What it owns
- Demand-to-capability retrieval
- Port/dependency/namespace compatibility
- Packaging and runtime resolution
- Assembly graph and glue budget
- UNDERDETERMINED questions
P12 · Assembly knowledge node
Deterministic composition planner
Retrieve candidates, eliminate incompatible sets mechanically, then let AI choose among feasible plans and write only bounded glue.
Problem boundary
What this subsystem owns and what it must hand to the rest of Actionist.
What research already found
- Phase 8 selected plan-then-fill over monolithic generation
- A model-free solver and 26 fixtures were designed
- Blocks can be fitted without AI once contracts and bindings are explicit
Open first-principles questions
- What is the optimization objective?
- When should a whole service beat several blocks?
- Which fields are necessary for deterministic compatibility?
- How should optional recommendations be presented to the client?
Inputs
- ProductSpec
- Capability shelf
- Archetype
- Host contracts
- Taste profile
Outputs
- AssemblyPlan
- CompatibilityProof
- ClarifyingQuestions
Connected knowledge nodes
P02 · Outcome and product specificationP03 · Curated capability shelfP05 · Living UI component layerP08 · Archetypes, shells and layoutsP09 · Data and state planeP10 · Identity, settings and navigation hostP11 · Connector and external-action planeP13 · Preview, editor and change loopP14 · Runtime, verification and release
Scientific research protocol
Every research owner uses the same evidence funnel. No padding is allowed when a denominator is inapplicable or evidence is unavailable.
Prior evidence auditRead the canonical spine, all domain-linked artifacts, the 17-industry catalogue and existing contradictions before searching again.
Commercial 100 → 10Define a denominator, map approximately 100 companies/products, and produce first-party dossiers for the strongest 10.
GitHub 100 → 10Search and dedupe approximately 100 relevant OSS projects, then deeply inspect the strongest 10 at the capability and architecture level.
Local estate joinInspect AutoSaaS, Great Library, SISOCRM, 21st stores and other domain-specific precedents already on the laptop.
Innovation 100 → 10Generate a wide divergent idea register, score it against the problem and retain the 10 strongest falsifiable innovations.
First-principles synthesisDecompose the irreducible problem, challenge inherited assumptions, reconcile contradictions and state the minimum sufficient design.
Contracts and experimentsReturn the proposed contract, dependency edges, unknowns, falsifiers, experiments, stop rules and decisions.
Independent verificationA separate verifier checks counts, links, source claims, overlap, contradictions and unsupported conclusions before promotion.
Evidence and agent outputs
Existing research remains source evidence. New agent packets are written to one owned workstream directory and linked here after verification.
| Artifact | Location | State |
|---|---|---|
| Composition agent architecture | research/actionmodel-builder-research-2026-08-26/phase-8/lanes/05-composition-agent-evals/outputs/composition-agent-architecture.md | existing |
| Composition eval suite | research/actionmodel-builder-research-2026-08-26/phase-8/lanes/05-composition-agent-evals/outputs/composition-eval-suite.json | existing |
| Synthetic pilot plan | research/actionmodel-builder-research-2026-08-26/phase-8/lanes/05-composition-agent-evals/outputs/synthetic-pilot-plan.md | existing |
| Future verified agent packet | research/workstreams/P12-deterministic-composition-planner/runs/<run-id>/ | planned |