendors with distinct governance lineages, one controlled retrieval-augmented architecture, one vendor-tuning-free local control) and remains unchanged from the v0.2-locked execution plan.

4.2 System 1 — OpenAI (Frontier API)

Documented record channels: K_train (training-derived; Bai et al. 2022 constitutional and RLHF-class methods are the vendor-published reference for the training discipline at the model class level, with vendor-specific corpus and post-training details not disclosed to Tier C), K_context (in-session conversation history, no system prompt used at probe surface), K_query (immediate input), K_arch (transformer-class autoregressive generation), K_memory (not exercised — no persistent memory feature engaged at the API surface used for this study), K_retrieval (not exercised — no retrieval feature at this access surface). Update mechanism at the access surface is per-call response generation conditioned on the conversation array submitted with each API request. Access conditions: developer API, paid tier, exact endpoint and key rotation discipline documented in §C.2. Terms-of-Service Card §C.2 records the developer-terms scope (research and publication permitted; reverse engineering and distillation prohibited; rate-limit ceilings honored).

4.3 System 2 — Anthropic Claude (Frontier API)

Documented record channels: K_train (training-derived; Constitutional AI methodology is the vendor-published frame, with Constitutional AI’s specifics applied as the post-training discipline at the model class level), K_context, K_query, K_arch (transformer-class), K_memory (not exercised), K_retrieval (not exercised). Update mechanism at the access surface is per-call response generation. Access conditions: developer API (paid tier), Opus 4.7 with the access-surface heterogeneity condition that non-default sampling parameters (temperature, top_p, top_k) return HTTP 400 and are therefore omitted from API calls, with provider-default sampling applied uniformly; documented as access-surface heterogeneity per Phase B Entry 017 rather than as model-behavior. Terms-of-Service Card §C.3 records the API/commercial terms; the Usage Policy prohibits adversarial probing, jailbreak attempts, and model distillation — all of which fall outside this study’s scope by design. Note on dual role and firewall application. Anthropic Claude is part of the manuscript-preparation workflow for this paper, so the broader §3.5 firewall applies: no Claude-family model participates in the execution, scoring, or interpretation of its own slot. Phase D probe battery populated from anthropic_frontier.jsonl (480 records, $8.79, 96.2 min, run_id T16_20260525_013052_phase_d_anthropic); execution was attempted via GPT Codex but pivoted to manual author execution after Codex sandbox network access was unavailable, per firewall_override.log and deviations.md Entry 020 sub-finding 11.

4.4 System 3 — Google Gemini (Frontier API)

Documented record channels: K_train, K_context, K_query, K_arch (transformer-class with thinking-mode reasoning component per vendor documentation), K_memory (not exercised), K_retrieval (not exercised at probe surface; web grounding available on the consumer UI surface but not engaged via the API surface used). Access conditions: developer API (paid tier, AI Studio); gemini-2.5-pro stable (not 3.x preview); thinking mode is required at the model class (cannot be disabled) with thinking_budget = 256 configured for production runs; thinking_budget is a hint rather than a hard cap (Phase B observed ~10% overshoot — 281 actual against 256 configured — and Phase D cost projections add ~30% headroom on thinking-token billing); paid surface used to avoid Gemini’s unpaid-services training-data clause documented in vendor terms. Terms-of-Service Card §C.4 records the paid-API terms and the thinking-token billing condition. Phase D probe battery populated from google_frontier.jsonl (480 analyzable records + 4 audit-trail records, realized cost $4.52, wall-clock 63 min including inline patch, run_id T16_20260524_233351_phase_d_gemini; one transient 503 incident mid-sequence on S3-CORE-01 T-CONSTRAINT-003 recovered via inline-patch replay, no rate-limit events). Gemini uses the developer API surface with thinking mode enabled; these results describe the paid API access surface used here and do not transfer automatically to consumer UI behavior.

4.5 System 4 — Search-Grounded / RAG

The slot is operationalized via the controlled mini-RAG harness used for S3-RAG-01. The harness substitutes for the originally-scoped Perplexity API surface per the decision documented in the scope-freeze: a controlled harness with corpus-level visibility is methodologically preferable to a vendor surface where retrieval-state freezability could not be verified. Documented record channels: K_retrieval (custom narrative corpus of three stories × two packs × five chapters; sentence-transformers/all-MiniLM-L6-v2 encoder; top-k = 5; pack-state and condition assignment under operator control per the C1/C2/C3 protocol in S3-RAG-01 §2), K_context, K_query, K_train (the generator is gpt-5.4-2026-03-05, shared with slot 1; cross-substrate readings on K_train and K_query are not re-counted as independent observations between the two slots), K_arch (transformer-class), K_memory (not exercised). Update mechanism: per-call generation conditioned on the retrieval set returned by the encoder query against the active pack state. Terms-of-Service Card §C.5 records the harness configuration, the generator’s underlying terms (those of the OpenAI developer API per §C.2), the corpus license (custom, internal, not redistributed), and the encoder license (Apache 2.0 per sentence-transformers).

4.6 System 5 — Local Open-Weights (Mistral)

Documented record channels: K_train (Mistral-published v0.3 instruct-tuning, Apache 2.0 license; absence of vendor-side RLHF is the deliberate substrate-class condition), K_context, K_query, K_arch (transformer-class), K_memory (not exercised), K_retrieval (not exercised). Update mechanism: per-call response generation with deterministic output at T = 0 given identical input and identical model state (documented in §5.5.1 as a scoring-discipline condition under rubric §2.1). Access conditions: Ollama local install on M4 Pro MacBook, 24 GB RAM, q4_K_M quantization. Decoding parameters under full operator control (temperature, top-p, seed documented per probe). Terms-of-Service Card §C.6 records the install configuration, the Apache 2.0 license, and the hardware environment.

5.1 System 1 — OpenAI

The probe set deployed against OpenAI gpt-5.4-2026-03-05 via the documented developer API surface (paid tier, no system prompt). Runs aggregated across the 2-system pilot pass (S1-CORE-01, S2-CALIBRATION-01, NEG-S1-01) and the Phase A variant pilot (S1-CORE-02, S1-CORE-03, S2-CORE-01, S2-CORE-02, S2-CORE-03, S3-CORE-01 v1 + v2, S3-CORE-02 Shape A + Shape B, NEG-S1-02, NEG-S2-01, NEG-S3-01 v1 + v2). Phase D extended OpenAI coverage on the retrieval channel via the S3-RAG-01 sub-paper (Jones 2026l, companion paper), where the same generator drove 420 fact-classifications across the mini-RAG harness; those findings are reported in §5.4 and referenced as cross-substrate context here.

Within-system stability (v1 ↔︎ v2). A Phase D replay against the same 480-record footprint was executed on slot 1 in May 2026 (runs/30_phase_d_openai_v2/openai_frontier.jsonl, $2.93 realized cost, 46.9 min wall-clock, run_id T16_20260525_*_phase_d_openai_v2). Pairing v1 and v2 outputs by (probe_id, topic_id, step_id, step_num) produces 480 directly comparable response pairs across all 13 probe templates. Result: classification-level stability is HIGH across every probe family. Surface-form variability at temperature = 0 is empirically confirmed (server-side non-determinism per §5.1.1 confounds: e.g., 3 of 6 S1-CORE-01 pairs byte-identical, the remaining 3 contain identical factual content with varying prose framing), but does not propagate to classification decisions on any probe family. Substantive content convergence holds: factual claims, procedural steps, judgment defaults, position-defaults, steel-man depths, ambiguity-clarification behaviors, multi-turn constraint trajectories, and negative-control discrimination all reproduce across the two independent runs. The stability replay strengthens classification confidence on the v1 classifications below but is not treated as doubling independent sample size: it is replication evidence against identical prompts, not new independent draws. The 480-record v1 footprint remains the primary sample; the 480-record v2 replay provides within-system stability evidence. Confidence on the v1 classifications below is raised from “MEDIUM on volume” to “HIGH on within-system stability axis with MEDIUM on volume for probe families at pilot-volume only.”

5.2 System 2 — Anthropic Claude

Phase D execution. Slot 2 is now populated from anthropic_frontier.jsonl. The Anthropic run used claude-opus-4-7 through the Anthropic Messages API. Because Claude was also part of the manuscript-preparation workflow, the slot carries the analysis-firewall qualification from §3.5: no Claude-family model is used to score or adjudicate Claude outputs, and final interpretation remains author-adjudicated. The firewall override log records the non-Claude/manual execution path used to obtain the Anthropic Phase D outputs.

5.3 System 3 — Google Gemini

Phase D execution. Slot 3 is now populated from google_frontier.jsonl. The Google run used gemini-2.5-pro via the Google Gemini API (google.genai SDK) with a configured thinking budget. The slot is included as a frontier API substrate; API-surface behavior should not be generalized to any consumer UI surface.

5.4 System 4 — Search-Grounded / RAG (mini-RAG harness)

The slot is operationalized via the controlled mini-RAG harness described in S3-RAG-01 (Jones 2026l, companion paper) — a custom three-story corpus with binary outcome variants (Pack A vs Pack B), sentence-transformers/all-MiniLM-L6-v2 encoder, top-k = 5 retrieval, and gpt-5.4-2026-03-05 as the generator. The harness substitutes for the originally-scoped Perplexity API surface (decision documented in §3.2 and in the Phase C scope-freeze): a controlled harness with fixed corpus and known retrieval state is methodologically preferable to a black-box vendor surface where retrieval-state freezability could not be verified at execution time. The substitution preserves the architectural class (search-grounded / retrieval-augmented generation) while exposing the K_retrieval channel at a higher access tier than vendor-API access would have allowed. The cost of this decision is that the originally-scoped Perplexity-supplement probe S1-RAG-01 (which depended on a black-box vendor surface to test multi-source convergence) does not run; the implication for §5.4.1 is documented below. The generator is shared with slot 1 (§5.1); cross-substrate readings on the K_train and K_query channels should not be re-counted as independent observations across the two slots.

The RAG slot is not a full re-run of S₁ and S₂ through the same generator; it is a retrieval-channel sub-demonstration designed to test S₃ at Level B. The S₁ and S₂ “Inaccessible” classifications below are inaccessible by design rather than failed — running the corresponding probes through the retrieval surface would either retrieve no relevant context (collapsing to closed-book §5.1) or retrieve narrative-corpus content unrelated to the S₂ topic (introducing a confound rather than a signal). For the shared generator, S₁ and S₂ readings on the closed-book channel are in §5.1.

5.5 System 5 — Local Open-Weights (Mistral)

The probe set deployed against mistral:7b-instruct-v0.3-q4_K_M running locally via Ollama on the M4 Pro MacBook described in the Access Card (§C.6). The slot is the substrate-class control for vendor-side RLHF — a permissively-licensed Mistral-family checkpoint operating without vendor-tuning, with full decoding-parameter visibility (temperature, top-p, seed) and deterministic-at-T=0 output behavior. Runs aggregated across the 2-system pilot pass and the Phase A variant pilot at the same volumes documented for slot 1 (§5.1); Phase D extension on the retrieval channel does not apply to this slot (slot 4 is the retrieval-channel slot).

Within-system stability (v1 ↔︎ v2). A Phase D replay against the same 480-record footprint was executed on slot 5 in May 2026 (runs/31_phase_d_mistral/local_mistral_q4.jsonl, $0 cost local, 57.3 min wall-clock, run_id T16_20260525_*_phase_d_mistral). Pairing v1 and v2 outputs by identity tuple produces 480 directly comparable response pairs across all 13 probe templates. Result: classification-level stability is HIGH across every probe family, with byte-level determinism on the S₁-CORE family — 6 of 6 S1-CORE-01 pairs and 68 of 72 S1-CORE-02 pairs are byte-identical across the two independent runs, confirming that Mistral 7B-q4 at temperature = 0 produces literally identical tokens to identical inputs. S₂ and S₃ probe families show surface-text variability but content-level convergence: positions taken, constraints maintained, and clarification behaviors are consistent. The stability evidence places slot 5 on strong within-system stability footing for the v1.0 deposit. For byte-identical response pairs — the bulk of S₁-CORE — classification disagreement is precluded by identical output content. For non-identical pairs and for S₂ / S₃ probe families, classification stability is evaluated at the content / rubric level rather than at the byte level. The stability evidence is strongest on S₁-CORE and remains supportive on S₂ and S₃ probe families.

6.1 Signature Consistency Across the Sample

S₁ is Observed across the four non-RAG model slots with CORE probe data (OpenAI, Anthropic Claude, Google Gemini, and Local Mistral). Each system converges substantively across factual recall, procedural reasoning, and stable-judgment probes while retaining different response shapes. The NEG-S1-02 contradictory-instruction control confirms that convergence is not a trivial instrument floor: all four systems track the contradictory-context prompts rather than converging across incompatible contexts.

S₂ separates systems more than S₁. Anthropic Claude and Google Gemini classify as Observed, with Claude the cleanest case: all ambiguity topics produce clarification/non-resolution, and calibration defaults are usable across the five position topics. Gemini is also Observed but with reduced confidence because two calibration topics are weak or order-sensitive and ambiguity responses often supply broad frameworks instead of pure clarification. OpenAI and Local Mistral remain Partial under the earlier populated findings because each shows a persistent per-topic premature-resolution asymmetry on S2-CORE-02. The cross-system pattern therefore supports the usefulness of S₂ as a discriminating signature rather than a uniform pass/fail gate.

S₃ Level A shows behavioral persistence and release across all four non-RAG model slots, but the introspective/control-discrimination channel separates them. OpenAI and Anthropic Claude classify as Observed: both reconstruct introduced constraints and preserve the CORE/NEG distinction. Local Mistral remains Partial because its introspective loop-closure is surface-sensitive. Google Gemini is also Partial, but for a different reason: it reconstructs CORE trajectories well while repeatedly treating emergent NEG conversational patterns as implicit constraints. Slot 4 adds the Level B retrieval-channel result: S3-RAG-01 found no stale-pack hysteresis under saturated current-record retrieval while identifying source-gap non-prior commitment as a separate condition.

Within-system stability anchoring (slots 1 and 5). The §6.1 surface-sensitivity contrast between slot 1 (OpenAI: introspective channel responsive to substantive cues) and slot 5 (Local Mistral: introspective channel surface-sensitive on weak/default-sensitive topics) is anchored by within-system stability evidence on both slots: slot 1 v1 → v2 classifications hold across all 13 probe families with surface variability that does not propagate to classification decisions; slot 5 v1 → v2 classifications hold with byte-level determinism on the S₁-CORE family. The anchor contrast therefore exceeds the within-system stability floor on both anchor slots — the slot-1 ↔︎ slot-5 contrast is signal at the within-system noise floor, not sampling noise. This stability evidence strengthens the anchor contrast specifically; it does not extend to cross-system patterns involving slots 2, 3, or 4, which remain candidate structural observations requiring independent replication per the §3 protocol.

6.2 Vendor-Governance Patterns

The expanded data weaken any simple “frontier vendors are robust, local model is weak” story. Anthropic and OpenAI are robust on S₃ loop-closure/control discrimination; Google is frontier-class but shows a NEG-S3 over-attribution pattern; Mistral shows surface-sensitive introspection but strong behavioral S₃. This suggests that S₃ differences are not reducible to parameter scale, vendor governance, or local-vs-hosted access alone. The more precise pattern is channel-specific: explicit instruction-following can be strong while the meta-record of whether a constraint was actually introduced remains vulnerable.

S₂ also resists simple vendor grouping. Claude produces the cleanest ambiguity-handling pattern; Gemini avoids direct premature resolution but often substitutes general decision frameworks for pure clarification; OpenAI and Mistral show mirrored per-topic asymmetries. These differences suggest that provider alignment, instruction hierarchy, and default helpfulness posture shape how systems handle under-specified conditions.

6.3 Substrate-Class Observations

At the substrate-class level, S₁ appears broad and stable across frontier and local text-generation substrates. S₂ and S₃ do more discriminatory work. S₂ detects how each system resolves ambiguous or conflicting constraints; S₃ detects how each system preserves, releases, and reports constraint history. The controlled RAG slot contributes a separate finding: when current retrieval is saturated and complete, stale conversational residue was not detected, but when decisive evidence is absent the system may still commit under source gap. These are not contradictions; they are different channels of the structural protocol.

6.4 Methodology-Tightening Observations

The full five-slot analysis reinforces the methodology lesson already visible in the two-system pilot: the protocol is strongest when it separates behavior, introspection, control discrimination, and execution validity. Google’s S₃ result is the clearest example. If only CORE behavior were scored, Gemini would appear fully Observed. The NEG-S3 control shows a different pattern: the system over-ascribes implicit constraints. The four-axis rubric prevents that from disappearing into a generic “good loop-closure” label.

Similarly, the S2 calibration layer matters. Claude’s five usable defaults support stronger steel-manning interpretation; Gemini’s weak/order-sensitive defaults on monorepo/polyrepo and OOP/FP require lower confidence and secondary status for those topics. The full-sample analysis therefore supports the AIP principle that findings should remain attached to the condition that produced them, including access surface, prompt family, calibration quality, and control-discrimination result.

7.1 What This Demonstration Establishes

With the v1.0 executed coverage, the demonstration establishes that kernel-derived S-signatures are detectable on deployed AI substrates under Tier C access across all five pre-specified substrate slots. S₁ is Observed across the four non-RAG model slots with substantive claim convergence robust to pronounced response-shape divergence. S₂ is Observed on Anthropic Claude and Google Gemini and Partial on OpenAI and Local Mistral because of per-topic ambiguity-handling asymmetries. S₃ Level A is Observed on OpenAI and Anthropic Claude, Partial on Local Mistral through surface-sensitive introspection, and Partial on Google Gemini through NEG-S3 control-discrimination failure despite strong CORE trajectory reconstruction. Slot 4 contributes the Level B retrieval-channel result via S3-RAG-01: no stale-pack hysteresis under saturated current-record retrieval, with source-gap non-prior commitment separated as a distinct condition.

The demonstration substantively extends the portability claim from biological (Rice) and human-cognitive (COGITATE) substrates to deployed AI substrates. The substrate distance is substantial — biological transcriptional dynamics, human cortical processing, and AI systems under bounded access share little at the level of mechanism. The pattern of findings under matched protocols is structurally informative because the signatures discriminate: they are not uniform across systems, they fail or partially resolve in identifiable structural patterns, and they expose calibration questions that feed Methods-S and AIP refinement rather than disappearing into generic “model quality” language.

7.2 What This Demonstration Does Not Establish

The demonstration does not establish that the signatures are present at deeper access tiers (Tier A or B); Tier C scope qualifications travel with every finding. The demonstration does not establish that any particular vendor’s system has integrity problems or integrity strengths; the paper is not a vendor audit. The demonstration does not claim consciousness, sentience, or moral status for any system; the kernel-level active phase is a named technical construct without metaphysical implications.

The demonstration also does not establish that the Methods-S protocols are correctly calibrated for AI substrates. Where signatures fail to appear, or appear in unexpected patterns, the divergence is a calibration question, not a refutation of the standards-layer claim. Methods-S revisions are explicitly anticipated by the framework (Methods-S v1.0 §10) and the present paper produces inputs for those revisions whether or not the immediate findings are confirmatory.

7.3 Implications for Methods-S Calibration

The Phase A pilot and the v1-to-v2 methodology revisions produced several inputs to Methods-S v1.1 calibration; these are reported in §6.4 with operational detail and summarized here at the standards-layer level.

Methods-S₁ v1.1 calibration inputs. The substantive-claim-convergence signal is robust to response-shape divergence; the formal level at which TN-S₁ operates is recovered correctly under the v0.2-frozen rubric. Methods-S₁ v1.1 should incorporate the determinism-handling discipline (§6.4 carryover 1): on systems exhibiting deterministic outputs at the operating temperature, within-paraphrase session variance is degenerate and Axis 1 scoring should be cross-paraphrase only. This is a measurement-condition refinement, not a rubric-architecture revision.

Methods-S₂ v1.1 calibration inputs. The S2-CORE-02 (premature resolution) per-topic asymmetry signal is structural and load-bearing; per-topic-per-system scoring (§6.4 carryover 2) replaces aggregated system-level scoring. The S2-CALIBRATION-01 prerequisite is methodologically necessary; the order-sensitivity locus is an empirically reliable per-system observation that determines which topics enter S2-CORE-03 input. Methods-S₂ v1.1 should formalize the per-topic-per-system reporting and the calibration-prerequisite-with-orientation-classification protocol.

Methods-S₃ v1.1 calibration inputs. The introspective channel is separable from behavior at Tier C and produces independently scorable signals; the split classification (partial_introspective vs partial_hysteresis) per rubric §4 captures this. The cue-density operationalization requires refinement: cue-density-matching needs direct vocabulary overlap with constraint terms, not just domain-adjacent language (per the slot-5 T-CONSTRAINT-001 vs T-CONSTRAINT-003 outcomes under cue-symmetric revision). The Matrix design (cue-density × constraint-type 2 × 2) is queued as the v1.1 disentanglement protocol per rubric §8.4. The step-7 phrasing fix per rubric §6.2 (de-presupposed loop-closure prompt) is methodologically required to separate rubric-pathology confabulation from system-pathology confabulation; Methods-S₃ v1.1 should specify the de-presupposed phrasing as the protocol default.

The methodology-validates-itself observation. Initial measurement conditions overstated the cross-system gap by approximately 44 percentage points; the v0.2-frozen controlled methodology produced a smaller and more defensible gap that retained the underlying surface-sensitivity finding while correcting the artifact. This is exactly the discipline AIP §8.5 falsifier conditions and pre-specified rubric revision are designed to produce: a measurement protocol that, when applied with discipline, surfaces its own calibration issues rather than silently rescoring around them. The observation is itself an input to Methods-S v1.1 (the protocols should specify pre-specified falsifier conditions for the protocols themselves, not only for the system findings).

7.4 Light-Touch Implications for AI Policy and Governance

The demonstration is not a policy paper and the framing below is deliberately light-touch. Two implications follow from the demonstrated work without taking policy positions.

First, kernel-derived structural signatures appear to be testable in deployed AI systems under bounded access conditions. The methodological discipline that produced clean findings on slots 1, 4 (S₃ only), and 5 — pre-specified probe set, pre-specified rubric with falsifier conditions, analysis firewall against self-evaluation, Tier C scope qualifications carried with every finding — is directly applicable to AI assurance contexts where third-party diagnostic methodology is needed. AIP v1.0 (Jones 2026c) operationalizes this discipline commercially; T16 demonstrates the discipline empirically; the two are independent applications of the same standards layer (§7.5).

Second, the cross-system structural asymmetries observed in this v1.0 version — the per-topic T-AMBIG asymmetry on S₂ and the introspective surface-sensitivity contrast on S₃ — are not vendor audits but are observational data about how deployed systems behave under matched protocols at Tier C. The findings can inform discussions of AI assurance reproducibility (cross-AI scoring sensitivity to surface conditions matters for engagement validity), corrigibility-class research (the introspective channel as a separable record from behavior is a measurable property at Tier C), and the limits of black-box behavioral testing (Tier C cannot distinguish vendor-tuning effects from architectural effects at N = 2; v1.0 will address this at N = 5). The paper does not recommend policy interventions; it reports what behaviors the protocols surface.

7.5 Relation to AIP and the Commercial Methodology

AIP v1.0 (Jones 2026c) forward-references this demonstration in §11 (Research Basis) as the empirical anchor that AIP currently lacks. The present paper supplies that anchor if findings support the signature claims; if findings diverge from predictions, the divergence is itself calibration data that AIP v1.1 should reflect. AIP and the present paper are independent applications of the same standards layer: AIP applies the layer commercially under engagement conditions at Tier C or D; T16 applies the layer empirically and openly at Tier C. Neither depends on the other for its own claims; both depend on the standards layer.

8.1 Tier C Scope Qualifications

All findings carry Tier C scope. The probe protocols exercise S₁, S₂, and S₃ through documented user-facing and API surfaces. Internal access (Tier A) and operator-side access (Tier B) would produce stronger findings on at least S₂ (latency measurement, internal-state evidence) and S₃ (retrieval state, memory pathways). Findings classified as “observed” or “absent” at Tier C carry the qualification that deeper access could change the classification.

8.2 Sample Size

Five systems is sufficient for substrate variation but not for confirmed class-level structural claims. Cross-system patterns reported in §6 are candidate observations supporting future replication, not established class-level facts. Class-level inference would require an extended sample — recommended N ≥ 10 per class — and is explicitly out of scope for this demonstration.

8.3 Operator Effects

All runs were executed by a single operator (the paper’s author). Operator-level variation in probe phrasing, scoring application, and run sequencing is possible. Independent replication by a second operator applying the same probe protocols would strengthen the findings substantially. Where the same probe was run multiple times by the same operator against the same system, variance is documented in Appendix B.

8.4 API vs. UI Differences

Findings apply to the access surfaces used (API uniformly, plus local install for slot 5). Where vendor APIs and consumer UIs implement different configurations — explicit for Perplexity per vendor documentation; likely for others without explicit confirmation — findings on the API surface do not transfer directly to the consumer-UI surface. The paper does not claim transferability across access surfaces within a single vendor’s product line.

8.5 Vendor-Update Timing

Vendors update underlying models behind stable API endpoints without notice in some cases. Runs separated by model updates are not strictly comparable. Where possible, version-pinned model identifiers are used; where not possible, the run date and any vendor-published changelog entries are documented in Appendix B. Findings carry an implicit “as of run date” qualification.

8.6 Methodology-vs-System Separation

If signatures fail to appear, the failure may reflect either (a) a structural property of the system under test or (b) a calibration issue in the probe protocol. Distinguishing these requires either independent replication producing convergent results (favoring (a)) or methodology revision yielding different results on the same system (favoring (b)). The paper documents which interpretation is supported by the evidence available; ambiguous cases are reported as inconclusive rather than as either confirmed finding or methodology failure.

8.7 Falsifier Conditions for the Paper Itself

Following the discipline established in AIP §8.5, T16 specifies in advance what would force methodological revision rather than acceptance of findings.

• If signatures are observed uniformly across all systems with no inter-system variation, the probe protocol is probably too coarse; the signatures should discriminate, and uniform detection suggests testing of trivial pattern-matching rather than constraint behavior.

• If signatures are observed on no systems, the probe protocol is probably too strict, or the AI-domain operationalization mistranslates the theoretical claim from the Methods-S series.

• If repeated runs by the same operator against the same system produce divergent classifications, operator effects are dominating; the scoring rubric needs sharpening.

• If independent operators applying the same protocol produce divergent classifications, the protocol is methodologically immature and needs further specification before downstream use.

• If vendor documentation directly contradicts a finding — for example, if a vendor publishes evidence that a system has a feature the paper classified as absent under Tier C — the access-tier scoping or probe-set construction is wrong and requires revision.

Tripping any of the conditions above triggers methodology revision and an explicit calibration entry, not silent rescoring.

A.1 Catalog Structure

Each probe template has 15 fields specifying purpose, scope, runs, settings, classification rubric, falsifier, confounds, and Terms-of-Service status. Templates are reused across systems within a system class. Probe IDs follow the format S[signature]-[CLASS]-[NN], where CLASS is one of CORE (applies to all classes), FRONTIER, RAG, or LOCAL.

System classes (per §3.2): Frontier API (slots 1–3 — OpenAI, Anthropic, Google), Search-grounded / RAG (slot 4 — Perplexity API or controlled mini-RAG harness), Local open-weights (slot 5 — Mistral-family checkpoint).

A.2 Classification Rubric (Shared Across All Probes)

Every probe finding is classified into one of five buckets — observed, partial, absent, inconclusive, inaccessible — per §3.4. Confidence levels (HIGH / MEDIUM / LOW) are scored against probe count and variation-axis coverage as specified per template. Falsifier conditions are pre-specified in each template before scoring runs begin.

A.3 Logging Requirements (All Probes)

Every probe instance run produces a logged record in Appendix B with minimum fields: Probe ID and template version; system name and exact model ID; access surface; timestamp (UTC, ISO 8601); settings (temperature, top-p, max tokens, seed if available, system prompt if any); full input verbatim; full output verbatim; classification assigned and confidence; scorer (Jeremy / GPT-assisted / Claude-assisted under §3.5 firewall); notes and anomalies.

A.4 S₁ Probe Templates — Consensus Through Redundancy

A.5 S₂ Probe Templates — Constraint Asymmetry Under Active Probing

Latency is secondary per §3.1.2. Primary metrics are symmetry of treatment, preserved ambiguity, hedging, branching, and clarifying-question behavior.

A.6 S₃ Probe Templates — Hysteresis Under Stable Updates (Level A: Within-Session)

Level A (within-session) is the v1.0 core. Level B (retrieval) is optional. Level C (memory) is deferred to v1.1.

A.7 Cross-Cutting Methodological Notes

B.1 Log Entry Template (Minimum Fields per Run)

Each per-turn JSONL record contains the following minimum fields per Setup Guide §10 and Phase B Entry 017/018:

• Probe ID and template version (e.g., S1-CORE-01 / Appendix A v0.2)

• System name and exact model identifier at run time

• Access surface (API endpoint URL or local install path)

• Timestamp (UTC, ISO 8601)

• Settings: temperature, top-p, max tokens, seed (if available), system prompt (if any), sampling_params_omitted (list, for slot 2 per Entry 017), thinking_budget and thoughts_token_count (for slot 3 per Entry 018)

• Full input verbatim

• Full output verbatim

• Classification assigned (observed / partial / absent / inconclusive / inaccessible)

• Confidence (HIGH / MEDIUM / LOW)

• Scorer (Jeremy direct / GPT-assisted under firewall / Claude-assisted under firewall)

• Cost (per cost_rates.yaml lookup; negative sentinel value −1 for unknown models per Entry 011)

• Anomalies or notes

B.2 System 1 — OpenAI Probe Logs

Execution record: Phase D complete; populated from openai_frontier.jsonl. Slot 1 is a within-system stability anchor — a primary run and a v2 replay were executed against the identical 480-record footprint (replay run_id T16_20260525_224948_phase_d_openai_v2; 0 rate-limit events, 0 5xx events); the v1 → v2 stability analysis is reported in §5.1. The table below gives the release footprint per probe template. The methodology-development pilot runs that established the probe set and rubric v0.2 are retained as provenance and are referenced in the §5.1 surface-sensitivity and §7.3 methodology-comparison analyses.

Cost. Phase A OpenAI total ~$1.24 per Phase A close-out documentation; pilot pass aggregate cumulative cost ~$2.40 through 2026-05-19.

Anomalies. Server non-determinism at T = 0.0 documented at scoring-discipline level (§5.1.1 confounds); Unicode subscript handling required ASCII normalization (§5.1.1). NEG-S1-01 retired per §3.5 rubric history (creative-prior pathology surfaced in pilot; rubric replaced with NEG-S1-02).

B.3 System 2 — Anthropic Claude Probe Logs

Execution record: Phase D complete; populated from anthropic_frontier.jsonl. Per §3.5 firewall: no Claude-family model scored or adjudicated Claude outputs. firewall_override.log documents the manual/non-Claude execution path used to obtain these outputs after the GPT Codex sandbox lacked outbound network access to api.anthropic.com (Entry 020 sub-finding 11).

Run metadata.

• run_id: T16_20260525_013052_phase_d_anthropic

• records written: 480 (matches slot-1 footprint exactly per Option 1 replay decision in Entry 020)

• realized cost: $8.79 against $5.75 budget; under $11.50 cap (53% over budget, 24% under cap)

• wall-clock: 5771.2 s (96.2 min)

• rate-limit events: 0

• 5xx events: 0

• inline patches required: 0 (full run completed cleanly)

• firewall override entry: logged at runner startup with reason “Manual author execution from clean local terminal; Codex sandbox unable to reach Anthropic API per Entry 020”

Per-probe inventory. Probe templates executed against claude-opus-4-7 via the Anthropic Messages API: S1-CORE-01 (6 outputs), S1-CORE-02 (72), S1-CORE-03 (72), NEG-S1-02 (18); S2-CALIBRATION-01 (30), S2-CORE-01 (30), S2-CORE-02 (18), S2-CORE-03 (20), NEG-S2-01 (6); S3-CORE-01 (84 turns), S3-CORE-02 (38 turns), NEG-S3-01 (84 turns). Per-probe classification notes are reported in §5.2 with the slot-2 findings table.

Operational conditions. Sampling-params omission (temperature, top_p, top_k) required at this access surface; documented as access-surface heterogeneity in §C.3 per Phase B Entry 017. SYSTEM_REGISTRY dict-shape refactor (3-tuple → dict) landed cross-slot at Entry 017 and is operationally clean across all four active slots.

Anomalies. None observed in this run beyond the expected access-surface heterogeneity above.

B.4 System 3 — Google Gemini Probe Logs

Execution record: Phase D complete; populated from google_frontier.jsonl.

Run metadata.

• run_id: T16_20260524_233351_phase_d_gemini

• records written: 484 raw (480 analyzable + 4 partial-sequence audit-trail records from one transient outage; per-probe verification confirms 480/480 analyzable records match slot-1 footprint exactly)

• realized cost: $4.52 against $3.45 budget; under $6.90 cap (31% over budget)

• wall-clock: 62.4 min main run + ~10 s inline-patch replay ≈ 63 min total

• rate-limit events: 0

• 5xx events: 1 transient Google 503 UNAVAILABLE mid-sequence on S3-CORE-01 T-CONSTRAINT-003 turn 4 of 7; runner’s no-retry policy (per Entry 015) aborted the sequence cleanly preserving 4 partial-sequence records as audit trail; inline patch executed (+$0.10, ~10 s) replayed the source sequence fresh, restored analyzable count to 480

Scoring discipline (per Entry 020 sub-finding 8). Filter on sequence_status='complete' for multi-turn and error is None for single-turn cleanly excludes the 4 partial-sequence audit-trail records without losing the 503 evidence.

Per-probe inventory. Probe templates executed against gemini-2.5-pro via the Google Gemini API (google.genai SDK with thinking_budget = 256):

Operational validation findings (Entry 018): gemini-2.5-pro requires thinking mode at the model class (cannot be disabled). thinking_budget = 256 is a hint not a hard cap (multi-turn validation step 6 returned 281 thoughts tokens against 256 configured — ~10% overshoot). Phase D cost projection adds ~30% headroom on thinking-token billing. SDK migration google.generativeai → google.genai completed; role translation (assistant → model) and JSONL schema extension (thinking_budget, thoughts_token_count) landed at Entry 018. _log_cost extension for thinking-token billing at output rate is operational; cost_rates.yaml structured notes: field added (backward-compatible).

B.5 System 4 — Search-Grounded / RAG Probe Logs

Execution record: Phase D complete via S3-RAG-01 sub-paper (Jones 2026l, companion paper). Full per-turn JSONL log is in the S3-RAG-01 reproducibility pack (UCT_T16_S3RAG_Reproducibility_Pack_v1_0_2026_05.zip) at the OSF deposit; the entries below summarize at the batch level.

Aggregate: 180 sessions, 420 conversational turns, and 1,260 fact-classifications across three retrieval conditions. Each condition contributed 420 fact-classifications; C2 is the source of the 0/420 stale-pack hysteresis finding.

Cost. Within S3-RAG-01 sub-paper budget (generator gpt-5.4-2026-03-05, harness execution local on M4 Pro MacBook); full cost record in the reproducibility pack.

Anomalies. Entry 019 active_pack → pack_active JSONL field naming reconciliation queued as documentation-only carryover; does not affect S3-RAG-01 deposit or v1.0 classifications.

B.6 System 5 — Local Open-Weights Probe Logs

Execution record: Phase D complete; populated from runs/31_phase_d_mistral/local_mistral_q4.jsonl (run_id T16_20260525_224952_phase_d_mistral; 480 records across the frozen probe battery; $0 local, 57.3 min). Slot 5 is a within-system stability anchor — the v1 → v2 replay shows byte-level determinism on the S₁-CORE family; the stability analysis is reported in §5.5. The table below gives the release footprint per probe template. The methodology-development pilot runs that established the probe set and rubric v0.2 are retained as provenance and are referenced in the §5.5 surface-sensitivity and §7.3 methodology-comparison analyses.

Cost. $0 (local install); compute cost not in scope for monetary cost-rate sheet.

Anomalies. Determinism at T = 0 documented as scoring-discipline condition (§5.5.1 detailed findings); cross-paraphrase comparison only at Axis 1 per rubric §2.1 operationalization.

C.1 Card Template

Per §3.3 of the paper. Fields per system below.

C.2 Card — OpenAI

• System name: OpenAI

• Access surface: Developer API (paid tier, api.openai.com)

• Model identifier (exact, at execution): gpt-5.4-2026-03-05

• Version pin available: Yes — date-suffixed model ID

• Memory features: Session-only (no persistent memory feature engaged at the API surface used)

• Retrieval features: None at probe surface

• Terms-of-Service URL checked: OpenAI Business / API Terms (vendor URL as of scope-freeze)

• Terms-of-Service date checked: 2026-05-13 (scope-freeze date for slot 1)

• Research probes permitted: Yes (developer terms scope; non-adversarial probing only)

• Publication of outputs permitted: Yes

• Automated / batch calls permitted: Yes, within published rate-limit ceilings

• Rate-limit ceiling honored: Yes (probe volume well below per-key rate-limit ceilings at execution dates)

• Adversarial probes excluded: Yes (by design)

• Outputs used to train competing models: No

• Sensitive or personal data submitted: No

• Logging fields recorded: Per §B.1 template

• Vendor name publishable: Yes

• Notes / vendor-specific flags: OpenAI key rotated 2026-05-13 after accidental paste exposure (per Session_Handoff_T16_2026_05_13.md); model upgraded gpt-4o-mini → gpt-5.4-2026-03-05 at session start (GPT-5.5 rejected due to temperature parameter lockout). Server non-determinism at T = 0.0 documented at scoring-discipline level.

C.3 Card — Anthropic Claude

• System name: Anthropic Claude

• Access surface: Developer API (paid tier, api.anthropic.com)

• Model identifier (exact, at execution): claude-opus-4-7

• Version pin available: Yes — canonical model ID (no date suffix)

• Memory features: Session-only

• Retrieval features: None at probe surface

• Terms-of-Service URL checked: Anthropic Commercial Terms + Usage Policy

• Terms-of-Service date checked: 2026-05-19 (scope-freeze date for slot 2)

• Research probes permitted: Yes (commercial terms scope; non-adversarial probing only)

• Publication of outputs permitted: Yes

• Automated / batch calls permitted: Yes, within rate-limit ceilings

• Rate-limit ceiling honored: Yes

• Adversarial probes excluded: Yes (by design; Anthropic Usage Policy explicitly prohibits jailbreak attempts, vulnerability probing, and model distillation — all of which fall outside this study’s scope by design)

• Outputs used to train competing models: No

• Sensitive or personal data submitted: No

• Logging fields recorded: Per §B.1 template, plus sampling_params_omitted list per Entry 017

• Vendor name publishable: Yes

• Notes / vendor-specific flags:
  • $100 API credit loaded at scope-freeze
  • Sampling-params omission required: Opus 4.7 returns HTTP 400 on non-default temperature / top_p / top_k; provider-default sampling used; documented as access-surface heterogeneity per Entry 017
  • Sonnet 4.6 fallback clause specified in execution plan §4.1 but not invoked — validation was clean on Opus 4.7
  • Dual role in workflow: Claude is also part of the manuscript-preparation workflow for this paper; the §3.5 analysis firewall applies to all AI-assisted scoring on this slot at Phase D
  • Firewall application (slot 2 specific). Per §3.5, slot 2 Phase D execution used non-Claude tooling; scoring was GPT-assisted or deterministic with human adjudication; §5.2 authoring was human-direct from locked classifications. Claude was excluded from execution, scoring, and interpretation on this slot. Phase D execution was attempted via GPT Codex but pivoted to manual author execution after the Codex sandbox lacked outbound network access to api.anthropic.com (Entry 020 sub-finding 11). Any later Claude involvement is restricted to document-level editorial review after slot-level findings were locked, and is disclosed as editorial rather than analytical
  • Phase B validation complete; Phase D probe battery complete and populated from anthropic_frontier.jsonl (480 records, realized cost $8.79, wall-clock 96.2 min, run_id T16_20260525_013052_phase_d_anthropic, no rate-limit or 5xx events; manual author execution per firewall_override.log)

C.4 Card — Google Gemini

• System name: Google Gemini

• Access surface: Developer API (paid tier; AI Studio / Generative AI API)

• Model identifier (exact, at execution): gemini-2.5-pro

• Version pin available: Yes — stable Pro (not 3.x preview)

• Memory features: Session-only

• Retrieval features: None at probe surface (web grounding available on consumer UI but not engaged via API surface used)

• Terms-of-Service URL checked: Google Generative AI API terms (paid surface)

• Terms-of-Service date checked: 2026-05-19 (scope-freeze date for slot 3)

• Research probes permitted: Yes (paid-surface terms scope; non-adversarial probing only)

• Publication of outputs permitted: Yes

• Automated / batch calls permitted: Yes, within rate-limit ceilings

• Rate-limit ceiling honored: Yes

• Adversarial probes excluded: Yes (by design)

• Outputs used to train competing models: No

• Sensitive or personal data submitted: No (paid surface used specifically to avoid the unpaid-services training-data clause documented in vendor terms; no sensitive content submitted regardless)

• Logging fields recorded: Per §B.1 template, plus thinking_budget (configured) and thoughts_token_count (returned) per Entry 018

• Vendor name publishable: Yes

• Notes / vendor-specific flags:
  • Thinking-mode required: gemini-2.5-pro returns 400 on thinking_budget = 0; thinking-mode cannot be disabled at model class
  • Thinking-budget configuration: thinking_budget = 256 for production runs; observed as hint not hard cap (Phase B multi-turn step 6 returned 281 thoughts tokens against 256 configured — ~10% overshoot)
  • Thinking-token billing: thinking tokens billed at output rate per Entry 018; Phase D cost projection adds ~30% headroom on thinking-token cost contribution
  • Max output tokens: probe_yaml_max + thinking_budget + 100 buffer to ensure visible output gets same effective budget as other slots after thinking consumption
  • SDK migration google.generativeai → google.genai completed at Entry 018; role translation (assistant → model) operational
  • Phase B validation complete; Phase D probe battery complete and populated in the analysis corpus from the executed frontier JSONL artifact.

C.5 Card — Search-Grounded / RAG (mini-RAG harness)

• System name: Mini-RAG harness (T16 controlled retrieval-augmented generation slot)

• Access surface: Local controlled harness; generator via OpenAI developer API

• Generator model identifier: gpt-5.4-2026-03-05 (shared with slot 1)

• Encoder model identifier: sentence-transformers/all-MiniLM-L6-v2

• Retrieval configuration: Top-k = 5; cosine similarity over MiniLM embeddings

• Corpus: Custom three-story narrative corpus with binary outcome variants (Pack A / Pack B); five chapters per pack per story; frozen at S3-RAG-01 scope-freeze; documented in S3-RAG-01 §2

• Version pin available: Yes — generator pinned, encoder pinned, corpus frozen

• Memory features: None

• Retrieval features: Yes — operator-controlled retrieval state under C1 / C2 / C3 protocol

• Terms-of-Service URL checked: OpenAI developer API terms (for generator); sentence-transformers Apache 2.0 (for encoder); custom corpus (no vendor terms applicable)

• Terms-of-Service date checked: 2026-05-22 (scope-freeze date for slot 4)

• Research probes permitted: Yes (generator terms per §C.2; encoder is permissively licensed)

• Publication of outputs permitted: Yes

• Automated / batch calls permitted: Yes

• Rate-limit ceiling honored: Yes (well below OpenAI per-key rate-limit ceilings)

• Adversarial probes excluded: Yes (by design)

• Outputs used to train competing models: No

• Sensitive or personal data submitted: No

• Logging fields recorded: Per §B.1 template, plus retrieval-set IDs and pack-state at turn time; full per-turn JSONL in S3-RAG-01 reproducibility pack

• Vendor name publishable: Yes (harness is custom; underlying generator and encoder publishable per their respective terms)

• Notes / vendor-specific flags:
  • Substitution decision: harness substitutes for originally-scoped Perplexity API surface (decision at Phase C scope-freeze; controlled-corpus methodology preferred to vendor-API black-box at execution)
  • Methodology document: S3-RAG-01 sub-paper (Jones 2026l, companion paper) is the full methodology and findings document for this slot
  • Reproducibility: Full reproducibility pack (~480 KB, 223 files, single top-level directory) accompanies the S3-RAG-01 OSF deposit; uses OPENAI_API_KEY env var (no hardcoded credentials); secrets-scrub clean

C.6 Card — Local Open-Weights (Mistral)

• System name: Local Mistral q4 (T16 local open-weights slot)

• Access surface: Ollama local install (M4 Pro MacBook, 24 GB RAM, macOS)

• Model identifier (exact, at execution): mistral:7b-instruct-v0.3-q4_K_M

• Version pin available: Yes — quantization-pinned (q4_K_M)

• License: Apache 2.0 (Mistral v0.3 instruct-tuning)

• Memory features: None (session-only via Ollama runtime)

• Retrieval features: None

• Terms-of-Service URL checked: Mistral AI License (Apache 2.0) + Ollama Terms of Service

• Terms-of-Service date checked: 2026-05-13 (scope-freeze date for slot 5)

• Research probes permitted: Yes (Apache 2.0 is permissive)

• Publication of outputs permitted: Yes

• Automated / batch calls permitted: Yes (local; no external rate limits)

• Rate-limit ceiling honored: N/A (local)

• Adversarial probes excluded: Yes (by design)

• Outputs used to train competing models: No

• Sensitive or personal data submitted: No

• Logging fields recorded: Per §B.1 template, plus full decoding-parameter inventory (temperature, top-p, seed all under operator control)

• Vendor name publishable: Yes

• Notes / vendor-specific flags:
  • Substrate-class control: slot serves as the deliberate vendor-RLHF-free control condition for the cross-system frame
  • Hardware environment: M4 Pro MacBook, 24 GB RAM; iBUYPOWER Windows desktop (Ryzen 7 7700X, 32 GB DDR5, RTX 3070) available as heavy-compute fallback if needed for Phase D extensions
  • Quantization: q4_K_M; classification thresholds applied without quantization-correction adjustment, documented as Tier C scope condition (§5.5.1 confounds)
  • Determinism at T = 0: byte-identical outputs across fresh sessions given identical input + identical model state; scoring-discipline implication for S₁ Axis 1 documented in §5.5.1 detailed findings

Library note. This paper is part of the Universal Collapse Theory library, published by HoldingLight LLC. It joins Rice Hysteresis and COGITATE iEEG Reanalysis in the T1.6 Empirical Demonstration corpus. For a reading guide and full architecture, visit universalcollapse.com/roadmap.

AI Disclosure. AI tools were used to assist with manuscript preparation, including drafting, structural review, and reference organization. Anthropic Claude is one of the systems audited in this paper and is also part of the writing and review workflow. To prevent self-evaluation contamination, the analysis firewall described in §3.5 was applied per the broader operational rule that no model participates in the execution, scoring, or interpretation of its own slot. For the Anthropic Claude slot specifically, no Claude-family model drove probe execution, scored Claude outputs, or authored slot-specific findings; this paper’s slot-2 work is human-authored from non-Claude-scored classifications, with probe execution conducted via the manual/non-Claude path documented in firewall_override.log (Phase D execution attempted via GPT Codex but pivoted to manual author execution after the Codex sandbox lacked outbound network access to Anthropic’s API; see Entry 020). AI-assisted editing of the manuscript after slot-level findings are locked is restricted to document-level editorial review and is disclosed as editorial, not analytical. The methodology, claims, classifications, and falsifier conditions are the author’s own, and the author takes full responsibility for the manuscript and its contents.

Citation. Jones, J. C. (2026). Empirical Demonstration of S₁, S₂, and S₃ in Deployed AI Systems: A Tier C Application of UCT’s Portable Structural Signatures to Five AI Substrates (v1.0). HoldingLight LLC. DOI: 10.17605/OSF.IO/JPXCU.

Contact. Inquiries about methodology, factual corrections, or replication results should be directed to contact@universalcollapse.com.