Skip to content
InvariantEngine
Proposed Extension / Not Implemented 2026-05-14 by Joshua Couto

IE-004: Memory-Grounded Review Councils

Defines parallel critic councils whose findings are treated as evidence, not truth.

review-councilcriticsgovernance
Specification context: Runtime sections describe the target contract unless the implementation-status section explicitly marks them as implemented. See What works today for the private-alpha implementation boundary.

Implementation Status

Proposed Extension / Not Implemented. Review Councils are not currently implemented in the Engine runtime. There is no council runner, critic dispatcher, fresh-context synthesis pipeline, or council event/projection model today. This document remains a proposed governance pattern.

Abstract

Memory-Grounded Review Councils provide structured, multi-perspective evaluation of agent output. Multiple critics evaluate the same output independently, each grounded in durable memory. Their findings are aggregated into a council verdict that is treated as evidence for the Session Integrity Layer to consider, not as an authoritative decision. This specification defines how councils are composed, executed, and integrated into the runtime.

Problem

Single-perspective evaluation of agent output is brittle. A single critic may miss issues that another perspective would catch. Conversely, a single critic may raise false alarms that would be moderated by other viewpoints. Without structured multi-perspective evaluation, the system either over-relies on a single validation path or has no systematic way to reconcile conflicting assessments.

Definition

A Review Council is a set of independent critic evaluations that run in parallel against the same agent output. Each critic is grounded in durable memory and evaluates the output from a specific perspective (correctness, consistency, source grounding, invariant compliance, etc.). The council produces a verdict that summarizes findings and recommends actions.

Motivation

Treating any single evaluation as truth creates a new single point of failure. Review councils distribute evaluation across multiple perspectives and treat the aggregate result as evidence to be weighed, not as a final answer. This mirrors real-world review processes where multiple reviewers contribute findings that are synthesized into a decision.

Runtime Behavior

  1. Compose: A council is assembled from a configured set of critics, each with a defined perspective and grounding in durable memory.
  2. Execute: All critics evaluate the agent output independently and in parallel.
  3. Aggregate: Individual critic findings are collected and synthesized into a council verdict.
  4. Report: The verdict is provided to the Session Integrity Layer as evidence.
  5. Audit: All individual findings and the aggregate verdict are logged.

Inputs

  • Agent output to be evaluated
  • Council configuration (which critics to run, their perspectives)
  • Durable memory state
  • Relevant negative memory entries
  • Active invariants

Outputs

  • Individual critic findings
  • Aggregated council verdict
  • Confidence level
  • Recommended action
  • Audit log entries

Failure Modes

  • Groupthink: All critics share similar biases and miss the same issues.
  • Noise amplification: Minor findings from multiple critics aggregate into a false alarm.
  • Grounding gaps: Critics are not adequately grounded in relevant memory.
  • Latency: Parallel execution still introduces unacceptable delay.

Open Questions

  • How should critic perspectives be selected for different types of output?
  • How should conflicting critic findings be reconciled?
  • Should council composition be static or adaptive based on session context?
  • What is the minimum number of critics needed for a meaningful verdict?

Non-goals

  • A Review Council is not a debate system. Critics do not negotiate consensus with each other.
  • It is not a majority-vote truth mechanism. Verdicts are not decided by counting votes.
  • Critic outputs are not authoritative — they are evidence for synthesis and governance review.
  • Council verdicts do not override human authority or the Session Integrity Layer’s governance policies.
  • A Review Council is not a replacement for human review where policy requires it.

Relationship to Other IE Primitives

  • IE-001 Session Integrity Layer consumes council verdicts as evidence.
  • IE-002 Hallucination Gauge may incorporate council findings into drift scoring.
  • IE-003 Negative Memory provides context for critics to check against.
  • IE-005 Memory Proposal Workflow may be triggered by council recommendations.