19 · The dynamic path — explore, steer, fan out

ADD 3.0 shipped one fixed route: every request became a contract, every contract a build, every build a gate. That is the right route for work you already understand — and the wrong first move for work you don't. 3.1 makes the path adaptive while the trust spine stays byte-identical: uncertainty gets a lane, steering gets a stamp, research gets a receipt. Nothing here weakens a floor; adaptivity changes the route, never what passes.

The patterns are field-imports, not inventions: the workflow/agent split and the research-loop anatomy (clarify → budget → query → read → reflect → refine → compress → stop) from the published agent-engineering literature, plan-and-execute replanning, and parallelize-reads-serialize-writes. What ADD adds is the part the field leaves informal: every dynamic move lands as a recorded, refusable artifact.

Route by uncertainty, not only by size

Intake (intake.md) always sized a request by scope and sensitivity. It now tallies a third axis first: the unknowns whose answer would change the contract's shape. Trivia and build detail don't count; a latent requirement you cannot yet state as a measurable target does.

• The closed floor is checked first and always wins — security · data · architecture size up, whatever the tally says.
• Among the lanes the floor allows, uncertainty dominates size: ONE contract-shaping unknown already argues Explore-first. Freezing a contract on a guess is how a plausible task ships the wrong thing with perfect receipts.
• The human still vetoes the route, as with every lane.

The Explore lane — when the answer is the deliverable

Some work's deliverable is an answer, not an edit: investigate a defect, evaluate a library, survey prior art. That work now has a first-class lane on the existing Task lifecycle — no new node type, no new verb:

add new Task <slug> --title "..." --kind explore

• Questions are the contract. Each Must in ## RULES is a question the explore must answer, stated so "answered" is judgeable. Questions freeze like contracts — rewriting a frozen question to fit the answer found is the same inversion as weakening a check.
• The budget is hard. ## PLAN carries one hard number (tool calls, sources, or wall-clock). The engine refuses to freeze a budgetless explore (R:UNBOUNDED). Overrun is a recorded re-freeze, never a silent continuation.
• One approval. The freeze seam is what will be asked and what it may cost — the same single human approval every task gets, at the same place.
• The loop is query → read → reflect → refine, stopped by sufficiency (every frozen question answered well enough to act on) or by the budget — both explicit.
• The deliverable is ## FINDINGS: a compressed, cited brief — one line per frozen question, F<n> (answers M<n>) · the finding · (evidence: <ref>). An uncited finding is an opinion, and the engine treats it as one.

The sufficiency gate is engine-enforced. A findings-only explore gates directly on its brief:

add gate <slug> PASS --by "<name>"

The gate refuses an unfrozen explore (R:UNFROZEN_EXPLORE — the questions+budget approval IS the lane's human seam), refuses when any frozen question lacks an evidence-carrying finding (R:HOLLOW_EXPLORE, naming the open ones), refuses post-freeze edits to the questions exactly as the drift tripwire does everywhere else, and records a passing verdict with evidence kind sources and the closed tally (closed: "3/3"). Security keeps its floor: a finding surfaced by research escalates to the human exactly as one surfaced by tests.

Downstream, the brief is a frozen fragment neighbors consume:

needs:
  - /tasks/<explore-slug>.md#findings

add brief compiles those findings into the next task's Direction prompt — assumptions arrive pre-discharged as evidence-carried facts. That is the lane's whole point: explore-first turns the next contract's unknowns into knowns before it freezes.

Micro-spikes — discharge an assumption instead of pricing it

Direction's ## ASSUMPTIONS always priced a guess (if wrong → consequence). When a guess is cheaply checkable, a micro-spike discharges it instead: run the two-minute probe, then record what was found on the assumption line — found: <what> with its evidence ref, on one physical line. A priced guess stays legitimate; a discharged one is simply better. An assumption that grows past a quick probe is the signal to route the question through the Explore lane instead.

Steering vs the contract — add replan

Builds discover things. A strategy that turned out wrong, a sequencing change, a constraint the plan didn't know — that is steering, and it now lands as a recorded, additive stamp:

add replan <slug> --note "pivoting to the sorted-merge approach"

The seal is untouched; the note joins the verified: trail as an act: replan entry, so the turn is visible in the record instead of silently absorbed. The engine refuses a replan on an unfrozen task (nothing is being steered) and on a done one (the trail is closed — that lesson belongs in add learn). The boundary is absolute: anything that would move a frozen gives:, a Must, or a check is a change-request — reopen and refreeze, never a replan. Steering steers the route; it never touches the seal (R:SEAL_TOUCH).

Read fan-out — facts merge, decisions serialize

Parallel work (chapter 08) serializes writes through waves and worktrees. Reads never needed that ceremony, and now the method says so: research streams, spec reads, and codebase surveys fan out freely — facts merge. The moment a stream would write — an edit, a stamp, a decision that binds a contract — it re-enters the serialized path (one write taints the stream). The floors are unchanged; fan-out widens what you can learn per hour, not who signs.

Planning wears a lens

Milestone drafts, intake proposals, and the loop's next-task proposal now load the best-fit persona whose flow: includes advisor before drafting starts, and the confirmed artifact records the lens (add advise <milestone> --persona <p>). The load is by fit and by roster — a bundle with no personas skips silently and behaves exactly as before. Chapter 10 covers the persona system; this is the one addition: planning is a loading surface too.

What never moved

The dynamic path is additive by construction, and the engine enforces the additivity:

floor	status under the dynamic path
security finding	HARD-STOP, everywhere — including findings surfaced by research
one human approval at the freeze	unchanged; the Explore lane's freeze IS that approval
receipt binding (covers: → evidence)	unchanged; findings bind questions the same way
frozen contracts move only by refreeze	unchanged; enforced on questions too
budgets (skill surface, task depth)	unchanged; explores add a hard research budget

A project that never uses --kind explore, never replans, and seeds no personas runs 3.1 exactly as it ran 3.0. The dynamic path is there for the day the work is uncertain — which is most days worth recording.