P126 � AIEP � Dissent, Outlier and Alternative Consensus Engine
Publication Date: 2026-03-13 Status: Open Source Prior Art Disclosure Licence: Apache License 2.0 Author / Organisation: Phatfella Ltd Schema: AIEP_OS_SPEC_TEMPLATE v1.0.1 � https://aiep.dev/schemas/aiep-os-spec-template/v1.0.1 Related Filings: GB2519711.2 (Core Protocol) � GB2519826.8 (Hardware Layer) Related Specs: P83 P106 P107 P110 P112
Framework Context
[0001] This disclosure operates within an Architected Instruction and Evidence Protocol (AIEP) environment as defined in United Kingdom patent application number GB2519711.2, filed 20 November 2025, the entire contents of which are incorporated herein by reference.
[0002] This disclosure defines a governed ideation and dissent engine � the AIEP Dissent, Outlier and Alternative Consensus Engine � that accepts a design problem or specification decision as input, instantiates a configurable pool of behavioural personas sourced from open-source models, runs a structured swarm ideation loop to generate independent positions from each persona, scores positions for divergence from consensus, surfaces outlier positions with full reasoning chains, and emits a structured dissent report carrying session-graded new art flags.
Background
[0003] Consensus lock-in is an architectural risk in any iterative design process driven by a single reasoning system. When the same model or the same reasoning pattern is applied repeatedly to adjacent problems, the output converges toward familiar mechanisms � not because better alternatives do not exist, but because the model’s training distribution assigns them lower prior probability than the consensus path.
[0004] Existing approaches to handling disagreement in distributed AI systems fall into two categories: (a) Byzantine fault-tolerant consensus mechanisms such as PBFT and Tendermint, which treat minority views as faults to be suppressed in order to achieve agreement; and (b) statistical evidence combination frameworks such as Dempster-Shafer Theory of Evidence, which aggregate conflicting evidence mathematically into a single combined belief. Both approaches eliminate the dissenting position as an artefact � the minority view is either overruled or absorbed into a combined estimate.
[0005] There exists a need for a governed dissent mechanism in which: (a) dissenting and outlier positions are preserved as structured, replayable, first-class artefacts rather than suppressed or aggregated; (b) the consensus view and all outlier positions coexist in the same output record, available for future arbitration when new evidence arrives; and (c) positions that reference mechanisms not present in the active corpus are flagged for human follow-up rather than silently discarded.
Summary of the Disclosure
[0006] This disclosure defines the AIEP Dissent, Outlier and Alternative Consensus Engine: a governed swarm ideation loop in which a configurable persona pool generates independent positions on a defined problem, positions are scored for divergence from consensus using a DivergenceScore computed across semantic embedding space, outlier positions are preserved with full reasoning chains as OutlierPosition artefacts, and a DissentReport is emitted carrying: the consensus view, ranked outlier positions, DivergenceScore per outlier, NewArtFlag list graded by scope, RecombinationScore for cross-domain positions, and a ResourceBoundary field cataloguing every model and corpus active in the session.
Part I � Epistemic Boundary Constraint
[0007] The engine operates under a non-negotiable epistemic boundary: it cannot surface positions that lie outside the combined training distribution and loaded resource of its active persona pool. This constraint governs the interpretation of all engine outputs and is a first-class design property.
[0008] Practical consequences of the epistemic boundary:
(a) OutlierPositions are outliers within the loaded space � a high DivergenceScore means far from consensus among active personas, not novel to the world;
(b) NewArtFlags are graded signals, not confirmations of novelty � a flagged position requires a formal prior art search; the flag is the prompt, not the result;
(c) Forking is combinatorial, not inventive � the engine recombines existing mechanisms in new assemblies; it does not generate knowledge that does not exist in its corpus;
(d) ResourceBoundary depth sets the ceiling � more diverse open-source models loaded widens coverage of the known space; the ceiling still exists and must be stated.
[0009] Every DissentReport carries a resource_boundary object cataloguing all models active during the session, all spec corpuses loaded, and the session timestamp. This makes the epistemic limit legible and auditable.
Part II � Persona Pool Architecture
[0010] The persona pool is model-agnostic and role-governed. Persona definitions are stored in a versioned PersonaPoolDefinition JSON document. Each persona entry declares:
persona_id� unique identifier within the sessionrole� aRoleDefinitionstring specifying reasoning behaviour (not model identity)model_binding� the open-source model assigned to this role at runtimesource_repository� GitHub / HuggingFace Hub repository reference for the assigned modeldomain_emphasis� optional domain bias (e.g.,"patent_law","systems_architecture","regulatory_compliance","end_user","adversarial")
[0011] Role definitions are composable. Example roles: challenge_assumptions, reason_from_first_principles, synthesise_cross_domain, apply_patent_counsel_frame, red_team_failure_modes, regulatory_reviewer, cost_and_feasibility_analyst, ethicist, end_user_advocate.
[0012] Model-to-role binding is a runtime configuration. The PersonaPoolDefinition schema is model-independent, so persona pools upgrade automatically as better open-source models become available without schema revision.
Part III � Swarm Ideation Loop
[0013] The swarm ideation loop proceeds in two rounds:
Round 1 � Independent position generation. Each persona generates a position and brief reasoning summary for the defined problem statement without cross-persona communication. No persona receives any other persona’s output during Round 1. Independence is enforced by the loop controller.
Divergence scoring. Positions are compared using semantic embedding distance to identify: (a) the ConsensusCluster � the majority-aligned position set; and (b) the ranked OutlierPosition list � positions with DivergenceScore = 0.65. DivergenceScore is a normalised distance from the ConsensusCluster centroid computed across the semantic embedding space of positions. Score range: 0.0 (identical to consensus) ? 1.0 (maximally divergent).
Round 2 � Chain-of-thought on top outliers. Top-ranked outlier positions are re-run with chain-of-thought reasoning to produce a full ReasoningChain per outlier. A RecombinationScore cross-domain check is performed for each outlier (defined in �IV). NewArtFlag evaluation is applied per outlier (defined in �V).
[0014] The loop is illustrated schematically as follows:
Problem Statement (input)
�
?
PersonaPoolDefinition (versioned JSON, model-agnostic)
�
+----------------------------------+
� Instantiate each persona independently �
+----------------------------------+
�
+----?----------------------------------+
� Round 1: Independent position generation �
� - No cross-persona communication �
� - Each persona generates position + �
� brief reasoning summary independently �
+---------------------------------------+
�
+----?----------------------------------+
� Divergence scoring �
� - ConsensusCluster identified �
� - OutlierPosition list ranked by �
� DivergenceScore (0.0 � 1.0) �
+---------------------------------------+
�
+----?--------------------------------------+
� Round 2: Chain-of-thought on top N outliers �
� - Full ReasoningChain per outlier �
� - RecombinationScore (cross-domain check) �
� - NewArtFlag evaluation per outlier �
+-------------------------------------------+
�
?
DissentReport (JSON + MD summary)
Part IV � RecombinationScore
[0015] For each OutlierPosition, a cross-domain mechanism check is performed: does the position assemble mechanisms from two or more distinct AIEP spec domains? Score range 0.0�1.0. Positions above 0.5 are flagged recombination_candidate.
[0016] The RecombinationScore definition is aligned with P106 (AGI Cross-Domain Recombination Engine) ConceptualDistanceScore. Positions flagged recombination_candidate are cross-referenced against the P106 ConceptualDistanceScore definition and may serve as RecombinantBranchCandidate inputs if the position is later promoted to a specification.
Part V � NewArtFlag Grading
[0017] NewArtFlags are graded by scope. Two grades are defined:
novel_within_session � the position references a mechanism not present in: (a) the active spec set; (b) the abandoned spec set; (c) any loaded corpus resource. Scope is limited to what the engine can see.
requires_prior_art_search � a novel_within_session flag that also involves a mechanism with potential real-world application outside the loaded corpus. This grade triggers a mandatory human follow-up: a formal prior art search must be conducted before the related specification is promoted.
[0018] Flags graded requires_prior_art_search must be cleared (recorded in the related spec’s frontmatter as new_art_flags: cleared) before promotion tooling will accept the related spec. Promotion checks enforce this as an error, not a warning.
[0019] NewArtFlag resolution workflow:
NewArtFlag generated
�
?
Grade: novel_within_session?
�
YES � NO ? discard (mechanism present in corpus)
�
?
Real-world application potential?
�
YES � NO ? log as novel_within_session, no block
�
?
Grade: requires_prior_art_search
�
?
Human follow-up ? formal prior art search
�
+---------+
� �
CLEAR RETAIN
� �
? ?
cleared: cleared:
true false ? blocks promotion
until resolved
Part VI � DissentReport Structure
[0020] The DissentReport is emitted as structured JSON. The schema is defined as follows:
{
"session_id": "string",
"pf_ref": "string",
"problem": "string",
"resource_boundary": {
"models_active": ["string"],
"spec_corpus_version": "string",
"abandoned_specs_loaded": "boolean",
"session_timestamp": "ISO8601",
"note": "Positions are outliers within this resource set only."
},
"consensus_view": "string",
"outliers": [
{
"rank": "integer",
"persona_id": "string",
"role": "string",
"model_binding": "string",
"position": "string",
"reasoning_chain": ["string"],
"divergence_score": "number (0.0�1.0)",
"recombination_score": "number (0.0�1.0)",
"recombination_candidate": "boolean",
"new_art_flags": [
{
"flag_id": "string",
"mechanism": "string",
"grade": "novel_within_session | requires_prior_art_search",
"cleared": "boolean",
"cleared_by": "string | null",
"cleared_at": "ISO8601 | null"
}
]
}
],
"generated_at": "ISO8601"
}
Part VII � Relationship to Prior AIEP Specifications
[0021] P83 (Recall Divergence Control System) � P126 uses P83 divergence scoring concepts adapted for ideation space rather than evidence-committed substrate operations.
[0022] P107 (AGI Outlier Fork Extension) � P107 governs outlier promotion within a live substrate reasoning session. P126 governs pre-specification outlier generation at design time. P126 positions may become inputs to P107-governed forks if subsequently promoted to a specification.
[0023] P106 (AGI Cross-Domain Recombination Engine) � P126’s swarm loop includes a recombination pass; outlier positions that combine mechanisms across distinct spec domains are flagged with a RecombinationScore aligned with P106’s ConceptualDistanceScore definition.
[0024] P112 (Atomic Claim Extraction and Cross-Verification) � dissent reports optionally invoke P112-style claim extraction on each outlier position to decompose compound claims before the new art flag evaluation.
[0025] P110 (Human-Participant Governed Session Swarm) � the persona pool architecture mirrors P110’s ParticipantCompatibilityScore design; P126 personas are analogous to P110 participant nodes except they are model-instantiated rather than human-operated, and session state is not committed to any substrate.
Drawings
Fig. 1 � Swarm ideation loop architecture
+---------------------------------------------------------+ � DISSENT ENGINE � � � � Input: Problem Statement � � � � � ? � � PersonaPoolDefinition (versioned JSON) � � +-----------------------------------------+ � � � P1 � P2 � P3 � P4 � P5 � Pn � � � �role �role �role �role �role �role � � � �model �model �model �model �model �model � � � +-----------------------------------------+ � � � � � � � � � � ? ? ? ? ? ? � � [Independent position generation � Round 1] � � � � � � � � � � +----------------------------------+ � � � � � ? � � Divergence Scoring � � ConsensusCluster ?---- majority � � OutlierList ?---- score = 0.65 � � � � � ? � � Round 2: CoT on top outliers � � RecombinationScore � � NewArtFlag grading � � � � � ? � � DissentReport (JSON + MD) � +---------------------------------------------------------+
Fig. 2 � NewArtFlag resolution
NewArtFlag generated
�
?
novel_within_session? --NO--? discard
�
YES
�
?
real-world application? --NO--? log (no block)
�
YES
�
?
requires_prior_art_search
�
?
formal prior art search
�
CLEAR --? cleared: true (unblocked)
RETAIN--? cleared: false ? blocks promotion
This specification is published as open-source prior art under the Apache License 2.0. Copyright 2026 Phatfella Limited. All rights reserved. Publication date: 13 March 2026.