Trajectories (Ring B)

Trajectories capture how decisions unfold (also known as reasoning traces or context graphs, separate from the state clock).

Contracts

• Hash-chained steps: { step_seq, step_id, prev_step_hash } (similar to chain-of-thought traces).
• Step taxonomy: base (host-emitted) vs overlays (pack/derived/human).
• Refs carry provenance: observed | inferred | asserted.
• Outcomes share the same refs shape.

Storage

• JSONL append-only logs for steps and outcomes.
• SQLite sidecar indexes for listing, filtering, pagination.
• Deterministic ordering and ids.

Runtime capture

• TrajectoryContext with budgets and volume controls.
• Host dispatchCommand emits key steps: trajectory_started, entity_written, trajectory_ended.
• PackContext exposes trajectory access.
• Redaction/summarization helpers: summarize tool args/results, redact secrets/PII, truncate as needed.

Projections

• Co-occurrence graphs (rolling window, top-K) (a form of context graph for entity relationships).
• Transition n-grams.
• WL-style structural signatures.
• Outcome correlation tables.

See also: Simulation for how trajectories feed evidence-backed queries.