From Consensus to Audit — Transforming multi-LLM reasoning from answer optimization into auditable decision evidence.
While current multi-agent reasoning frameworks (Multi-Agent Debate, ReConcile, ECON) focus on maximizing accuracy and driving agents toward consensus, high-stakes decisions require a different objective. Under risk, a definitive answer is less valuable than an auditable decision trace that explicitly surfaces hidden assumptions, conflicts, and minority warnings.
MRRA is a state-machine-based reasoning control architecture (S0–S7) that permits non-convergence, limits overconfidence, and prioritizes audit trail generation over forced consensus.
This repository contains the technical report, specification, analysis, and case studies for MRRA v0.1.1. The framework is implemented as the MRRA Core v0.8.0 skill for Hermes Agent.
MRRA operates across three tightly integrated layers:
S0 [User Query]
│
┌──────────▼──────────┐
│ 1. Problem Reframer │ ◄─── (S1 Heuristics)
└──────────┬──────────┘
│ Surfaced Assumptions
┌──────────▼──────────┐
│ 2. Triangulated │ ◄─── (S3 Solver/Skeptic/Alternative)
│ Evaluator │ ───► Interference Layer (4-Axis Analysis)
└──────────┬──────────┘
│ Competing Perspectives
┌──────────▼──────────┐
│ 3. Revision │ ◄─── (S4/S5 Persistent Registry)
│ Controller │ ───► Confidence Cap & Stopping Rules
└──────────┬──────────┘
│
S7 [Auditable Decision Evidence] (Conditional Recommendation)
Deconstructs and reframes questions using six cognitive heuristics (abstraction, decomposition, lateral reframing, estimation, projection, assumption elicitation) to surface implicit assumptions before evaluation begins.
Employs three role-specialized agents — Solver (constructs the strongest solution), Skeptic (detects risks and contradictions), and Alternative (proposes reframings) — to evaluate candidate directions. Uses the Interference Layer to classify cross-role outputs into four categories:
| Classification | Definition | Downstream Action |
|---|---|---|
| Agreement | Independent convergence across roles | Accepted as high-confidence evidence |
| Conflict | Contradictory conclusions | Triggers Revision Controller return |
| Complement | Each role addresses different facets | Integrated to strengthen conclusion |
| Absence / Absent Warning | A high-impact concern raised by only one role or absent from dominant consensus | Preserved and must be reviewed before final aggregation (protects minority views from majority pressure) |
When comparative decision-making is needed, candidates are scored via MCDA (six axes: independent agreement, evidence strength, logical consistency, implementability, risk controllability, originality).
Maintains a persistent Assumption Registry (audit trail of every assumption with confidence, importance, and verification status) and applies structural confidence governance:
- Confidence Cap: Critical + Low assumptions cap final confidence at Medium.
- Stopping Conditions (clean termination): all critical assumptions verified, conflicts within threshold, absent warnings reviewed, MCDA margin satisfied (when active).
- Conditional Termination: when clean conditions fail, recommendation is conditional on resolving residual uncertainties. MRRA permits non-convergence rather than forcing an unverified conclusion.
-
C1: Reasoning Control Architecture — A structured, recursively-controlled state machine (S0–S7) designed for traceability rather than simple conversational debate.
-
C2: Interference Analysis & Absent Promotion — A structured evaluation primitive that prevents minority warnings (Absence) from being erased by majority pressure.
-
C3: Conservative Confidence Governance — A mechanism that caps final confidence based on unverified assumptions, conflicts, and absent warnings, producing conditional recommendations under residual risk.
| File | Description |
|---|---|
MRRA-Technical-Report-v0.1.1.pdf |
Compiled technical report published on Zenodo (v0.1.1, June 10, 2026). |
MRRA-Technical-Report-v0.1.1.md |
Markdown source of the Technical Report / Extended Abstract. |
MRRA-Technical-Report-v0.1.1-package.zip |
Complete snapshot bundle for reproducibility. |
case-study-zero23w-v0.1.md |
Case Study 2: Web access-control design evaluation (domain contrast against crypto trading). |
related-work-matrix-v0.1.md |
13-paper × 10-dimension positioning matrix, mapping MRRA against the SOTA. |
related-work-structure-v0.1.md |
Literature review section structure, contribution mapping, and 7-element uniqueness table. |
extended-abstract-v0.1.md |
Extended abstract summarizing the MRRA framework. |
references.bib |
Comprehensive BibTeX database of surveyed literature, including the MRRA Zenodo DOI. |
Among surveyed works, MRRA is designed with a focus on auditability and structural uncertainty control rather than benchmark-optimized convergence:
| Dimension | Debate Frameworks (MAD / DMAD / ReConcile) | Game-theoretic (ECON) | Calibration (UQ / FermiEval) | MRRA (This Work) |
|---|---|---|---|---|
| Primary Goal | Accuracy / consensus | Equilibrium / efficiency | Statistical coverage | Auditability / decision trace |
| Assumption Tracking | None | None | None | Persistent Registry (Audit Trail) |
| Minority Protection | Weak | None | None | Absent Promotion (Mandatory Review) |
| Confidence Handling | Self-reported or calibrated | Implicit (belief) | Probabilistic calibration | Structural Capping under Unresolved Risk |
| Evaluation Target | Benchmark accuracy | Benchmark performance | CI coverage | Assumption coverage / risk detection |
"Among surveyed works, we did not identify prior work that jointly treats these mechanisms as first-class components of a reasoning-control architecture."
The full 13-paper comparative matrix with explicit MRRA positioning is available in related-work-matrix-v0.1.md.
A L2.5 strategy change for a crypto trading bot was evaluated through MRRA Deep mode. The Interference Layer detected 3 Conflicts, 1 Agreement, 2 Complements, and 4 Absence warnings. MCDA scored the best candidate at 4.4/10. The Solver's proposal was rejected — Confidence Cap activated due to Critical+Low assumptions about market regime persistence.
A proposed policy change (retroactively restricting publicly accessible magazine issues to premium-only) was evaluated. The Interference Layer identified a key Absent Warning — the notification process design was missing from all perspectives. Two Critical+Low assumptions triggered the Confidence Cap. Final recommendation was conditional on three preconditions.
| Metric | bitflyerbot | zero23w |
|---|---|---|
| Assumptions surfaced | 9 (Deep) | 9 (Deep) |
| Critical risk detected | 3 | 2 |
| Absent warnings preserved | 4 | 1 |
| Confidence Cap activated | Yes | Yes |
| Recommendation type | Conditional | Conditional |
If you use MRRA in your research, please cite the Technical Report as follows:
@techreport{maruko2026mrra,
title={MRRA: An Auditable Reasoning Control Architecture for Role-Specialized LLM Ensembles},
author={Maruko, Yoshifumi},
institution={Zenodo},
year={2026},
number={v0.1.1},
doi={10.5281/zenodo.20618374},
url={https://doi.org/10.5281/zenodo.20618374}
}MRRA is implemented as a production-grade skill set for Hermes Agent. The core implementation includes:
- mrra-core (v0.8.0) — Full S0–S7 state machine, Assumption Registry, CLI tools
- triangulated-reasoning (v4.2.0) — Role-specialized Solver/Skeptic/Alternative evaluation
- marukoshiki-reasoning — Six-cognitive-heuristic problem reframing
- Interference Layer — 4-axis structured cross-perspective analysis
- MCDA Module — 6-axis multi-criteria decision scoring with observation preset profiles
Implementation source is available under the MRRA Core Skill repository.
- Technical Report (Zenodo): 10.5281/zenodo.20618374
- Case Study (zero23w Web Access Control):
case-study-zero23w-v0.1.md
This work is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).
For detailed specifications, refer to the MRRA Technical Report PDF or contact the author through the repository issues.