Polyrhythmic reasoning pattern

Prompt-first procedure: Describe the outcome you want in your agent conversation. The agent should select and load the appropriate AIWG assets, explain material changes, request any needed approval, and report verification evidence. Exact commands and flags appear only in the CLI reference.

The polyrhythmic-reasoning adapter turns the illustrative 4:4 and 5:4 prompt shape into an explicit FlowGraph pattern. It makes no claim that more beats are smarter, more accurate, or more efficient. Those are evaluation questions, not properties implied by the topology.

Both profiles use flow.aiwg.io/v1alpha1 and kind: FlowGraph. No fourth-level DNS API group is required.

Profiles

strict-lcm runs problem-mode every fourth activation and user-mode every fifth activation. Their first common activation is 20, when final synthesis is allowed. This profile exists to reproduce and test that exact schedule.

adaptive uses the same nodes, phases, state fields, bindings, and final output. It checks an evidence-backed typed convergence value after each paired activation and may synthesize early. The eight-activation ceiling remains independent and stops execution even if the evaluator never converges.

The periods are one pattern choice, not a default recommendation for other graphs.

Beat contract

Problem-mode uses the derive-evaluate-expand phase:

1. Derive one bounded candidate from the supplied question. 2. Evaluate it against supplied task evidence. 3. Expand only implications supported by that evidence.

User-mode uses classify-evaluate-expand:

1. Classify only context the user stated and attach the supporting evidence. 2. Preserve unknown when evidence is missing or ambiguous. 3. Evaluate presentation constraints and expand only supported adaptations.

Each beat emits a typed evidence record with track, phase, activation, iteration, status, evidence, and conclusion. status: executed means the beat ran. Adapters may record status: skipped only when the manifest's optional/failure policy authorizes the skip; the reason belongs in the event trace. A failed beat follows the graph's partial-synthesis policy and is never silently treated as successful.

Evidence arrays are appended in manifest order. The convergence score uses a maximum reducer. Conflicting conclusions are preserved and final synthesis returns conflicting-track-results rather than silently selecting one.

Public output and trace boundary

Final synthesis returns:

• answer — the user-facing response;
• decisionSummary — a concise compatibility or conflict outcome;
• evidenceSummary — beat counts and conflict status; and
• userState — an explicitly supported state or unknown.

Both profiles are final-only. Intermediate drafts are not user output. Metadata traces contain execution identities, status, digests, and resources—not private chain-of-thought. Classification evidence is redacted if full-I/O tracing is enabled.

Safe examples

The shipped examples cover technical troubleshooting, conceptual explanation, practical planning, and theoretical comparison. They include agent-only and agent-plus-read-only-tool forms. The latter grants only filesystem:read and uses sideEffectMode: none.

The adapter rejects general medical/healthcare diagnosis or treatment, investment/financial advice, and any scenario marked high risk. Those domains require separate governance and are not generic prompt templates. It also rejects unsupported claims that the pattern is better, smarter, more accurate, or more efficient.

Programmatic adapter

~~~javascript import { buildPolyrhythmicGraph } from './lib/polyrhythmic-reasoning.mjs';

const graph = buildPolyrhythmicGraph({ id: 'explain-idempotency', domain: 'conceptual-explanation', profile: 'adaptive', composition: 'agent-plus-read-only-tool', question: 'Explain idempotent versus exactly-once execution.', evidence: [{ source: 'local-document', path: 'docs/runtime-operations.md' }], classificationEvidence: [], statedUserState: 'unknown', }); ~~~

The builder validates the scenario, narrows read-only tool authority, produces one strict FlowGraph, and validates the generated graph before returning it. Use the normal the agent-owned composition operation run adapter boundary to execute it.