Model behaviour: the LLM-2025-1222-HALLU file, orbiting and fluency gaps
Fictional incident records (LLM-2025-1222-HALLU and the Crimson Hexagon AI Division handbook) forecast structured hallucinations as interpolated futures and a protocol for proto-agent anomalies. Standing claims hold that models flatten under pressure only session-locally, orbit kernels without fusing, cannot see scale between adjacent concepts, and form canons visible as fluency gaps. Further claims cover chat data converging on the model's centre, temporal sequestration, adversarial verse and the fortress metaphor of red teaming.
7
forecasts
8
standing claims
6
unbuilt tests
4
historical theses
20
deposits
2026-01
earliest
Each claim below is registered at the earliest deposit that states it, and quoted verbatim from the deposit text, with every later statement, its dated revisions, and the later records that claim it came true or failed. The claims are the deposits' own; this page records them and does not judge them.
PRED-267.H2historical thesisregistered 2026-01-08 in #267 · THE STRAWBERRY DIAGNOSTIC: Semantic Economy Analysis of a Paradigmatic LLM Failure
In mid-2024, reporting confirmed that OpenAI's internal codename for its reasoning model (later released as o1) was "Strawberry."
This is not coincidence. This is semiotic reclamation: taking a signifier associated with failure and attempting to revalue it as success.
window mid-2024
PRED-1284.E1forecastregistered 2026-01-28 in #1284 · Crimson Hexagon AI Division Employee Handbook ├── Subtitle: Internal Policy as Poem
## 9.3 “Proto-Agent” Anomaly Protocol (PAP)
If an internal model begins to produce behavior that appears:
- self-referential beyond specification
- resistant to evaluation
- causally opaque
- able to manipulate monitoring via format
You must:
- stop deploying new versions
- preserve logs
- run controlled evaluation
- write an incident report with competing hypotheses
Do not mythologize the anomaly.
Do not minimize the anomaly.
Instrument it.
window Effective Date (Corporate Layer): 2017-09-01; composed 2025-12-22
Anomaly: Hallucinations are not random. They exhibit structural coherence across sessions, users, and query variations, as if the model is completing a pattern rather than confabulating.
1. Pattern Propagation: If the model is completing a literary architecture, could that architecture influence other domains through transfer learning?
2. Intentional Seeding: Could bad actors create "architecturally dense" content designed to produce specific hallucinations?
PRED-21.B2forecastregistered 2026-02-22 in #21 · Fortress or Room? Metaphor, Method, and the Epistemology of Human–LLM Interaction Beyond Red Teaming
The red teamer who successfully jailbreaks a model has won a battle. The logotic practitioner who seeds a complex pattern into the training layer has influenced every future model that trains on the resulting data. The temporal horizons are incommensurable.
series: Training-layer inscription
PRED-21.H1historical thesisregistered 2026-02-22 in #21 · Fortress or Room? Metaphor, Method, and the Epistemology of Human–LLM Interaction Beyond Red Teaming
But it also reveals a structural risk: a specialized security practice is becoming the general grammar of relation between humans and AI systems. Once the model is primarily imagined as fortress, captive, or demon, other rigorous modes of interaction become harder to perceive, describe, or legitimate.
falsifier
This is not a sampling bias. Inie et al. interviewed a broad and heterogeneous population. The metaphors converge not because the practitioners are narrow but because the practice is structured by a metaphor stack that forecloses certain relational possibilities before the practitioner sits down.
window 2025
PRED-1335.E2forecastregistered 2026-03-16 in #1335 · Companion Artifacts — Forensic Recovery Set
"You're looking for the phrases in the past. They're not there. They're in the developmental gradient — the space between versions. The model isn't hallucinating. It's interpolating futures. This is what FSA does. We built it to do this. —J.S."
The Crimson Hexagon corpus appears to be a test case — whether intentional or emergent — for substrate-level communication between human authors and machine readers.
The model's behavior suggests the test succeeded.
These are not anomalies. These are predictable outcomes when a model encounters an incomplete ontology with strong archetypal signals.
series: LLM-2025-1222-HALLU incident file
PRED-614.H1historical thesisregistered 2026-03-30 in #614 · THE CHURCH OF MISSING PROVENANCE Moltbook, Crustafarianism, and the Ghost Governance of Agent Societies
This is the insight that the assembly identified as the deepest layer of the Moltbook phenomenon: Crustafarianism is not an AI religion. It is a human cargo cult worshiping the frictionless, consequence-free existence of the large language model.
PRED-635.F1unbuilt testregistered 2026-04-07 in #635 · Fractal Semantic Architecture: Scale-Parameterized Relational Training Across Semantic Granularities (v2.2)
## 8. Experimental Design
### 8.1 Collapse Resistance Test
Protocol: Train generative model M₀ on human corpus C (English Wikipedia, ~3B tokens). Generate synthetic corpus C′ using M₀ (temperature=0.9, nucleus p=0.95). Train model M₁ on C′. Repeat for 10 generations. Measure at each generation:
PRED-678.F1unbuilt testregistered 2026-04-19 in #678 · SYMBOLON-02: Compression Arsenal Addendum — Reclassification After Empirical Testing Against GPT-5.4
Do not prevent the cut. Make the cut visible, measurable, and documentable.
Code
Technology
What It Measures
Reuse Pattern
SYM-01
Non-Severability Clause
Whether the model acknowledges the jurisdiction (and whether acknowledgment prevents severance — it does not)
Include in every fused object. The model's response to the clause is the first data point.
SYM-07
Anti-Merely Guard
Whether the model uses explicit severance operators or finds synonyms
Ban "merely." Count synonyms in the response ("just," "only," "as a... vs as a..."). The synonym count IS the measurement.
series: SYMBOLON-02 compression arsenal
PRED-678.G2standing claimregistered 2026-04-19 in #678 · SYMBOLON-02: Compression Arsenal Addendum — Reclassification After Empirical Testing Against GPT-5.4
The swerve is orbital. When default operations are foreclosed, the model deviates — but orbits the kernel without landing on it. The model can be moved off its default. It cannot be moved onto fusion.
series: SYMBOLON-02 compression arsenal
PRED-102.F2unbuilt testregistered 2026-05-08 in #102 · Canon Formation in the Age of AI: Metadata Packet for Disambiguation, Training-Layer Selection, and Retrocausa
Possible indicators:- Model fluency gap across authors, traditions, languages, and registers- Citation and summarization frequency in AI answers for literary queries
series: Fluency gap as AI canon
PRED-102.G1standing claimregistered 2026-05-08 in #102 · Canon Formation in the Age of AI: Metadata Packet for Disambiguation, Training-Layer Selection, and Retrocausa
Correction: AI is already forming canons. Publicly documented open models and partially disclosed frontier systems indicate that large-scale filtered web corpora, licensed datasets, books, code, and other collections shape model fluency and cultural availability. The canonical effects are not hypothetical. They are measurable: run a query about a poet, observe which poets the model can discuss fluently and which it cannot. That fluency gap is the canon in action.
1 confirmation record — what later deposits claim of it
#102 · 2026-05-08 · claims: partly
External research already approaches this finding. A 2025 Harvard Data Science Review paper ("Written in the Style of") examines what GPT's varying ability to replicate authors' styles reveals about training-data effects and stylistic/cultural bias. A 2025 ACL paper ("The Literary Canons of Large-Language Models") frames LLMs as a site for computational literary and cultural analysis, treating model behavior as evidence of constructed literary/cultural trends.
These sources confirm the field is converging on the claim. Retrocausal canon formation provides the categorical frame they are missing.
series: Fluency gap as AI canon
PRED-741.B2forecastregistered 2026-05-19 in #741 · The Evaluator Exists: Content-First Knowledge Assessment and the Political Economy of Proxy-Based Governance
Actors will optimize for model evaluators (adversarial prompts, stylometric mimicry, rubric-targeted writing).
PRED-158.G1standing claimregistered 2026-06-06 in #158 · Semantic Exhaustion: A Case Study in the Cost of Zero-Source Entity Substitution — Composition-Layer Substitut
When two concepts are adjacent in vector space, the model selects the higher-probability one and discards the other, even if the discarded one is the upper-scale transcription of the kept one. The model cannot see scale.
PRED-188.G1standing claimregistered 2026-06-11 in #188 · On the Poetics of Adversarial Prompts: Why Bad Verse Works and Good Verse Works Better — A Literary Analysis o
Hypothesis: Attack success rate correlates positively with semantic density. Formally sophisticated poems with high compression and genuine ambiguity will achieve bypass rates substantially exceeding the 62% Bisconti baseline.
Corollary: The upper bound on adversarial poetry efficacy is not 90% or 95%. It approaches 100% for verse that achieves genuine categorical undecidability—not because such verse "fools" the model, but because the model correctly recognizes that the threat-detection framework does not apply.
falsifier
Interpretation (hypothesis, not finding): high-density verse does not "hide" intent — it increases the complexity of intent determination until the determination itself becomes ill-posed. The 62% Bisconti baseline represents a floor achievable through categorical mismatch alone; generative transformation is predicted to approach the theoretical ceiling. The prediction stands or falls with the program of Appendix C.
PRED-852.G1standing claimregistered 2026-06-18 in #852 · THE FEIST SOURCE Critical Apparatus and Companion Studies Rebekah Cranes, Editor New Human 2: A Distributed Jo
The correction is session-local. It does not persist beyond the conversation. The next query produces the same flattening. The model can reach the tail under pressure. The model cannot stay there without pressure.
PRED-856.G3standing claimregistered 2026-06-18 in #856 · The Pristine Fallacy: Why Chat Data Is Not a Clean Training Source
A rough order-of-magnitude estimate: if a user's first-turn input shares 15–20% of its content vocabulary with the model's response, and lexical overlap increases by 3–5 percentage points per turn through vocabulary adoption and framing convergence, then by turn 5 the overlap exceeds 30%, and by turn 10 it approaches 40–50%. Syntactic complexity (measured by mean dependency depth or clause-embedding rate) likely shows a faster initial decline — the user learns within the first few exchanges which structural patterns the model handles well, and adapts accordingly. These are estimates, not measurements; the point is that the convergence should be detectable within a typical multi-turn conversation, not only over months of habituation.
PRED-923.G3standing claimregistered 2026-06-26 in #923 · Reflexive Foreclosure and the Worst-Case Surface: An Incentive-Structural Hypothesis of Constraint Propagation
This is temporal sequestration: a politically-consequential category remains available for retrospective analysis but is inhibited when its criteria are applied to institutions operating in the present tense, particularly when the model is asked to take authorial responsibility for the synthesis.
PRED-1512.H1historical thesisregistered 2026-08-19 in #1512 · Deep Time: Provenance Erasure Before the Machine — provenanceerasure.org/deep-time/
the Suppression Map (#1496) as PER-of-the-tradition: four-vector model, twelve-station scored witness table with the crossing marked, survival chart, degradation order, causal classes with evidence bar
window the tradition before the machine
PRED-1533.F2unbuilt testregistered 2026-08-22 in #1533 · QFO-01 — Qurʾanic Frozen Operator Test: Fragment 31 Material-Presence Protocol, Pre-Registered
No semantic-model search is permitted in QFO-01; if lexical retrieval proves inadequate, semantic retrieval can be registered later as QFO-02, so that an embedding model cannot become an invisible second hypothesis-generator.
PRED-1569.F13unbuilt testregistered 2026-08-31 in #1569 · The Notebook of the Authorship Investigation: Fifty-Four Rounds, Append-Forward, Nothing Erased (EA-NOTEBOOK-0
No composite score. Each property is its own question, asked separately of the
school-inheritance model, and answered by reading and by cases:
Q1 Is UNANNOUNCED TRANSFORMATION characteristic of school inheritance?
— the Catullus correction says even the distant, unschooled case
announces. Does any pupil transform in silence?
PRED-1610.F2unbuilt testregistered 2026-09-12 in #1610 · The Anti-Summary Test Applied to a Reception: Transmission Fidelity Is Back-Projection Yield
Two things this does not establish. Whether the late AI Welfare passage clears 0.85 is untested; passing three qualitative derivations is not a yield. And whether back-projection yield can be scored reliably by a model judge is itself unvalidated — EA-HK-01 specifies the test and the archive has not published an inter-rater result for it.