Research program governance

Automo owns mutable scientific research state under .automo/. GetDone, when used for project delivery and software development, owns project direction, roadmap/milestones, and .agent/. Automo must not duplicate those milestones as scientific state.

A fresh automo init creates the scientific skeleton:

text .automo/ ├── project.yaml └── research-program/ ├── program.yaml ├── model-graph.yaml ├── hypotheses/ ├── candidates/ └── composition-experiments/

Registry, research-run, pool, refresh, and capability state also live below .automo/ as they are created.

Two separate graphs

Automo deliberately separates system structure from scientific claims.

The model graph describes computation/information structure. Models may be independent, composed as submodel → meta-model inputs, or related only by correlation/complementarity.

The hypothesis graph describes falsifiable scientific claims. A hypothesis may be scoped to one primary model, reference related models, or test cross-model/system composition. It is not required to mirror the model graph.

Example:

bash automo research model-add ranking --role submodel automo research model-add market --role submodel automo research model-add meta --role meta automo research model-relation-add ranking meta --kind input automo research model-relation-add market meta --kind input

A flat informational relation can be represented without creating an execution dependency:

bash automo research model-relation-add ranking market --kind correlated

Model candidates and exact composition

A ModelComponent is the logical role in the system; a model candidate is one immutable researched version of that component. Candidate records pin the model_spec_id, point at the runtime/registry artifact, and, for composed models, pin the exact upstream candidate versions used to construct it.

```bash automo research candidate-add RANK-v1 \ --model ranking \ --model-spec-id ranking-v1 \ --artifact-id MODEL-RANK-v1

automo research candidate-add MARKET-v2 \ --model market \ --model-spec-id market-v2 \ --artifact-id MODEL-MARKET-v2

automo research candidate-add META-market-v3 \ --model meta \ --model-spec-id meta-linear \ --artifact-id MODEL-META-market-v3 \ --input market:MARKET-v2

automo research candidate-add META-market-rank-v3 \ --model meta \ --model-spec-id meta-linear \ --artifact-id MODEL-META-market-rank-v3 \ --input ranking:RANK-v1 \ --input market:MARKET-v2

automo research candidate-select META-market-rank-v3 ```

Selection is per logical model. Selecting a new candidate replaces the previous selected candidate for that model but does not delete historical candidates or evidence.

Composition and ablation experiments

A composition experiment asks whether a different composition or target-model specification adds value while holding the committed comparison dimensions explicit. Automo rejects experiments that change both dimensions at once. This supports two common designs:

1. Incremental-input ablation: compare two immutable target-model candidates that differ only by one pinned upstream input.
2. Composition-algorithm comparison: compare two target-model candidates that pin the same upstream inputs but use different target-model artifacts/implementations.

For an input ablation:

```bash automo research hypothesis-create H-RANK-INCREMENTAL \ --statement "Ranking adds incremental information beyond market." \ --primary-model meta \ --related-model ranking \ --objective log_loss:parent:minimize \ --evaluation-depth parent

automo research composition-create EXP-RANK-ABLATION \ --hypothesis H-RANK-INCREMENTAL \ --target-model meta \ --control-candidate META-market-v3 \ --treatment-candidate META-market-rank-v3 \ --metric log_loss \ --metric brier \ --rationale "Isolate ranking's incremental contribution." \ --expected-effect "Improve OOS probability quality without calibration degradation." \ --falsification "No committed metric improves materially." ```

Automo validates candidate composition at candidate-registration time: every pinned input candidate must belong to the named upstream model, and an input relation must exist from that upstream model to the target model. A composition experiment then compares two distinct immutable candidates of the same target model. Correlated/complementary/alternative relations describe information structure but do not authorize runtime composition.

The model graph and candidate graph therefore answer different questions: the model graph says which logical components may interact; candidate records say exactly which researched versions were used; composition experiments say which controlled comparison produced the evidence.

Hypothesis lifecycle

Create a root claim and child claims as needed:

```bash automo research hypothesis-create H-ROOT \ --statement "The system provides useful out-of-sample predictive value." \ --primary-model meta \ --evaluation-depth system

automo research hypothesis-create H-RANK \ --parent H-ROOT \ --statement "Ranking provides incremental predictive information." \ --primary-model ranking \ --related-model meta \ --evaluation-depth parent ```

Activate the immediate scientific context before governed automated research:

bash automo research hypothesis-activate H-RANK

After evidence is sufficient:

bash automo research hypothesis-conclude H-RANK \ --status supported \ --conclusion "Parent-level evaluation shows incremental value."

Valid conclusion states are supported, rejected, and inconclusive. A local improvement is not sufficient when the hypothesis requires parent or system evidence.

GetDone boundary

GetDone/project roadmap material answers what direction matters next. Automo answers what scientific claims are unresolved and what evidence is needed. Automo may consume project direction as prioritization context but does not persist milestone lifecycle state.

When research requires missing software capability, Automo persists a bounded capability request under .automo/capabilities/requests/ with program/hypothesis/experiment provenance. Optional GetDone delegation may implement that capability, using .agent/ for development workflow state. Automo protects research evidence from delegated mutation and independently validates the returned implementation result before research resumes.