P251 — AIEP — Dissent-Aware Compute Scheduler

Applicant: Neil Grassby Classification: Patent Application — Confidential Priority: Claims priority from GB2519711.2 filed 20 November 2025 Architecture Layer: AIEP Phase 2 Multi-Agent Scheduling Layer

Framework Context

[0001] This specification operates within an AIEP environment as defined in GB2519711.2 and GB2519798.9. The present specification defines a compute scheduling mechanism that adjusts compute allocation between agents in the multi-agent reasoning ensemble in response to dissent patterns, favouring agents producing high-value dissenting signals.

Field of the Invention

[0002] The present invention relates to dissent-aware compute scheduling for multi-agent AI reasoning ensembles.

Background

[0003] In multi-agent reasoning ensembles, agents with minority positions may produce critical dissent signals that prevent premature convergence on incorrect conclusions. Standard compute schedulers treating all agents equally or scheduling purely for throughput may starve dissenting agents of compute at critical moments, suppressing valuable dissent signals.

Summary of the Invention

[0004] The invention provides a Dissent-Aware Compute Scheduler (DACS) that: monitors dissent output history of each agent in the ensemble; maintains a dissent value score per agent based on how frequently its dissent signals have been validated by subsequent evidence; assigns compute priority weights that increase for agents with high dissent value; and reallocates compute budget dynamically when high-value dissent is in progress.

[0005] Under standard conditions, agents receive equal compute allocation. When high-value dissent is detected, the dissenting agent’s allocation is increased, and the DRREE (P235) is notified.

ASCII Architecture

Multi-Agent Ensemble (P209)
         |
         v
+-------------------------------------------+
| Dissent-Aware Compute Scheduler (DACS)    |
|                                           |
|  Per-agent dissent value score           |
|  Compute priority weight calculation     |
|  Dissent detection                       |
|  Dynamic compute reallocation            |
|  DRREE notification (P235)              |
+-------------------+-----------------------+
                    |
                    v
  Adjusted compute allocation per agent
  → Resource Allocation Engine (P213)

Detailed Description

[0005] Dissent Value Score. The DACS maintains a per-agent dissent value score. The score is incremented when an agent’s dissent position is subsequently validated by new evidence admissions or designated as high-impact by the governance review process. The score is decremented when dissent positions are found to be evidentially unsupported after resolution.

[0006] Priority Weight Assignment. Compute priority weights are computed as: weight = base_allocation + (dissent_value_score * dissent_bonus_coefficient). Weights are normalised to sum to 1.0 across the ensemble.

[0007] Active Dissent Reallocation. When an agent produces a live dissent signal, the DACS temporarily doubles its compute allocation (subject to total budget constraint) to ensure the dissenting position is fully developed before the ensemble closes on a conclusion.

[0008] Integration with DRREE. When dissent value is high and dissent is active, the DACS flags this to the Dissent-Responsive Reasoning Escalation Engine (P235) for potential escalation action.

Technical Effect

[0009] The invention provides a compute resource allocation mechanism that systematically rewards evidentially validated dissent in multi-agent AI ensemble reasoning, ensuring that historically productive dissenting agents receive proportionally greater compute for future reasoning while controlling total budget expenditure. By temporarily doubling compute allocation for agents with live dissent signals, the engine ensures minority positions receive sufficient computation to be fully developed before the ensemble concludes. This reduces the risk of premature consensus suppressing important evidential exceptions, improving reasoning quality without requiring constant escalation.

Claims

1. A method of dissent-aware compute scheduling for a multi-agent evidence-bound artificial intelligence ensemble, comprising the steps of: (a) maintaining a per-agent dissent value score, incrementing it when an agent’s prior dissent position is subsequently validated by new evidence admissions or designated high-impact by governance review, and decrementing it when a prior dissent position is found evidentially unsupported after resolution; (b) computing a compute priority weight for each ensemble agent as: weight = base_allocation + (dissent_value_score × dissent_bonus_coefficient), normalising all weights to sum to 1.0; (c) on detection of a live dissent signal from an agent, temporarily doubling the agent’s compute allocation, subject to total session budget constraint, until the dissent is resolved or a maximum dissent exploration budget is reached; (d) when high dissent value score and an active dissent signal coincide, flagging the event to the Dissent-Responsive Reasoning Escalation Engine for potential escalation; (e) recording all allocation decisions and dissent value score updates as evidence artefacts admitted to the evidence ledger.

2. The method of claim 1, wherein the dissent_bonus_coefficient is a configurable parameter set in the governance policy document.

3. The method of claim 1, wherein weight normalisation at step (b) is recalculated on each reasoning step, reflecting current dissent value scores of all ensemble members.

4. The method of claim 1, wherein the maximum dissent exploration budget at step (c) is expressed as a percentage of total session compute budget.

5. The method of claim 1, wherein sustained allocation of doubled compute to an agent that subsequently fails to produce evidentially supported conclusions results in a recalibration step reducing the dissent_bonus_coefficient for that agent.

6. A dissent-aware compute scheduler for a multi-agent evidence-bound artificial intelligence system, comprising: a dissent value score ledger maintaining per-agent historical dissent outcomes; a priority weight calculator; a live dissent allocation module providing temporary doubled compute allocation; and an escalation interface to the Dissent-Responsive Reasoning Escalation Engine.

7. A computer-readable medium carrying instructions for implementing the method of any preceding method claim.

Abstract

A dissent-aware compute scheduler for multi-agent evidence-bound artificial intelligence maintains per-agent dissent value scores reflecting the historical evidence-validation rate of each agent’s dissent positions. Compute priority weights proportional to dissent value scores are assigned to ensemble members. Agents producing live dissent signals receive temporarily doubled compute allocation to ensure full development of dissenting positions before the ensemble concludes. High-value active dissent events are flagged to the Dissent-Responsive Reasoning Escalation Engine. All allocation decisions are recorded in the evidence ledger.