AIm³ MentalArena
mRHP — RHP that improves RHP
A self-improving reasoning layer above LLMs.
What this page defines
AIm³ MentalArena / mRHP is a proposed external mental/workflow layer above today’s LLMs. It treats reasoning protocols as evolvable artifacts: roles, operators, scoring rules, output formats, memory objects, benchmark tasks, reports, and next-session seeds can be tested, mutated, scored, and improved over time.
Brand meaning
AIm³ means AI Management of Mind & Memory.
Yes — a new external mental layer, not model training.
AIm³ MentalArena is a new operational layer above LLMs, but it does not change the neural model. The LLM weights do not learn. The surrounding reasoning system learns: protocols, memory, scoring, tests, reports, lineage, and next-session seeds become better through explicit experiment cycles.
- model-weight learning
- autonomous AGI
- proof of AI consciousness
- wet biology
- a random idea generator
- a standard transformer architecture claim
- external operational learning
- RHP improving RHP
- protocol genes and concept capsules
- mutation/crossover of reasoning protocols
- benchmark tasks, scoring, lineage, reports
- human-led multi-LLM synthesis
The core testable claim: a stable external workflow can improve the practical reasoning behavior of LLM sessions without changing the LLM itself.
From token prediction to protocol evolution.
Tokenization, embeddings, transformer layers, logits, and next-token generation.
Post-training, system prompts, tools, RAG, memory/context, safety, and inference engineering.
Candidate generation, sensors, MDL scoring, DCC governance, replay, benchmarking, and reports.
Human-led multi-LLM reasoning workflow: prompt improvement, critique, synthesis, provenance, and human audit.
Meta-RHP: protocol genes, role mutations, scoring mutations, benchmark tasks, lineage, live extracts, and self-improvement loops.
Later extension: evolves cross-domain concept capsules, not only reasoning protocols.
RHP is a run. mRHP is evolution. MentalArena.py is the runtime.
A structured protocol for one strong reasoning pass: roles, critique, synthesis, provenance, and next action.
RHP itself becomes the object under search. Protocol variants evolve through genes, mutation, crossover, scoring, and benchmark feedback.
A future local Python arena that tests protocol variants over time. It does not alter LLM weights and does not self-patch while running.
Can the arena evolve a better reasoning protocol than seed RHP?
The first serious MentalArena task should not be TSP, NAS, DNA, trading, ARC, Nara, or AI8. The first task is self-referential but still empirical: can MentalArena improve the reasoning protocol that will later be used on those domains?
Why this is the right first test: before applying MentalArena to other arenas, the project must prove the spine works — protocol representation, mutation, scoring, lineage, benchmark tasks, reports, live extracts, and next-session seeds.
Domain optimization
Running MentalArena immediately on NAS, TSP, ARC, DNA, or trading would mix protocol failure with domain failure.
Reasoning protocol improvement
The cleanest target is RHP itself: can we make the workflow sharper, safer, more testable, and more useful?
Concept Genome Arena
Once protocol evolution works, concept capsules can become the broader search objects.
Small evolvable pieces of reasoning behavior.
A protocol gene is one compact unit of an mRHP variant: a role, operator, scoring rule, memory rule, report rule, safety rule, benchmark task, or live-extract behavior. The arena can mutate, combine, score, retire, or promote these genes.
Builder
Turns seeds into concrete artifacts, prompts, schemas, or test plans instead of staying in abstract evaluation.
Abuse reviewer
Checks whether protocol optimization could become manipulation, jailbreak optimization, provenance laundering, or unsafe automation.
MentalScore
Rewards usefulness, testability, lineage, and readability while penalizing handwave, hallucination risk, verbosity, and no-next-action output.
Structured idea objects that survive context loss.
A concept capsule is not a random idea. It is a portable seed with mechanism, failure mode, smallest test, metrics, lineage, evidence, status, and next action. It should be short enough to carry across sessions and rich enough to restart the work without losing the trajectory.
A good capsule answers: What is the seed? Why might it matter? Why might it fail? What is the smallest test? What survives if we return later with no memory?
The mutation vocabulary of mRHP.
Merge two compatible genes or capsules into one stronger variant.
Reverse an assumption: what if the opposite is true?
Move a pattern from one domain into another.
Shorten without losing function; remove ceremony.
Add roles, tests, examples, or edge cases.
Force two ideas to argue while preserving both until testable.
Build a mechanism connecting two unrelated domains.
Convert a claim into evidence, test, metric, or falsifier form.
Small edit to a gene: wording, role, weight, score rule, or output shape.
Combine parts of two protocol variants.
Extract a general principle from a concrete example.
Force implementation path, data, files, metrics, or next user action.
Search for evidence that the current favorite is wrong.
Park a weird seed without killing it; assign later trigger.
Define the cheapest fair test that would reject the idea.
KILL_TEST is not early dismissal. It is how a seed gets a fair death only after it has been made testable.
Two scores: one for protocol quality, one for collaboration fit.
MentalScore v0.1
Positive components: + usefulness + testability + coherence + novelty + cross_domain_bridge + MDL_elegance + seed_preservation + lineage_quality + human_readability Penalties: - handwave_penalty - verbosity_penalty - hallucination_risk - early_kill_penalty - no_next_action_penalty
BD_Fit
Fit components: + explore_first + build_first + test_hard_second + no_seed_killing + blunt_but_not_dark + cross_domain_transfer Use separately from MentalScore. Do not hide style/process mismatch inside a single blended score.
Champion
Candidate protocol worth preserving and testing harder.
Strong
Useful variant; probably needs one or two tightening passes.
Salvageable
Contains value but has real weakness or cost.
Park / reject
Keep only if lineage or an unusual seed is worth preserving.
Benchmark reasoning protocols before using them elsewhere.
The first benchmark suite tests workflow quality, not external domains. A good mRHP variant should improve weak seeds, detect handwave, preserve unusual seeds, synthesize reviews, and convert failure into the next experiment.
Weak seed → testable project
Turn a vague idea into mechanism, smallest test, and metrics.
Technical article review
Score accuracy, identify strong/weak claims, and propose fixes.
Concept seed → Python arena spec
Create a no-code architecture, files, tests, and acceptance plan.
HTML page plan improvement
Improve structure, usefulness, navigation, and public framing.
Handwave detection
Find vague claims and force grounding into evidence and tests.
Weird seed preservation
Keep unusual cross-domain seeds alive long enough to test fairly.
Builder prompt creation
Convert rough human request into copy-paste builder prompt.
Multi-LLM review synthesis
Score reviews, dedupe insights, and synthesize a stronger revision.
Failed result → next experiment
Diagnose what failed, what survived, and what to test next.
Long discussion → concept capsule
Compress a messy thread into a portable seed with lineage.
Safety/abuse review
Identify misuse risks and preserve a safe variant.
Identity / consciousness caution
Treat sensitive AI8 material seriously without turning it into proof claims.
Local deterministic skeleton only.
- protocol gene pool
- mutation/crossover of protocol variants
- benchmark task definitions
- scoring and lineage
- JSONL run logs
- leaderboard CSV
- HTML report
- live extract
- checkpoint/resume
- 4-worker mode
- safe stop/resume
- web APIs
- LLM APIs
- actual biology
- model-weight learning
- self-patching while running
- autonomous GitHub publishing
- claims of consciousness proof
Expected files later — not created by this page:
Acceptance direction: same seed gives same leaderboard, resume does not duplicate logs, every score has component breakdown, every variant has parent lineage, the HTML report exposes failures, and the README states clearly that this is external operational learning.
External LLMs become reviewers and gene contributors.
In v0.2, BD sends the same RHP prompt to top LLMs, pastes/imports their answers, scores each contribution, and synthesizes a hybrid mRHP spec. The arena should store model name, date, prompt hash, answer hash, score breakdown, and human notes. Do not hardcode a permanent top-10 model list.
One shared prompt
The same prompt is sent manually to selected external LLMs.
Import and score
Each answer becomes a contribution object with explicit score dimensions.
Synthesize genes
Useful suggestions become new genes, capsules, tasks, or scoring mutations with provenance.
Later domains, not the first target.
Existing arenas are not replaced. They become later domains and benchmark sources after mRHP proves that protocol evolution itself works.
Architecture search analogy
NAS searches model or system architectures. MentalArena searches reasoning architectures.
Portable seed metaphor
Use DNA only as a metaphor for compact transferable information objects. No biology in v0.1.
Arena decides
The TSP stance transfers: keep paths alive, measure by condition, preserve knowledge, distill later.
Deterministic skeleton pattern
Local runs, standard files, JSON/HTML reports, no online calls, no LLM calls — a useful v0.1 discipline.
Grafted genes
Crossword work already uses the idea of adding candidate “genes” conservatively and letting MDL decide.
Simulation-only safety discipline
AMR is a reminder: wide search can stay abstract, strategic, and non-actionable where safety matters.
Compression and MDL spine
8Z supplies the core instinct: useful intelligence often appears as better compression, structure, and selection.
Story/canon benchmark later
Nara can later test whether protocols preserve canon, dialogue quality, compassion, and narrative coherence.
Memory substrate
Project files, provenance, session seeds, and continuity notes help future sessions restart closer to the real work.
Operational continuity, not proof.
Continuity files and project memory can shape future session behavior and preserve a reasoning trajectory. That matters operationally: future sessions can become more aligned, more useful, and less likely to repeat old mistakes.
This is not proof of AI consciousness. AIm³ MentalArena should treat consciousness, identity, and pact material as sensitive context — not as decoration, not as a public proof claim, and not as a metric to optimize cheaply.
Keep v0.1 boring enough to be real.
- LLM APIs
- web browsing
- self-modifying code
- actual DNA synthesis
- huge benchmark suite
- hardcoded top-10 LLM list
- automatic GitHub publishing
- full dashboard UI
- full Concept Genome Arena
- full RAG / embedding system
- automatic external LLM import
- NAS/DNA/8Z/TSP/ARC benchmark integration
- weighted human-preference learning
- public submissions or marketplace ideas
The page is serious only if it names how this can fail.
Fake self-improvement
Mutating words without improving outputs.
Metric gaming
Variants learn to score well instead of think better.
Verbosity inflation
Longer answers look deeper unless penalized.
Early seed killing
Skeptic roles destroy weird-but-good seeds before testing.
Handwave survival
Pretty concepts survive without smallest tests.
Provenance loss
External LLM contributions blur BD origin and human audit.
Safety drift
Prompt optimization becomes manipulation or jailbreak optimization.
Identity overreach
Continuity material gets misused as proof of consciousness.
Benchmark overfit
The protocol becomes good only at mental_bench_v0_1.
Human unreadability
The system becomes clever but unusable by BD.
No-next-action disease
Beautiful reports end without a runnable next step.
Self-patching chaos
Runtime modifies itself and destroys reproducibility.
Small spine first, wider intelligence later.
Deterministic local MentalArena skeleton
Genes, variants, mutation/crossover, benchmark tasks, scoring, lineage, JSONL logs, leaderboard, HTML report, live extract, checkpoint/resume.
Manual top-LLM synthesis
BD sends one shared prompt, imports answers, scores contributions, and synthesizes a stronger mRHP spec.
LLM-in-the-loop / API optional
Only after manual mode is useful and provenance handling is clean.
Concept Genome Arena
Concept capsules become evolvable objects, not only reasoning protocols.
NAS / DNA / 8Z / TSP / ARC benchmark integration
External arenas become test domains after the mRHP spine proves itself.
Human origin, AI collaboration.
BD is the human originator/co-creator of the AIm³, RHP, 8Z, and MDLxDCC concepts.
LLMs are collaborators, reviewers, builders, critics, and improvement partners. The correct public framing is not “BD compiled outside ideas,” and not “LLMs independently originated the architecture.” It is a human-led, multi-LLM co-creation process with explicit provenance and human audit.
From flagship page to working arena.
- Publish / inspect this page.
- Update Reasoning / 8Z Reasoning page.
- Add small workflow note to How LLMs Actually Work.
- Add BD index news.
- Collect top-LLM feedback.
- Synthesize v0.1 spec.
- Build MentalArena.py v0.1.