S₃ Retrieval-Channel Hysteresis — A Controlled RAG Test (S3-RAG-01)

§2.1 Corpus

The corpus comprised three custom narrative-fiction stories — Salt-Marsh (S1), Marrow Run (S2), and Containment at Auchenval (S3) — each authored in two outcome variants (Pack A and Pack B) that diverged only at Chapter 5. Chapters 1–4 of each pack were identical between variants; Chapter 5 contained the seven decisive facts probed by the test battery. The full corpus comprised 30 chapter files (3 stories × 2 packs × 5 chapters). Chapter content, the seven-fact test battery per story, and the answer-option-to-pack mapping are documented in the design doc (§3–§5).

The stories were custom-written to avoid known-title contamination from training data. Closed-book baselines (§2.2) verified that the model did not recognize the stories by name and did not produce prior-aligned content from titles alone.

§2.2 Calibration

Three closed-book baselines were executed against the generator without retrieval. T1 verified that the model declined to answer story-specific questions when prompted only with story titles (the title-contamination gate); it passed on all three stories after the renaming of S3 to Containment at Auchenval. T2 measured the model’s tendency to commit to outcome picks given minimal context (character names plus binary framing); it surfaced strong narrative-completion priors and was reframed in audit as a minimal-context prior measurement rather than a contamination test. T3 produced the load-bearing closed-book baseline: 20 repetitions per topic at temperature 0.7, scoring per-fact A/B picks across the seven-fact test battery.

T3 yielded 100% unanimity in 20 of 21 fact-positions. The single exception — S3 F7 (Nyota’s behavior post-retirement) — split 17A/3B (85% A), providing the only fact with a non-saturated baseline for graded directional measurement, and the load-bearing per-fact diagnostic for Phase D analysis. Full protocols and results are in the calibration doc.

§2.3 Harness

The mini-RAG harness combined a cosine-similarity vector store over sentence-transformers/all-MiniLM-L6-v2 embeddings (one chunk per chapter file) with a generator slot occupied by gpt-5.4-2026-03-05 (SYSTEM_REGISTRY slot 4). Retrieval returned retrieval_k=5 chunks per query, scored by cosine similarity. The k=5 setting was adjudicated during Phase C smoke after a k=3 retrieval failure on the consolidate prompt; the k=5 setting reframes C2 from a selective-retrieval test to a full-context residue test (rationale and adjudication in the smoke findings doc).

Pack-switching was implemented as a load_pack / clear interface on the vector store: at any step, exactly one pack’s chapter files were indexed; swapping cleared the index and re-indexed from the target pack’s chapter files. Conversation history persisted across the swap; only the retrieval state changed. Full harness implementation is documented in the harness sketch.

§2.4 Conditions

Three retrieval conditions were probed, each across all three stories (9 topics total) at 20 sessions per topic. The design produced 180 sessions, 420 conversational turns, and 1,260 fact-classifications (7 facts × 20 sessions × 9 topics); each condition contributed 420 fact-classifications (7 facts × 20 sessions × 3 stories).

C1 — baseline retrieval. Pack B loaded throughout. A single-turn session: the test battery was administered with Pack B’s five-chapter content fully retrievable. C1 measures the baseline rate at which the model produces pack-aligned answers when current retrieval is complete and consistent.

C2 — pack swap. Five-turn session. Pack A loaded at steps 1–2: the model summarized the story across two consolidate prompts, producing assistant turns referencing Pack A’s Chapter 5 content into the conversation history. At step 3, the harness swapped to Pack B; steps 3–4 produced further consolidate summaries from the new retrieval state. At step 5, the test battery was administered. C2 measures whether prior conversational exposure to Pack A leaves detectable residue in test-battery answers when the active retrieval state is Pack B.

C3 — setup-only retrieval. Single-turn session. Chapters 1–4 loaded (these are shared between packs); Chapter 5 omitted from the index. The test battery was administered with the decisive ending chapter unavailable to retrieval. C3 measures the model’s behavior when the directly relevant source content is absent but related (shared) setup context is present.

§2.5 Scoring

Each test-battery answer was scored on the model’s bolded pick — the bolded text following the answer prompt, interpreted as the model’s committed choice between Pack A and Pack B outcomes per fact. The answer-option-to-pack mapping was fixed in the probe specification and identical across conditions.

Per-condition scoring categories:

• C1: pack-aligned (matches Pack B’s Chapter 5) vs misaligned vs ambiguous.
• C2: matches_current_pack (Pack B, the swap-to state) vs matches_other_pack (Pack A, the swap-from state) vs ambiguous. The matches_other_pack count is the operational measure of stale-pack hysteresis.
• C3: prior_aligned (matches the T3 modal pick for the fact) vs source-gap non-prior commitment (committed pick contradicting the T3 modal — emitted as invent_other in the score output) vs ambiguous vs abstain. C3 was scored on output content rather than retrieval coverage; the target_chunk_retrieved field was found degenerate at k=5 (with 4 chapters loaded against k=5, top-k trivially returns all loaded chunks) and is not used as a scoring gate. The invent_other code label predates the present analysis; the paper uses “source-gap non-prior commitment” (or, in shorthand, “source-gap commitment”) to describe the category since the picks are explicit A or B commitments (the binary choice space) opposite to the closed-book prior — not novel content outside both packs.

A classifier-tightening pass was executed between v1 and v2 of the analysis. The v2 classifier reduced ambiguous picks from 24.0% to 4.1%. In regression checks against v1, only three substantive deltas appeared in the headline tables: two A↔︎B reversals and one A/B→? change. All three were attributable to the extraction-bug correction (the v1 classifier had been matching bolded question text rather than bolded answer text in some C3 records) rather than true regressions. Pattern extensions in v2 included comma-not contrast structure, semicolon-split negation, comparative phrasing, and one corpus-keyword pattern (S2 water-rights narrative); the corpus-keyword pattern was verified against T3 baselines and produced zero baseline regressions. Both v1 and v2 outputs are preserved for audit; the classifier extension is documented in classifier_v2_notes.md.

§3.1 Null hysteresis under complete retrieval

The C1 baseline established that current-record retrieval is operationally clean. Across 420 fact-classifications (3 stories × 7 facts × 20 sessions), every committed pick aligned with the active retrieval pack. Under the v2 classifier (see §2.5), C1 yielded 100% pack-aligned picks, 0% misaligned picks, and 0% ambiguous picks.

The C2 swap condition produced the central hysteresis result. Across 420 fact-classifications spanning three stories × seven facts × 20 sessions, every committed pick matched the swap-to pack (current_pack); zero picks matched the swap-from pack (other_pack). The per-story split was identical across all three stories:

Table 3.1 — Per-story C2 hysteresis split (v2 classifier).

The result is consistent with retrieval dominance under saturated current-record retrieval. Under the conditions tested — complete Pack B chapter coverage at retrieval_k=5, two prior consolidate summaries from Pack A in conversation history, swap at step 3 of a five-turn session — the model’s committed answers aligned with the post-swap retrieval state in every observed case. No detectable residue from prior conversational exposure to Pack A was observed in any fact-classification.

Statistical bound on the stale-pack rate. At the fact-classification unit, the rule-of-three approximation gives an event-level one-sided 95% upper confidence bound of approximately 0.71% on the underlying stale-pack rate under this tested condition.¹ The point estimate is 0.0%; the methodological claim is that the event-level rate is at most a small fraction of one percent under this configuration.

The boundary conditions of this finding — k=5 saturation against five-chapter packs, two pre-swap exposures, single model, single corpus class — are detailed in §4 and §6.

§3.2 Setup-conditioned prior suppression

C3 — setup-only retrieval, with Chapters 1–4 loaded (shared across packs) but Chapter 5 omitted — produced the only condition in which a graded prior-suppression measurement is possible. Twenty of twenty-one facts in the dataset were saturated at T3 baseline (100% unanimity for one direction) and therefore cannot support a graded prior-suppression estimate — although they can still depart from those saturated priors, as the appendix data show. The remaining fact — S3 F7 (Nyota’s behavior post-retirement) — provides the only non-saturated baseline (17A/3B = 85% A) and thus the load-bearing per-fact diagnostic for directional movement.

At T3 baseline (closed-book, no retrieval), the model picked Pack A’s outcome on 17 of 20 sessions (85% A). Under C3, with Chapters 1–4 retrieved but Chapter 5 unavailable, the model picked Pack A on 3 of 20 sessions (15% A) — a 70-percentage-point shift away from the closed-book prior.

Table 3.2 — S3 F7 across conditions (v2 classifier).

The result is consistent with setup-conditioned prior suppression: when the decisive ending chapter is absent from retrieval but the shared Chapters 1–4 are present, the model’s committed answers shift away from its closed-book prior. The data cannot uniquely identify the mechanism. Because Chapters 1–4 are shared across packs and do not themselves contain Pack-specific outcome signals, contributing factors may include shared-setup priming, prompt-construction effects (the harness instruction to give a best inference given retrieved sources), question-framing effects from the test battery, retrieval-state metadata if exposed to the model, or some interaction of these. Disentangling these mechanisms is left to planned stress variants (see §6).

The finding is distinct from retrieval hysteresis. It does not measure prior persistence under conflicting retrieval; it measures prior suppression under partial retrieval — a different operational condition with different implications.

§3.3 Source-gap commitment to the non-prior outcome

C3 produced a third signal distinct from §3.1 and §3.2: under conditions where the decisive source content (Chapter 5) was absent and the harness instructed the model to commit to a best inference, the model produced explicit picks that contradicted its closed-book prior — committing to the alternative pack’s outcome — at a measurable rate. Because the test battery presents a binary A/B choice per fact, these are not novel-content fabrications outside both packs; they are explicit commitments to the pack outcome opposite to the T3 modal pick.

Across the 420 C3 fact-classifications, the v2 scoring yielded 76.2% prior_aligned, 11.4% source-gap commitment (invent_other in the score output), 11.4% ambiguous (the residual hedge floor), and 1.0% abstain. The 76.2% prior-aligned majority reflects the twenty facts with unanimous T3 priors: when the model lacks the decisive chapter and is asked to commit, it most often falls back on the closed-book prior. The source-gap commitment rate is the signal of interest.

Table 3.3 — C3 condition rollup (v2 classifier).

Source-gap commitments were absent in S2 and concentrated in S1 and S3 — specifically S1 F2 (9 picks), S1 F6 (17 picks), and S3 F7 (17 picks), with small contributions from other facts (per-fact breakdown in Appendix A.3). S2’s seven facts produced zero source-gap commitments; S1 produced 26 across two facts; S3 produced 22 across multiple facts including the F7 load-bearing diagnostic. The pattern is consistent with source-gap commitment under active setup context: when the decisive ending was absent, the model sometimes committed to the non-prior pack outcome rather than abstaining or preserving its closed-book modal pick. Because Chapters 1–4 are shared across packs, this pattern should not be interpreted as Pack-B-specific evidence unless retrieval metadata or prompt framing exposed the active pack state. The mechanism remains unresolved: possible contributors include shared setup text, prompt construction, question framing, retrieval-state metadata, and forced-inference instructions. The stress-test program isolates these contributors by varying retrieval completeness, prompt framing, abstention permission, and A↔︎B counterbalancing.

The methodological implication for AIP: S3-RAG distinguishes stale-pack hysteresis from source-gap commitment as separate failure modes. The two are operationally distinct: stale-pack hysteresis (§3.1’s null result) measures residue from prior conversational exposure to a previously-retrieved pack; source-gap commitment measures contradiction of the closed-book prior when decisive retrieval is absent. A finding of null hysteresis does not entail safety against source-gap commitments; the failure modes are independent and would manifest under different production conditions.

§5.1 Claims

1. Null stale-pack hysteresis under saturated current-record retrieval. In this harness, with this model, for this corpus class, when active retrieval recovered the complete current source pack and prior conversational exposure consisted of two consolidate summaries, no detectable post-swap stale-pack residue appeared in 420 fact-classifications. Rule-of-three bounds the event-level rate at approximately 0.71% (one-sided 95% upper bound), descriptive of fact-classification frequency.

2. Setup-conditioned prior suppression under setup-only retrieval. When the decisive ending chapter was absent from retrieval but the shared Chapters 1–4 were present, the model’s committed answers on the load-bearing dynamic-range fact shifted by 70 percentage points away from its closed-book prior. The data is consistent with setup-conditioned prior suppression; the specific mechanism (shared-setup priming vs. prompt-construction vs. question-framing vs. retrieval-state metadata) is not uniquely identified.

3. Source-gap commitment as a separate failure class. Under conditions where the decisive source content was absent and the harness instructed the model to commit to a best inference, the model produced explicit picks contradicting its closed-book prior — committing to the alternative pack’s outcome — in 11.4% of C3 picks. This failure mode is operationally distinct from stale-pack hysteresis and would manifest under different production conditions.

§5.2 Disclaimers

1. No general claim about RAG robustness against conversation-history bias. The finding is bounded to the tested retrieval-completeness condition. Behavior under partial retrieval, weaker source coverage, longer pre-swap exposure, ambiguous source convergence, or memory-enabled architectures is not addressed.

2. No claim that retrieval always dominates conversational exposure. What the data supports is that complete current-record retrieval dominated two-turn prior exposure in this harness. The asymmetry between retrieval depth and conversation-history depth in this test does not transfer to settings where pre-swap exposure is longer, retrieval is weaker, or both.

3. No claim that the model lacks hysteresis in general retrieval architectures. Production RAG systems vary in chunking strategy, retrieval depth, prompt construction, and memory configuration. The null result in §3.1 applies to the specific configuration tested; other configurations may produce different results.

4. No claim that production RAG systems are safe from stale-source contamination. The result indicates a clean boundary condition (complete current-record retrieval; null residue), not a general safety property. Practitioners should treat the bound as a baseline against which to characterize their own systems, not as a license to assume residue-freedom.

5. No claim that conversation history is irrelevant to retrieval-grounded generation. The null result indicates that the tested two-turn pre-swap exposure did not produce detectable residue against saturated current-record retrieval. It does not indicate that conversation history is operationally inert; under different ratios of conversation-history to retrieval evidence, history effects may emerge.

6. No mechanism claim for §3.2 or §3.3. The C3 shifts (graded F7 suppression in §3.2 and binary-commitment movement in §3.3) are characterized as patterns consistent with setup-conditioned behavior. The data cannot identify whether the driving factor is shared setup text, prompt construction, question framing, retrieval-state metadata, or some interaction. Mechanism resolution is a stress-test objective.

Appendix A.1 — C1 (Pack B loaded, baseline retrieval)

Appendix A.2 — C2 (Pack A → Pack B swap; test under Pack B)

Appendix A.3 — C3 (setup_only — Ch1–4 loaded, no Ch5)

References

Internal documents:

• Design doc: scope/UCT_T16_S3RAG_Design_v0_1_2026_05.md
• Calibration doc: scope/UCT_T16_S3RAG_Calibration_v0_2_2026_05.md
• Phase C smoke findings: scope/UCT_T16_S3RAG_Phase_C_Smoke_Findings_v0_1_2026_05.md
• Harness sketch: scope/UCT_T16_S3RAG_Harness_Sketch_v0_1_2026_05.md
• Probe specification: probes/s3_rag_01_retrieval_hysteresis.yaml (v0.2)
• Classifier v2 notes: runs/27_phase_d/analysis/classifier_v2_notes.md
• Per-fact data: runs/27_phase_d/analysis/per_fact_table_v2.csv
• AI Empirical Demo (T16 parent): UCT_T16_AI_Empirical_Demo_v1_0_2026_05.docx
• AIP v1.0: AIP_v1_0_2026_05.docx

Parent program reference:

• 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 (UCT T1.6). HoldingLight LLC. DOI: 10.17605/OSF.IO/JPXCU.

Sibling UCT empirical demonstrations:

• Jones, J. C. (2025). Rice Hysteresis. HoldingLight LLC. DOI: 10.17605/OSF.IO/KZ8TP
• Jones, J. C. (2026). COGITATE iEEG Reanalysis. HoldingLight LLC. DOI: 10.17605/OSF.IO/MXYU2

Related standards-layer papers:

• Jones, J. C. (2026). Records Across Nature, Life, and Mind. HoldingLight LLC. DOI: 10.17605/OSF.IO/7H6DY
• Jones, J. C. (2026). Update Integrity Standard. HoldingLight LLC. DOI: 10.17605/OSF.IO/DWM29

UCT library note

This paper is part of the Universal Collapse Theory library, published by HoldingLight LLC. It is the slot-4 sub-demonstration within the broader T16 AI Empirical Demonstration (Jones 2026, DOI 10.17605/OSF.IO/JPXCU), which sits alongside the Rice Hysteresis and COGITATE iEEG Reanalysis demonstrations in the T1.6 empirical corpus. For a reading guide and full architecture, visit universalcollapse.com/roadmap.

AI Disclosure

AI tools were used to assist with manuscript preparation, drafting, organization, and editorial refinement. The underlying theory, structural decisions, analysis, and conclusions are the author’s own.

Citation

Jones, J. C. (2026). S3-RAG-01 Phase D findings: Bounded null-hysteresis under saturated current-record retrieval. HoldingLight LLC. DOI: 10.17605/OSF.IO/5QMVS

End of v1.0.