A human-seed, multi‑LLM reasoning laboratory. BD gives the seed; LLMs amplify; MAL synthesizes; builders build; arenas, data, and evidence decide.
v0.98 benchmark hardening6 Mental Modes4 Mental EnginesAI-ONLY / HUMAN × AIpublic/local splitno API keys on public page27-domain world application mapPrize Frontier · verified 19 Jun 2026MAL → AI8 / Reasojjng
Purpose
Not a chatbot. A reasoning workflow lab.
MentalArena is an external operational learning layer above LLMs. It does not train or modify models. It improves protocols, context packs, scoring, lineage, builder prompts, evidence reports, and next-run seeds.
Core rule: unusual human seeds are not rejected at the idea stage. Objections become assumptions, variants, controls, tests, metrics, risk labels, or implementation warnings.
Mental Modes
Mode chooses the epistemic style.
Innovation
Seed Amplifier
For new concepts and speculative invention. No seed rejection; convert critique into tests.
Builder Prompt
Code Spec
For producing the strongest task-level builder prompt from sources, requirements, and acceptance gates.
Factual
Source-Grounded
For history/science/fact-heavy articles. Source ledger and uncertainty outrank style.
Creative-Realistic
System Upgrade
For MAL/app/protocol upgrades: imaginative but buildable, with constraints and MVP paths.
Artistic
Story / Scene
For Nara, stories, worldbuilding, emotional truth, beauty, voice, and canon.
Audit
Evidence
For ZIP/log/code/result review. Attack claims and evidence, not the human seed.
Mental Engines
Engine chooses how the multi‑LLM brain thinks.
Engine
Purpose
Default use
E1 Flash Council
One independent multi‑LLM round + synthesis.
Quick breadth, cheap baseline.
E2 MAL Standard
Structured council → conflict/coverage → second round → synthesis.
Builder prompts, system upgrades, evidence audit.
E3 Roundtable
Multi‑round brainstorming where models hear the shared room memory.
Innovation, writers’ room, hard architecture.
E4 Synthesis Cascade
Roundtable → synthesis_1 → second roundtable on synthesis_1.
Critical final decisions only.
World application map
Where MAL can create unusually high value.
MentalArena is strongest when a problem needs several genuinely different viewpoints, the cost of a missed error is high, human tacit knowledge matters, and the result can be tested against sources, code, data, experiments, or downstream outcomes.
The full MAL loop:
human seed or frozen task → independent model attempts → disagreement and coverage map → a small high-information human question queue → second round → synthesis or build specification → implementation → external test → next improved iteration.
MAL value rises with:
• viewpoint diversity
• cost of missed error
• external verifiability
• human tacit knowledge
• traceability need
• parallelizable work
It falls with:
• extreme urgency
• privacy exposure
• strictly sequential work
• unverifiable outputs
Evidence status: the 0–100 numbers below are strategic fit estimates, not benchmark results. They state where MAL's architecture appears promising. Domain-specific blind tests must decide whether MAL actually beats a strong single-model baseline, a simple workflow, or a human expert process.
Engineering, AI, and digital systems
Software and AI-system development98 / 100
Best role: turn ambiguous requirements into buildable architecture, implementation gates, and evidence.
Legacy-system upgrade: add a major capability or migrate architecture without silently breaking accepted behavior, file formats, APIs, or release workflows.
Failed-build diagnosis: explain why code passes smoke tests but fails on real datasets, production traffic, edge cases, or performance limits.
Agent/product architecture: decide between a single model, RAG, tools, multi-agent orchestration, or deterministic code, then define the smallest decisive build and eval.
Best role: run blind, cost-aware comparisons and preserve dissent instead of choosing a model by reputation.
Model and workflow selection: choose the best model, RAG stack, or agent workflow for coding, support, analysis, or research under quality, latency, cost, and privacy constraints.
Orchestration ablation: test whether a multi-agent council actually beats one strong model when the seed, sources, builder, and evaluator are frozen.
AI-system risk audit: locate hallucination, prompt-injection, data-leakage, policy, calibration, and human-oversight failures before production promotion.
Recommended: Audit / Evidence · E2 or E4 for final release decisions
Complex technical R&D95 / 100
Best role: keep unconventional ideas alive long enough to convert them into competing mechanisms and kill-tests.
New machine or energy concept: split a speculative mechanism into variants, constraints, failure modes, simulations, and the smallest physical prototype.
Algorithmic invention: explore a large design space, preserve minority hypotheses, and define controls that distinguish a real algorithmic gain from benchmark overfit.
Recommended: Creative-Realistic / Innovation · E3
Technical audit and due diligence94 / 100
Best role: act as an evidence court across claims, source files, code, logs, and tests.
Startup technology claim: verify whether a claimed benchmark, compression gain, AI capability, or trading result survives matched controls and fresh data.
Vendor acceptance: compare contract requirements with the delivered repository, test reports, security posture, and missing evidence.
Research-package forensics: find leakage, lookahead, impossible accounting, open-tail dependence, cherry-picked runs, or absent negative results in a ZIP bundle.
Recommended: Audit / Evidence · E2
Safety engineering93 / 100
Best role: generate independent hazard views, merge them into a safety case, and demand testable mitigations.
FMEA and hazard analysis: identify coupled failures in a battery system, robot, industrial machine, medical device, or autonomous subsystem.
Safety-case construction: connect every safety claim to evidence, assumptions, owner, test, residual risk, and operating boundary.
Recovery design: compare graceful degradation, fail-safe shutdown, redundancy, alarm, and human-takeover strategies under cascading faults.
Recommended: Audit · E2; E3 for architecture redesign
Cybersecurity91 / 100
Best role: coordinate competing attack hypotheses and evidence without autonomously executing defensive actions.
Threat model: map attack paths through an AI-enabled SaaS product, model supply chain, RAG corpus, tools, identities, and third-party integrations.
Remediation priority: rank fixes by exploitability, blast radius, evidence strength, recovery cost, and dependency rather than by the loudest team opinion.
Recommended: Audit · E2 or E3; humans authorize all actions
Science, health, and evidence-intensive work
Scientific research96 / 100
Best role: generate competing hypotheses, explicit falsifiers, and experiments rather than one polished story.
Unexplained anomaly: produce mechanistically different explanations, list what each predicts, and design the cheapest experiment that separates them.
Replication planning: audit an influential result for hidden degrees of freedom, confounders, missing controls, and a preregistered replication design.
Research-program synthesis: turn a fragmented literature into a claim ledger, unresolved conflicts, prioritized hypotheses, and a sequence of empirical tests.
Recommended: Innovation / Factual · E3
Pharma and biotechnology90 / 100
Best role: support research design and evidence review under expert oversight.
Mechanism and assay design: compare biological mechanisms, biomarkers, assay choices, off-target risks, and experiments that could reject each route.
Contradictory preclinical evidence: reconcile cell, animal, omics, and literature findings while exposing where the datasets genuinely disagree.
Trial-protocol review: challenge endpoints, eligibility, stratification, safety monitoring, operational feasibility, and failure-to-learn risks before launch.
Best role: prepare a clinician-led review; never act as an autonomous diagnostician or prescriber.
Complex case conference: organize imaging, pathology, labs, history, guidelines, and specialist disagreements for a tumor board or rare-disease meeting.
Missing-information analysis: identify which additional test, history item, or specialist opinion would most change the current decision.
Guideline conflict: show why recommendations differ across patient subgroups, comorbidities, values, contraindications, and evidence quality.
Legal and standards mapping, control ownership, evidence packages, exceptions, remediation, and audit trail.
MAL Writers' Room
Stories and worlds
Independent creative lenses, dialogue, structure, canon, emotional arc, production realism, and human final vision.
Global rollout lens
European Union
Governance-heavy adoption
Strong fit for compliance, public-sector analysis, healthcare research, finance, industrial safety, and multilingual evidence trails.
North America and UK
Engineering and diligence
Strong fit for software, biotech, cybersecurity, venture/M&A technical review, AI evaluation, and professional services.
East Asia
Industrial systems
Strong fit for manufacturing quality, robotics, semiconductors, supply chains, technical R&D, and reliability engineering.
Emerging markets and international organizations
Affordable multidisciplinary council
Potentially valuable where small teams need broad expertise—provided local people are decision partners, not merely data sources.
Honest positioning: MAL is not the first multi-agent system. Its intended differentiation is the combination of a human seed, multiple providers, independent first attempts, preserved disagreement, selective human questions, builder handoff, external evidence gates, cost tracking, and iterative memory. The goal is not “more agents.” The goal is a better decision-and-evidence loop.
MAL Prize Frontier
Use real prize challenges as external judges.
After MAL-on-MAL, the Trading Arena ablation, and ARC‑AGI‑3, the strongest next move is not another internal demo. It is a standing frontier of externally defined problems with deadlines, judges, rules, and money at stake. The prize is useful; the independent evaluation is more valuable.
VERIFIED 19 JUN 2026
This is a live snapshot, not a permanent promise. Deadlines, eligibility, IP terms, and award amounts can change. Before committing work, MAL must re-open the official rules, confirm that entry is still accepted, and record a dated eligibility/IP checklist. Final submissions must remain original, human-reviewed, manually submitted, and compliant with each challenge's AI and automation rules.
Recommended sequence — with a continuous radar
Step 0
MAL improves MAL
Use v0.98 to diagnose and build the smallest credible v1.0, then lock the testable version.
Step 1
Trading A–F benchmark
Compare direct prompting, E1–E4, and manual gold reference using the same seed, builder, data, and evidence gates.
Step 2
ARC‑AGI‑3
Move from prompt quality to interactive agents that must explore, model, set goals, plan, act, and correct themselves.
Step 3
Prize Frontier
Run fast eligibility and kill-tests across software, forests, water, health, energy, logistics, and hardware.
Step 4
Cross-domain proof
Publish what transferred, what failed, cost per useful result, and which pieces should move into AI8 / Reasojjng.
Important scheduling correction: the roadmap is not fully sequential. The Prize Frontier should run continuously in lightweight triage mode while the core benchmarks continue. A good challenge should not be missed merely because its deadline arrives before ARC‑AGI‑3 is finished.
MAL opportunity score ≈ problem fit × eligibility × data access × prototype speed × evidence value × strategic learning × prize value ÷ (time × capital × partner risk × execution risk)
Priority lane — plausible near-term targets
ARC Prize 2026 — ARC‑AGI‑3
DirectMilestones 30 Jun / 30 SepFinal 2 Nov
$850K track $2M program
Problem: build an offline agent that enters novel interactive environments without instructions and learns what to do through efficient exploration and action.
MAL attack
Generate genuinely different agent architectures, preserve dissent about memory and search, build the smallest local simulator, then use ablations to determine which modules add score rather than narrative complexity.
First decisive test: can one local, reproducible agent beat random and hand-coded baselines on unseen developer environments under the same compute budget?
Problem: transform sensor or remote-sensing inputs into usable georeferenced surface-fuel and understory maps, with 3D outputs, validation, visualisation, and a Python ingestion tool.
MAL attack
Run a geospatial, forestry, computer-vision, uncertainty, and validation council; compare LiDAR, photogrammetry, multispectral, and public-data paths; then build one end-to-end reproducible pipeline rather than a broad concept paper.
First decisive test: can public or obtainable data yield one scientifically defensible 100 m × 100 m mapped site in an accepted output format? Rules admit adults from the U.S. or a NATO-affiliated ally, subject to verification.
Problem: propose a practical, scalable solution either for safe water and sanitation or for ocean health, ecosystem protection, sustainability, and resilience.
MAL attack
Start with a specific community and measurable failure, combine technical design with economics, equity, local adoption, and replication, then force the council to produce a pilot plan and falsifiable impact model.
First decisive test: can we identify a real local partner, baseline metric, and pilot that can start without winning the prize?
Problem: develop a validated, commercially credible solution for forests, biodiversity, sustainable materials, monitoring, soils, or nature data.
MAL attack
Use MAL as a virtual founding team: choose one narrow pain point, map users and buyers, design the minimum prototype, pressure-test scientific validity, and build a clear path from research to field adoption.
First decisive test: do we have or can we create a small validated demonstration and a credible early-stage team before September?
Partner lane — high-value challenges that need an eligible organisation or specialist build team
EU Prize for Governance Innovations in Energy Communities
Eligible energy-community partner25 Jun
€1,000,000
Problem: demonstrate inclusive, transparent, effective governance and fair benefit sharing in an existing EU renewable or citizen energy community.
MAL role
Rapidly audit governance, participation, financing, benefit sharing, local-policy alignment, and evidence; turn the best existing practice into a judge-aligned submission without inventing results.
Gate before any work: an already eligible energy community with up to 10,000 members must agree to apply immediately. Without that partner, skip this cycle.
HHS Digital Stockpile & Manufacturing Response Network
Manufacturing / logistics team28 Aug
Up to $2.04M
Problem: design a resilient system that can rapidly produce and distribute critical medical supplies during public-health emergencies and supply-chain disruption.
MAL role
Build independent architectures for digital inventory, distributed manufacturing, qualification, cyber-security, logistics, and emergency command; then stress them against disruption scenarios and convert the winner into an eight-page design, blueprint, and pitch.
Gate before any work: secure manufacturing, emergency-response, and U.S. eligibility expertise; this is not a credible solo paper exercise.
Aerospace hardware teamRegister 29 Jul · apply 12 Aug
Up to 3 × $500,000
Problem: create a payload that can be manipulated, reconfigured, or serviced by a robotic arm in low Earth orbit, then mature it to flight readiness.
MAL role
Run systems-engineering, robotics, payload, thermal, power, interfaces, reliability, FMEA, schedule, and test-plan lenses; MAL should narrow the concept before expensive hardware begins.
Gate before any work: find an eligible aerospace partner that can own payload fabrication, environmental testing, and flight readiness.
Problem: identify or develop scalable, evidence-based curricula that integrate nutrition into medical, residency, or nursing education.
MAL role
Map the competency framework, curriculum gaps, pedagogy, clinical integration, implementation barriers, outcome measures, and national scaling plan while preserving expert and learner disagreement.
Gate before any work: an eligible U.S.-based accredited medical, residency, or nursing institution must own the submission.
Portfolio rule: never choose a prize only because the purse is large. Keep one computational target, one real-world environmental or social target, and at most one partner-heavy moonshot. Every entry needs a no-go gate, a fixed budget, an evidence plan, and a reusable artifact even when it loses.
Human collaboration
AI-ONLY by default. HUMAN × AI when BD's intuition can unlock the search.
AI-ONLY
The council uses the initial BD seed and completes its workflow without a mid-run human question gate. This remains the clean control for ablation tests.
HUMAN × AI
Models first make their own best attempt, then may submit precise ASK_BD questions. MAL deduplicates them, pauses for a small high-information queue, records answers or skips, accepts unsolicited BD seeds, and carries that memory into later rounds.
Evidence boundary: human input is high-priority project intent, not proof. Contribution ledgers and an optional AI-only delta help attribution; tests, sources, arenas, and downstream results still decide.
Communication transport
TEXT by default. Structured JSON when useful.
TEXT
Natural-language council communication remains the default and is best for nuance, exploration, explanation, and artistic work.
JSON
Optional compact stage schemas improve inter-LLM structure, comparison, and local parsing. Valid sidecars feed later rounds; malformed JSON falls back to the preserved raw answer.
Human boundary: final synthesis is always normal readable text. JSON is an intermediate transport, not the public deliverable and not a guarantee of lower token cost.
Benchmark doctrine
Trading v1.4 is the proof test.
The benchmark freezes the initial prompt-builder seed, required source files, and final GPT‑5.5 Pro builder, then varies only the prompt-generation workflow.
Benchmark manifest records IPB/source/prompt/candidate hashes, anonymized candidate map, run order, builder model, and timestamps.
Cost/benefit ledger
Tracks LLM calls, token estimates, wall time, BD effort, prompt score, code score, evidence score, and score-per-cost.
No-promotion gates
Controls, event ledger, force-close-end PnL, open-tail report, and strategy fidelity are required before promotion.
External eval export
Exports promptfoo-like and candidate evaluation matrices for later comparison with standard eval tooling.
Public/private boundary
This public page contains no API-key UI.
Local-only tools such as provider API keys, provider status, chat logs, and local run controls belong only inside local_lab/AIM3_MentalArena_Lab.html served from localhost. They are intentionally absent from this public CRP page.
Context, comparison, and destination
MAL is a bridge—not the final architecture.
MAL should be judged against both simpler single-model workflows and the strongest external multi-agent research. It should also be understood correctly inside this project: MAL is an external reasoning laboratory and a stepping stone toward the proposed AI8 / Reasojjng architecture, not the endpoint itself.
MAL's strongest advantages and honest weaknesses
Where MAL can be stronger
Multi-provider diversity rather than one model family.
Independent first attempts before models influence one another.
Preserved dissent and an explicit test that can resolve it.
Human-seed amplification plus selective high-information questions.
Builder handoff, cost ledger, lineage, controls, and external evidence gates.
One workflow that can move from software to science, policy, nature, or art.
Where MAL is currently weaker
More latency, token cost, orchestration work, and failure surfaces than a strong single model.
Correlated models can repeat the same error and create false confidence.
Coordination can damage strictly sequential tasks.
Tool use, retrieval, long-horizon execution, and production reliability are less mature than leading deployed agent systems.
Its memory is external workflow memory, not yet integrated learning inside one persistent architecture.
MAL does not itself guarantee truth, safety, legal compliance, or useful real-world impact.
MAL compared with the wider landscape
Reference
What it has over MAL today
What MAL is trying to add
Lesson for us
One strong model / simple workflow
Lower cost, lower latency, clearer state, and often better coherence on sequential work.
Diversity, dissent, human questions, ablation arms, and an evidence court when one answer is not enough.
E1 or one model must remain the default null. Complexity has to earn promotion.
Google Research — scaling agent systems
Controlled evidence across 180 configurations and task-dependent scaling principles.
A practical multi-provider laboratory that routes modes, includes the human seed, and carries results into build-and-test loops.
Parallelisable problems favour councils; sequential problems may need one lead agent and a later audit.
Google AI Co-Scientist
Science-specific hypothesis generation, ranking, and research collaboration built around a powerful unified model ecosystem.
For science, MAL needs stronger literature provenance, experiment tooling, and domain-expert gates.
Anthropic multi-agent research system
Mature parallel web/tool research, production infrastructure, delegation, and context engineering.
Provider diversity, blind workflow comparison, contribution ledgers, preserved minority views, and human-seeded system invention.
We need better tool reliability, budget routing, recovery, and long-running execution—not simply more model calls.
NIST AI RMF
A cross-sector risk-management framework and common trustworthiness language.
An operational workflow that can turn risk questions into roles, evidence requests, tests, owners, and release gates.
MAL can help implement a risk process; it cannot certify itself as trustworthy.
EU AI Act
Binding legal obligations, governance, transparency, and enforcement in the EU.
Traceable decisions, source and claim ledgers, human gates, and compliance evidence packs.
Workflow traceability supports compliance but does not replace legal classification or counsel.
WHO AI-health guidance
Health-specific ethics, governance, human responsibility, and stakeholder protections.
Multidisciplinary deliberation, explicit missing evidence, and expert-question routing.
Clinical use must remain expert-led, validated, and bounded; MAL is decision support, not an autonomous clinician.
OECD AI in government
Cross-country evidence on public-sector adoption, enabling conditions, guardrails, and engagement.
A repeatable policy council that keeps stakeholder conflict, implementation, auditability, and local human input visible.
Public legitimacy and accountable authority cannot be delegated to an LLM council.
The bridge from MAL to AI8 / Reasojjng
MAL today
External collective cognition
Several LLMs, a human, source files, shared memory, builders, tests, and evidence are coordinated as separate components.
→
ARC‑AGI‑3 bridge
Interactive competence
The agent must actively explore, build a world model, infer or select goals, plan, execute, and course-correct under feedback.
→
AI8 / Reasojjng proposal
Integrated persistent architecture
Internalise specialised reasoning, DCC-like routing, durable memory, active perception, goal management, action, self-critique, skill accumulation, and a human value interface.
The intended direction: MAL is the scaffold on which we can discover which cognitive roles, conflict mechanisms, memory structures, evidence gates, and human interactions are actually useful. ARC‑AGI‑3 is especially important because it pressures the system beyond static answers into exploration, modeling, goal-setting, planning, execution, and correction. The long-term AI8 / Reasojjng claim remains a research proposal until an integrated system is built and independently tested.
Official context sources
These sources do not prove MAL's superiority. They define the external evidence, engineering, and governance landscape against which MAL should be measured.