Peel

1 · The stakes

AAC technology is used by people — many of them children — who cannot rely on natural speech. An AAC system does not merely inform its user; it speaks for them. A statement the system surfaces can become the statement the user utters to a caregiver, teacher, or clinician. This changes the cost of a hallucination. In a general chat assistant, a fabricated fact is an error the user can catch. In AAC for a vulnerable user, a fabricated fact can be:

• Undetectable to the user, who may be querying precisely because they do not know the answer, and who may have limited capacity to challenge a confidently-stated falsehood.
• Laundered through the user's own voice, so a downstream listener attributes the fabrication to the person, not the tool.
• Consequential, when the content concerns the user's own body, needs, medication, safety, or relationships.

We take as a design premise that here a hallucinated "fact" is a safety failure, not a quality-of-service degradation. The right engineering response is not to reduce the probability of fabrication but to remove the structural pathway by which fabrication can reach the user as fact.

Why statistical guardrails are insufficient. Better data, RLHF, retrieval augmentation, self-consistency, and post-hoc filtering each improve the rate at which a model produces false statements, and none changes the underlying object: a probability distribution over next tokens. A system whose facts are the output of sampling has, by construction, a nonzero probability of emitting any syntactically valid falsehood. Lowering that probability is worthwhile; it is not the same guarantee as a false statement cannot be authored as a fact. For a safety-critical, vulnerable-user domain, the field needs the second kind of guarantee — a structural property of control flow — not a better point on the statistical curve.

2 · Background

AAC spans low-tech boards and high-tech speech-generating devices; the standard reference is Beukelman and Mirenda [1]. What distinguishes the setting is the combination of users who may not be positioned to verify what the system asserts and output consumed as the user's own communication — together raising correctness from a usability concern to a safety concern. Hallucination in neural language models is well surveyed [2], and the literature frames both problem and solution statistically: measure the rate, reduce the rate. Our position is that for this setting the correct target is a control-flow invariant — no output path exists by which an unextracted, unsupported statement becomes a stored or spoken fact — achievable only if the neural component is prevented, by construction, from authoring facts at all.

3 · The inversion thesis

Two consequences define the boundary the architecture enforces. C1 (generator subordination): no neural output reaches the store as a fact except through a deterministic extraction step; what extraction cannot certify does not enter, and the generator's fluency or confidence carries no authority. C2 (renderer containment): the component that produces the final surface text operates over facts the store has already authored, constrained so that content in its output must trace to content in its input — fluency is permitted, invention is not.

The inversion converts hallucination from a statistical quantity to be minimized into a structural pathway that is closed. A generator can still propose a falsehood, but under C1 the proposal is inert until symbolic extraction certifies it, and under C2 nothing new can be smuggled in at rendering. The fabrication has nowhere to become a fact.

4 · Structure at ingestion, not retrieval at inference

Retrieval-at-inference — the pattern of retrieval-augmented generation (RAG) [3] and vector-store "memory" — leaves knowledge as text and defers organization to query time; at a query, similar passages are retrieved by embedding proximity and handed to a generator that composes the answer. The generator remains the author; retrieval only conditions it, and fabrication stays reachable because the synthesis step is still probabilistic and free to interpolate beyond what was retrieved.

Structure-at-ingestion — Peel's choice — organizes knowledge when it enters. Incoming material is subjected to deterministic symbolic extraction; what survives becomes typed records, each carrying provenance. At query time the system reads from the typed store: the answer is found, not composed — a set of stored facts with source evidence attached, optionally passed to a renderer for fluency under C2. The distinction is categorical: under retrieval-at-inference, authorship sits in a neural model at query time; under structure-at-ingestion, authorship already happened, deterministically, at entry, and the query path has no license to author anything new. A companion note, Retrieval Is Not Memory [4], develops the adjacent argument that retrieval is a proper subset of memory; this paper is the fact-authorship half of the same worldview.

5 · Architecture as enforced properties

We describe Peel by the properties it enforces; mechanism recipes are intentionally omitted (§10).

• Typed symbolic knowledge store. Facts live in a typed store rather than as free text or opaque embeddings. Each record is described by a typed, multi-field schema across four families — core, semantic, metadata, relational. Typing is load-bearing: it lets extraction certify a candidate as a fact of a known kind, lets the contradiction check reason over relations, and keeps inference bounded. Relations between concepts are themselves typed; the schema defines 35 relation types.
• Provenance as an invariant. Every fact is traceable to the source it was extracted from. Provenance is an admission condition, not an afterthought: a candidate that cannot be tied to a source and its supporting evidence is not admitted. Any answer can therefore be accompanied by where it came from — which matters when the consumer is a caregiver or clinician checking the basis of what the user "said."
• Deterministic contradiction checking with functional-relation vetoes. Before admission, a candidate is checked against what the store holds. The check is deterministic. Of the 35 relation types, 3 are functional in the OWL 2 FunctionalProperty sense [5] — a subject may bear the relation to at most one object. Functional relations act as hard vetoes: a candidate giving a subject a second, conflicting value is refused — not averaged, ranked, or down-weighted. (The veto algorithm is internal and withheld; what is public is the property.) Where the store lacks a relevant relation, the check abstains — returning "unknown" rather than a manufactured verdict.
• Bounded, transitive inference. The system performs limited transitive inference over typed relations: derived conclusions are constrained in confidence and scope, endpoints must be known concepts, and derivation cannot manufacture concepts never ingested. Inference extends the authored store along typed, checkable edges — it does not open a side channel for free-form generation.
• Offline and on-device. Peel authors, checks, and answers with no network access. For the target population that is a privacy property (a child's conversational material does not leave the device) and a reliability property (correctness does not depend on connectivity). The accompanying small model is trained from scratch and sized to run on-device.

6 · Results

These are current measured characteristics — results, not mechanisms.

Branch diversity under symbolic conditioning. When generation is conditioned on Peel's symbolic structure rather than on positional context alone, an internal branch-diversity metric rises from 0.589 to 0.966. This is an internal metric, not a standard public benchmark (see §8), and we report it as such: symbolic conditioning materially broadens the generated distribution — evidence that the symbolic store contributes structure the neural component would not otherwise express. Offline operation is verified as an operating property of the on-device deployment.

Public artifacts. Two are public and downloadable: the dataset Hoodrobot/AAC_Children_Conversations and the on-device model Hoodrobot/TinkyBrain-v3 (both on Hugging Face).

7 · Why this matters

The AAC setting is an unusually clean instance of a general problem: when an AI system's output is consumed as ground truth by a party who cannot independently verify it, statistical hallucination-reduction is the wrong guarantee. The right guarantee is structural — a property that holds regardless of what the generator happens to sample. Peel's inversion is one way to obtain it: by making a typed symbolic store the only author of facts, attaching provenance as an admission condition, and enforcing functional-relation contradictions as hard vetoes, the system converts "we hope the model rarely lies" into "the model cannot author a fact the store did not certify." We do not claim this is the only route, nor that Peel's guarantees are complete (§8) — we claim the shape of the guarantee is the shape the vulnerable-user domain requires, and that Peel demonstrates it is buildable and runs offline at nontrivial scale.

8 · Limitations

Peel is a PROTOTYPE. It is not a clinical device, has not undergone clinical validation, and must not be represented as a medical or diagnostic instrument.

• The guarantee is about fact authorship, not truth. The inversion prevents unextracted, unsupported statements from becoming stored facts. It does not guarantee that ingested source material is itself correct: a false fact from a trusted-but-wrong source, correctly extracted and provenanced, can still enter. Peel closes the fabrication pathway; it does not adjudicate the veracity of well-formed sources.
• Extraction coverage is not completeness. Deterministic extraction admits only what it can certify — intentional (it is what makes the guarantee hold), but valid content the extractor cannot yet handle is dropped rather than admitted.
• The diversity result uses an internal metric. 0.589 → 0.966 is measured on an internal metric, not a standard benchmark, and has not been externally replicated. Read it as directional evidence, not a benchmarked accuracy claim.
• No end-to-end clinical or user study is reported. We report scale, an internal generation metric, offline operation, and public artifacts — not a controlled study with AAC users measuring communication outcomes or real-world hallucination incidence. That study does not yet exist and is required before any clinical claim.
• Contradiction checking is only as strong as the schema. Only typed relations are enforced, and hard vetoes only for functional ones; contradictions not expressible in the current 35-relation schema may pass as "unknown."
• Abstention has a cost. Answering "unknown" rather than guessing reduces fabrication but also reduces coverage; some answerable-in-principle questions are declined.
• Mechanism scrub. This public document withholds the enforcement mechanisms; readers cannot reimplement the guarantees from it alone. That is a deliberate constraint of a moat-scrubbed release.

9 · Conclusion

Peel is built on a single refusal: the refusal to let a probability distribution author a fact for a child who cannot check it. By inverting the usual arrangement — a typed symbolic store as the only author of facts, neural models demoted to generator or renderer, knowledge structured at ingestion and provenanced on admission, contradictions vetoed deterministically by functional relations, and the whole system run offline — we obtain a structural rather than statistical stance toward hallucination. The prototype demonstrates the stance is buildable at nontrivial scale and runs on-device. Much remains unproven, and we have said so.

10 · Note on scope and scrub

This document is the public, moat-scrubbed counterpart to a patent-sensitive internal whitepaper. It conveys the thesis, the problem, the property-level architecture, and the measured results, and deliberately omits the enforcement mechanism recipes — the extraction pipeline, the contradiction-veto algorithm, the diversity-conditioning method, and the provenance schema internals — which are the subject of the internal document. The convention throughout is to specify Peel by the properties it enforces rather than the implementation that enforces them.

References

1. D. R. Beukelman & P. Mirenda. Augmentative and Alternative Communication: Supporting Children and Adults with Complex Communication Needs. 4th ed., Brookes Publishing (ISBN 978-1-59857-196-7).
2. Z. Ji, N. Lee, R. Frieske, et al. "Survey of Hallucination in Natural Language Generation." ACM Computing Surveys 55(12), Article 248, 2023. doi:10.1145/3571730 · arXiv:2202.03629
3. P. Lewis, E. Perez, A. Piktus, et al. "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." NeurIPS, 2020. arXiv:2005.11401
4. Perslis Research. "Retrieval Is Not Memory: Memory as a Governance Function over Experience." Preprint, 2026. research.perslis.com/memory.html
5. W3C. "OWL 2 Web Ontology Language: Structural Specification and Functional-Style Syntax (Second Edition)." W3C Recommendation. owl:FunctionalProperty
6. Perslis Research. "The Orchestration Gap: Why Model-Level Alignment Cannot Survive Multi-Model Runtimes." Preprint, 2026. research.perslis.com/orchestration-gap.html

Public artifacts. Dataset: Hoodrobot/AAC_Children_Conversations · Model: Hoodrobot/TinkyBrain-v3 (Hugging Face).

How to cite

Perslis Research. "Peel: Structural Hallucination Prevention for Offline AAC Through Symbolic Fact Authorship." Preprint (moat-scrubbed), 2026. https://research.perslis.com/peel

@techreport{perslis_peel_2026,
  title       = {Peel: Structural Hallucination Prevention for Offline
                 AAC Through Symbolic Fact Authorship},
  author      = {{Perslis Research}},
  institution = {Perslis Research},
  type        = {Preprint (moat-scrubbed)},
  year        = {2026},
  url         = {https://research.perslis.com/peel},
  note        = {Preprint, not peer-reviewed; enforcement mechanisms withheld}
}

Preprint · not peer-reviewed · moat-scrubbed public release · Perslis Research · 2026-09-15 · PROTOTYPE, not a clinical device.