AMOS Organism

AMOS Organism is the governed learning architecture intended to become the intelligence backbone of AMOS Hosted. Qwen (and later other compatible models) provides general cognition; the organism accumulates reusable procedures that survive independent verification.

This repository begins at the smallest defensible organism boundary:

• Energy is mission-scoped permission to spend compute and tools.
• Fitness is cross-mission, clawbackable credit that vests only after a host-attested consumption path reaches the official verifier.
• Strategy genes are immutable, content-addressed procedures with lineage.
• Reputation is a contextual view of vested fitness, not a currency.
• Novelty protects a small archive slot; it never grants energy or fitness.
• Trust and authority remain host-owned constraints outside the organism.
• HRR/semantic state guides attention but can never become evidence.

The model may propose work. It cannot mint receipts, fitness, genes, evidence, or authority.

The loop

flowchart LR
    M["Mission + exact host state"] --> H["Shared HRR attention"]
    H --> S["Specialists carrying strategy genes"]
    S --> E["Energy-bounded actions"]
    E --> R["Host receipts and artifacts"]
    R --> V["Independent verifier"]
    V --> C["Consumption-graph settlement"]
    C --> F["Vested fitness + gene lineage"]
    F --> N["Selection, mutation, recombination"]
    N --> M

An AMOS procedure has crossed the organism threshold when it vests on one mission, is reused on a new mission without changing model weights, and earns fitness again. A procedure that was merely produced—but never consumed—cannot buy its own survival.

What is implemented

• Mission energy allocation, reservation, spending, refund, and reset.
• Provisional fitness escrow, verifier-gated vesting, decay/clawback, and contextual reputation views.
• Host-attested causal graphs using conservative consumption/citation credit.
• Content-addressed strategy genes with mutation/recombination parents, rights/contamination tags, verified outcomes, and novelty retention.
• Context-keyed pheromones that preserve signal type instead of collapsing to a global attract/repel scalar.
• A dual-channel world-state boundary: exact host facts versus non-authoritative HRR attention and typed transition predictions.
• A value-of-computation policy based only on vested improvement, with quality lexicographically ahead of speed.
• A durable, append-only JSONL event chain plus a host-gated AMOS/AWS trace intake boundary. Failed or ineligible runs become zero-fitness negative experience; verified procedures remain candidates until separately approved.
• Deterministic replay of fitness escrow, settlement, regression punishment, candidate lineage/promotion, admitted genes, host-attested expressions, and verifier outcomes, so neither reward nor punishment disappears on restart.
• Versioned canonical contracts under contracts/ for candidates, expressions, trace bundles, and consent-gated Platform Mission episodes.
• Contextual gene selection and a host gene-expressed receipt. A model cannot self-report which procedure deserves credit.
• Research-only mutation/recombination proposals that run without approval; admission, external authority, promotion, and success claims remain host/verifier controlled.
• Lexicographic gene selection. Genuine verified failure produces negative guidance; successful but unconsumed work is recorded separately as uncredited and cannot poison a procedure.
• Three initial strategy genes extracted and separately admitted from verified, AMOS-owned Qwen swarm traces: bounded provider recovery, typed-tool recovery, and lossless context compaction.

See First Principles, Model Manifest, and Roadmap.

The swarm experiment

The Qwen swarm self-learning experiment that produces the organism's traces now lives in this repository under swarm/: the holographic swarm kernel, the organism metabolism simulator, the Harbor agents, the frozen benchmark fixtures and recorded results, the AWS research plane (Terraform, trainer, adapter verifier, runner), and the program documents under docs/swarm/. The swarm imports the kernel's contracts and digest; npm run check exercises both.

Development

Requires Node.js 22.18 or newer. Development and CI are pinned to Node 24; npm install fails closed on unsupported runtimes because Node 20 cannot execute the repository's type-stripped TypeScript tests.

npm install
npm run check        # kernel typecheck + swarm syntax check + all tests
npm run test:swarm   # swarm suites only

Research trace bundles can be settled without trusting model-authored status:

node scripts/importTraceBundle.ts INPUT.json OUTPUT.jsonl

Status

Research-stage kernel. The first metabolism loop now exists, including canonical consent-gated Platform Mission episode intake after host/KMS attestation, but it has not yet demonstrated repeatable quality improvement on frozen holdouts. It does not replace the external AMOS verifier, train adapters, or make production decisions. Those boundaries are deliberate.

License

Apache-2.0.