Research Routes Plan

Create a route-neutral semantic context before choosing an implementation or experiment route. Compile it as research-pack/v1 when a durable Pack artifact is needed.

1. Classify the context as software, research-exploration, or hybrid and preserve the exact starting point.
2. Create a question/problem/frame/shared-backbone record. Do not invent a claim when the input only supplies a question.
3. Add one or more concurrent routes and branches. A route is a navigable hypothesis space, not a commitment.
4. Represent claims, hypotheses, probes, experiments, evidence, observations, unknowns, decisions, implementation, canonical state, continuation, and provenance as distinct nodes.
5. Label every evidence item with one of support, counter, null, boundary, conflict, failure, context, mixed, or unknown; label every relation with source-stated, deterministic-derived, candidate-inferred, or human-confirmed.
6. Preserve lifecycle events: Create, Fork, Park, Reopen, Revive, Rewind, Reject, Supersede, Historical Context, and Unresolved. Forks receive new IDs and rewinds preserve history.
7. Mark CounterPilot work as a read-only boundary. It may challenge scope or assumptions but cannot mutate route state.
8. Never present AI inference as fact. Canonical promotion and current continuation require an explicit HumanDecision with actor and provenance.
9. Keep FailureRecord polarity, RewindProposal, Lesson, ReopenedProblem, canonical downgrade, and Human route authorization in separate Pack records. A failure may narrow a regime without invalidating every route.

Use the shared schema at shared/schema/research-routes.schema.json, semantic runtime in shared/core/model.py, and read-only validator shared/router/validate_context.py. The companion reference explains the node and lifecycle vocabulary.