P117

AIEP — Parametric Unburdening Architecture for AI Reasoning Systems

Applicant: Neil Grassby / Phatfella Ltd

status: review — to file

Priority: Claims priority from GB2519711.2 filed 20 November 2025

Classification: Patent Application — Confidential

Proof of concept: Piea — piea.ai

FIELD OF THE INVENTION

[0001] The present invention relates to the architecture of artificial intelligence reasoning systems.

[0002] More particularly, the invention relates to a system architecture in which an AI reasoning model is relieved of all parametric knowledge storage obligations by binding it to a hash-addressed distributed evidence substrate — the AIEP substrate — such that the model’s parameter count is determined solely by its reasoning capability requirements, independent of the breadth or depth of factual knowledge required to perform its function.

[0003] The invention further relates to the use of a governed AIEP-compliant mirrored knowledge surface — the mirrored web — as the live, continuously-updated external knowledge corpus from which the model retrieves evidence by ContentHash on demand, replacing parametric knowledge storage entirely.

BACKGROUND

[0004] Large language models (LLMs) currently deployed in AI reasoning systems are large for a specific reason: they are required to carry knowledge in their parameters. The model must encode factual relationships, domain-specific information, temporal context, linguistic patterns, and world-state representations within its weights. This is because no governed external knowledge substrate exists from which the model can retrieve specific facts deterministically at inference time.

[0005] The consequence is a continuous and compulsory relationship between knowledge breadth and model size. A model that must know more must have more parameters. A model deployed across more domains must be larger. A model kept current must be retrained. There is no architectural path within the parametric paradigm to break this relationship.

[0006] The bloat problem — the sustained growth in model parameter counts required for capable general AI reasoning — is therefore not a technical inefficiency. It is a structural property of an architecture that has no governed external knowledge substrate. The model is large because it has nowhere else to put the knowledge.

[0007] Prior approaches to model compression — quantisation, distillation, sparse activation, mixture-of-experts — address inference cost within the parametric paradigm. They do not address the root cause. A distilled model is a smaller model that knows less. A mixture-of-experts model routes to specialist sub-models, each of which still carries parametric knowledge. None of these approaches eliminate the dependency between knowledge breadth and parameter count.

[0008] Retrieval-augmented generation (RAG) represents a partial externalisation of knowledge — providing a model with document context at inference time. RAG reduces the model’s need to memorise specific documents but does not provide a governed, hash-addressed, evidence-weighted substrate. Retrieved documents are injected into the context window without canonical identity binding, provenance declaration, evidence weight, or constitutional governance. The model cannot distinguish a fabricated document from a retrieved one. Hallucination remains possible. The context window grows with the volume of injected material.

[0009] No existing system provides: a governed hash-addressed external knowledge substrate where every artefact has a canonical ContentHash identity; a deterministic retrieval protocol that returns an artefact or null, with no interpolation path; evidence weighting for retrieved artefacts reflecting provenance, corroboration, recency, and schema validity; a mirrored web surface where AIEP-compliant publishers declare structured, hash-bound, provenance-declared knowledge artefacts; or a reasoning model architecture that requires zero parametric knowledge and whose parameter count is therefore determined solely by reasoning capability.

SUMMARY OF THE INVENTION

[0010] The invention provides a Parametric Unburdening Architecture for AI reasoning systems comprising three components in governed composition: a ParametricUnburdenedReasoningModel; a HashAddressedSubstrate; and a MirroredKnowledgeSurface. Together, these components eliminate the dependency between an AI model’s factual knowledge requirements and its parameter count.

[0011] A ParametricUnburdenedReasoningModel is an AI reasoning model whose parameters encode reasoning capability only — the capacity to evaluate evidence, form conclusions, detect contradictions, manage reasoning branches, and quantify uncertainty. The model encodes no factual knowledge. Its parameter count is determined by the complexity of these reasoning operations, not by the breadth or depth of knowledge required to perform them.

[0012] A HashAddressedSubstrate is a governed evidence corpus in which every evidence artefact is identified by a ContentHash computed as a deterministic cryptographic hash over the canonical form of the artefact. Two artefacts with identical content produce identical ContentHashes. A modified artefact produces a different hash. Retrieval is a deterministic lookup: a ContentHashRequest returns the artefact or a NullArtefactRecord. There is no interpolation path and no fabrication path.

[0013] A MirroredKnowledgeSurface is the aggregate of AIEP-compliant machine mirror endpoints published by content providers on the open internet. Each mirror endpoint publishes structured, hash-bound, provenance-declared representations of content artefacts in the canonical form defined by the AIEP protocol (P60–P63). The MirroredKnowledgeSurface is continuously updated as publishers update their mirror endpoints. It has no knowledge cutoff.

[0014] At inference time, when the ParametricUnburdenedReasoningModel requires a specific piece of knowledge to continue a reasoning branch, it emits a ContentHashRequest. The HashAddressedSubstrate resolves the request through a priority-ordered chain: local swarm participants, federated swarm, MirroredKnowledgeSurface crawler, open internet crawler. The model receives the artefact or a NullArtefactRecord. In the NullArtefactRecord case, the model records a P16 NegativeProofRecord — a cryptographic commitment that the evidence does not exist in reachable evidence space.

[0015] The technical effect of this architecture is the elimination of parametric knowledge storage from AI reasoning systems. A ParametricUnburdenedReasoningModel deployed over a rich HashAddressedSubstrate and MirroredKnowledgeSurface achieves equivalent or superior factual accuracy to a parametric LLM with substantially greater parameter count, because its factual accuracy is determined by the evidence quality of the substrate, not by the completeness of its training data.

The Parametric Bloat Equation — Eliminated

[0016] In a parametric LLM, the relationship between knowledge breadth K, knowledge depth D, and parameter count P is monotonically increasing: P = f(K, D). For any fixed reasoning capability R, increasing K or D requires increasing P.

[0017] In the Parametric Unburdening Architecture, knowledge breadth K and knowledge depth D are properties of the HashAddressedSubstrate, not the model. The model’s parameter count P is determined solely by reasoning capability R: P = f(R). K and D are substrate properties with no effect on P.

[0018] This is the core claim. The parametric bloat equation is eliminated by architectural separation. The model stops growing when knowledge grows. Only the substrate grows. The substrate grows at zero marginal model cost.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

1. ParametricUnburdenedReasoningModel

[0019] The ParametricUnburdenedReasoningModel is trained exclusively on reasoning operations within an AIEP substrate context. Training data comprises: constitutional governance constraint evaluation tasks; evidence weight assessment across domain-classified artefacts; branch formation, archival, and reactivation operations; contradiction detection and ContradictionRecord generation (P112); uncertainty quantification producing four-field UncertaintyRecords (P105); ContentHashRequest emission when knowledge is required; and NegativeProofRecord interpretation when hash resolution fails.

[0020] The model is explicitly trained not to generate factual claims from parametric memory. When the model encounters a query requiring a fact it has not retrieved, it emits a ContentHashRequest rather than generating a response from internal weights. This is enforced architecturally: the model’s output vocabulary distinguishes between reasoning operations and knowledge requests. Knowledge request tokens trigger the hash retrieval protocol. They do not generate text.

[0021] The parameter count of a ParametricUnburdenedReasoningModel in a preferred embodiment is in the range of 100M to 3B parameters. This range is sufficient for the reasoning operations enumerated in paragraph [0019]. It is substantially smaller than parametric LLMs deployed for equivalent factual reasoning tasks, which require 7B to 405B parameters to encode sufficient knowledge.

2. HashAddressedSubstrate

[0022] The HashAddressedSubstrate implements the AIEP Evidence Ledger (P80) as the canonical storage layer. Every artefact is admitted through the AIEP corpus admission pipeline (P10, P17, P24, P25, P112) and assigned a ContentHash on admission. ContentHash is computed as sha256(CanonicalForm(artefact)) where CanonicalForm applies P10 normalisation.

[0023] The HashAddressedSubstrate is partitioned into three access tiers: local substrate (artefacts admitted by this node’s own corpus loader, access in microseconds); local swarm (artefacts held by proximate swarm participants within P91-governed consent scope, access in milliseconds); and federated swarm (artefacts held by broader swarm participants within session scope, access in tens of milliseconds).

[0024] A ContentHashRequest that cannot be resolved from any swarm tier triggers the MirroredKnowledgeSurface crawler. A ContentHashRequest that cannot be resolved from any source generates a NullArtefactRecord comprising the requested ContentHash, the resolution attempt chain, and a NullHash = H(RequestedContentHash ‖ ResolutionAttemptChain ‖ SchemaVersionId). A P16 NegativeProofRecord is appended to the Evidence Ledger.

[0025] Evidence weight is assigned to each artefact on admission and updated continuously as corroboration accumulates across swarm participants. A highly corroborated artefact — one independently admitted by many swarm participants — carries higher evidence weight than a singly-held artefact. The model receives evidence weight alongside the artefact in every ContentHash response and incorporates it into branch evidence weight computation.

3. MirroredKnowledgeSurface

[0026] The MirroredKnowledgeSurface is the aggregate of AIEP mirror endpoints (P60–P63) published on the open internet. An AIEP mirror endpoint publishes, for each content artefact: a canonical structured representation of the content; a ContentHash binding the representation; a probability metadata envelope (P66) declaring provenance type, confidence score, attestation source, and freshness indicator; and plausibility constraint declarations (P67).

[0027] The MirroredKnowledgeSurface crawler accesses mirror endpoints with priority over raw HTML. A mirror-resolved artefact enters the HashAddressedSubstrate at substantially higher initial evidence weight than an artefact resolved from raw internet content, because the mirror endpoint’s ContentHash, provenance declaration, and plausibility constraints provide structural confidence in the artefact’s integrity that raw HTML does not.

[0028] The MirroredKnowledgeSurface has no knowledge cutoff. As publishers update their mirror endpoints, the updated artefacts become available for ContentHash resolution. The ParametricUnburdenedReasoningModel’s effective knowledge currency is therefore the currency of the MirroredKnowledgeSurface — continuously updated, with no training cycle required.

4. Proof of Concept — Piea

[0029] The present invention is demonstrated in proof-of-concept form by Piea (piea.ai), a governed AI reasoning system implementing the Parametric Unburdening Architecture over a Cloudflare Worker substrate with Cloudflare D1 Evidence Ledger and KV swarm cache. Piea’s reasoning core is implemented using a slimline open-source model (Qwen 2.5, Mistral Small 3.1 24B, DeepSeek-R1-Distill-Qwen-32B) operating within the AIEP substrate. The proof of concept demonstrates: ContentHash-addressed evidence retrieval from local swarm and MirroredKnowledgeSurface; hash-bound reasoning branches with founding tension lineage; P94 anticipatory surfacing; P112 atomic claim cross-verification; P90 swarm consensus in session context; and P116 always-on substrate continuity independent of client session state.

[0030] The Piea proof of concept validates the core claim: a slimline reasoning model operating over a governed hash-addressed substrate performs governed factual reasoning tasks without requiring parametric knowledge storage, and produces UncertaintyRecords and NegativeProofRecords that parametric systems cannot generate — because parametric systems have no architectural equivalent of a hash miss.

5. The Anti-Bloat Invariant

[0031] The Anti-Bloat Invariant is the formal statement of the invention’s core claim: for any knowledge domain D with breadth B and depth d, adding D to the HashAddressedSubstrate has zero effect on the ParametricUnburdenedReasoningModel’s parameter count. Parameter count is invariant under knowledge expansion.

[0032] This invariant holds because knowledge is never encoded in model weights. It is always retrieved by ContentHash. The model’s parameter count depends only on the complexity of reasoning operations, which does not increase as knowledge breadth increases.

[0033] The Anti-Bloat Invariant implies a compounding efficiency property: as the HashAddressedSubstrate accumulates more artefacts across more domains, the ParametricUnburdenedReasoningModel becomes more capable — in the sense that it can reason across more domains with greater factual accuracy — without any increase in inference cost, parameter count, or training requirement.

CLAIMS

1. A Parametric Unburdening Architecture for an AI reasoning system, the architecture comprising: a ParametricUnburdenedReasoningModel whose parameters encode reasoning operations only and encode no factual knowledge; a HashAddressedSubstrate in which every evidence artefact is identified by a ContentHash computed as a deterministic cryptographic hash over the canonical form of the artefact; and a hash retrieval protocol by which the ParametricUnburdenedReasoningModel emits ContentHashRequests when factual knowledge is required and receives in response the identified artefact or a NullArtefactRecord; wherein the parameter count of the ParametricUnburdenedReasoningModel is determined by reasoning capability requirements independent of the breadth or depth of factual knowledge held in the HashAddressedSubstrate.

2. The architecture of claim 1 wherein the hash retrieval protocol resolves ContentHashRequests through a priority-ordered chain comprising local substrate, local swarm participants, federated swarm, and MirroredKnowledgeSurface crawler, with no interpolation path and no knowledge fabrication path.

3. The architecture of claim 1 wherein a NullArtefactRecord is generated upon hash resolution failure, comprising the requested ContentHash, the resolution attempt chain, and a NullHash; and wherein a P16 NegativeProofRecord is appended to the Evidence Ledger, constituting a cryptographic commitment that the evidence does not exist in reachable evidence space.

4. The architecture of claim 1 wherein each artefact returned in response to a ContentHashRequest carries an evidence weight reflecting provenance, corroboration count across swarm participants, recency, and schema validity; and wherein the ParametricUnburdenedReasoningModel incorporates the evidence weight into branch evidence weight computation.

5. The architecture of claim 1 further comprising a MirroredKnowledgeSurface comprising the aggregate of AIEP-compliant machine mirror endpoints, wherein each endpoint publishes a ContentHash-bound structured representation of each content artefact with a probability metadata envelope and plausibility constraint declarations.

6. The architecture of claim 5 wherein artefacts resolved from AIEP-compliant mirror endpoints are admitted to the HashAddressedSubstrate at higher initial evidence weight than artefacts resolved from non-mirror sources.

7. The architecture of claim 1 wherein the ParametricUnburdenedReasoningModel’s output vocabulary distinguishes reasoning operations from ContentHashRequests, and wherein ContentHashRequest tokens trigger the hash retrieval protocol without generating text from model weights.

8. The architecture of claim 1 wherein the Anti-Bloat Invariant holds: for any knowledge domain added to the HashAddressedSubstrate, the parameter count of the ParametricUnburdenedReasoningModel is unchanged.

9. A method for eliminating parametric knowledge storage from an AI reasoning system comprising: training a model exclusively on reasoning operations within an AIEP substrate context without encoding factual knowledge in model weights; providing a HashAddressedSubstrate with deterministic ContentHash-based artefact retrieval; emitting ContentHashRequests from the model when factual knowledge is required during inference; returning artefacts or NullArtefactRecords in response to ContentHashRequests; and recording NegativeProofRecords for unresolvable ContentHashRequests.

10. A computing system comprising one or more processors and memory storing instructions which, when executed, perform the method of claim 9.

Drawings

FIG. 1 — Architecture diagram (see filed application for figures)

Figure 1 — ParametricUnburdenedReasoningModel and ContentHashRequest

User Query
     │
     ▼
┌─────────────────────────────────────┐
│  ParametricUnburdenedReasoningModel  │
│  encodes reasoning operations only  │
│  no factual knowledge in weights    │
└──────────────────┬──────────────────┘
                   │ requires factual knowledge
                   ▼
ContentHashRequest emitted
(ContentHash of required artefact)
     │
     ▼
HashAddressedSubstrate resolves

Figure 2 — HashAddressedSubstrate Resolution

ContentHashRequest(ContentHash)
     │
     ▼
┌─────────────────────────────────┐
│   HashAddressedSubstrate          │
│   lookup by ContentHash           │
└────────────┬────────────────────┘
           │ FOUND              NOT FOUND
           ▼                    ▼
      return canonical      NullArtefactRecord returned
      artefact              NegativeProofRecord committed
                            to Evidence Ledger
                            (cryptographic proof of absence)

Figure 3 — MirroredKnowledgeSurface (Anti-Bloat Invariant)

AIEP-compliant machine mirror endpoints (aggregate)
     continuously updated
     hash-addressed
     provenance-declared
     no training cutoff
     │
     ▼
MirroredKnowledgeSurface
     │
Anti-Bloat Invariant:
param_count(Model) = CONSTANT
regardless of knowledge expansion in substrate
Adding knowledge ──► zero effect on model size

Figure 4 — ContentHash-Addressed Artefact Lifecycle

External source (live web via Mirror)
     │
     ▼ AIEP Mirror ingestion
ContentHash computed over artefact bytes
Artefact committed to Evidence Ledger
with hash, timestamp, provenance
     │
     ▼
ContentHash stored in
HashAddressedSubstrate index
retrievable by any ContentHashRequest
from the ParametricUnburdenedReasoningModel

ABSTRACT

A Parametric Unburdening Architecture for AI reasoning systems is disclosed. A ParametricUnburdenedReasoningModel encodes reasoning operations only — no factual knowledge — and emits ContentHashRequests when factual knowledge is required during inference. A HashAddressedSubstrate resolves ContentHashRequests by returning the canonical artefact identified by ContentHash or a NullArtefactRecord when resolution fails. A NegativeProofRecord is committed to the Evidence Ledger upon resolution failure, constituting a cryptographic proof of evidence absence. A MirroredKnowledgeSurface — the aggregate of AIEP-compliant machine mirror endpoints — provides a continuously-updated, hash-addressed, provenance-declared external knowledge corpus with no training cutoff. The Anti-Bloat Invariant is established: the parameter count of the ParametricUnburdenedReasoningModel is invariant under knowledge expansion. Adding knowledge to the substrate has zero effect on model size. The architecture eliminates the dependency between factual knowledge breadth and model parameter count that drives LLM bloat. The proof of concept is Piea (piea.ai), a governed AI reasoning system demonstrating the architecture over a live AIEP substrate.

© 2026 Phatfella Ltd. All rights reserved. AIEP — Architected Instruction & Evidence Protocol. GB2519711.2 filed 20 November 2025.