DAWM Phase 9: cost-aware learned routing and budget-gated rollout

DAWM Phase 9 — cost-aware learned routing and budget-gated rollout

Three sub-phases take the DAWM cursor from a deterministic graph-traversal probe to a learned, cost-aware routing policy and finally to a budget-gated multi-step rollout controller. Net result: four findings, of which one is a positive mechanism and three are honest boundaries.

1. Deterministic cursor (preliminary probe)

Phase 9 asks whether rendering-invariant semantic state (established in the prior phase) can be executed as an activation-graph-plus-cursor trajectory rather than dense or stateless re-evaluation. The first probe builds a small controlled cell — semantic fact nodes, entity-continuation activation edges, a known short multi-hop gold cursor path, a distractor branch from the same start node, and cross-turn / cross-rendering support facts — and compares five systems: the activation-graph cursor, a dense persistent scan, a dense stateless scan, a budget-starved cursor, and a wrong-route control.

At a high level: the cursor matches dense-persistent answer quality with roughly half the activated nodes and under half the edge inspections; a stateless current-turn scan cannot solve split evidence; under insufficient budget the cursor abstains rather than hallucinates; and a wrong route is visible as an unsupported answer plus path mismatch. The result was confirmed on a separate multi-seed run with an accompanying passing test suite. It is explicitly framed as a mechanism sketch / diagnostic seed, not a closure — it does not claim learned routing, graph scaling, or hardware acceleration.

2. Routing economics (closed)

The second sub-phase fine-tunes the learned routing policy (inherited from the prior cursor) with three lightweight economic heads, adding only a small number of new parameters on top of the base policy:

• cost head: cursor state -> predicted remaining-cost-to-goal (scalar)
• early-stop head: cursor state -> stop logit (scalar)
• provenance-weighting head: candidate -> soft reliability weight in (0,1)

The cursor reranks each candidate by multiplying its candidate score by the candidate's provenance weight, shifting the decision boundary toward provenance-high (genuinely supporting) candidates when raw scores are close. Fine-tuning from the base checkpoint converged quickly per seed.

A multi-seed gate evaluation reported, at a high level: the activation-reduction hypothesis sat at ceiling (not applicable, see boundary below); multi-candidate gold-pick passed; trace-equivalence passed; and the tight-budget hallucination rate stayed at or below its target.

The cost-regression loss reached zero across seeds — the head reliably predicts cost-to-goal, the first concrete cost-aware behavior in DAWM: the cursor knows how much work remains. The economic layer adds new decision logic without breaking the trace-equivalence-by-construction property (unconditionally), and budget-bounded abstention stays hallucination-free.

Positive mechanism — a learned cursor-routing system extended with a cost-estimation head, an early-stop head, and a per-candidate provenance-weighting head, trained by supervised fine-tuning of a base routing policy while preserving trace-equivalence-by-construction. The cursor can now answer: why move, why stop, why expand, why abstain.

Boundary (activation-reduction utility ceiling) — the economic layer reduces activation only where the base policy over-activates above the structural minimum. Here the base already activates the minimum (the structural floor for a short multi-hop path), so the activation-reduction hypothesis sits at the floor with no headroom. This is structurally identical to the earlier delayed-commit ceiling: economic optimization is conditional on the baseline being above the structural floor. The two boundaries jointly form an economic-optimization operating-condition envelope.

3. Multi-step rollout + budget gating (partial closure)

The third sub-phase composes the frozen economic heads (no fine-tuning — kept uncontaminated by design) into a multi-step rollout controller with a budget gate and a provenance top-k register. The controller pre-empts (abstains) before activating any node when predicted remaining cost exceeds remaining budget, and — for the first time — wires the early-stop head as a live control signal (previously it was read but only no-op-passed; the cursor always walked the full path).

A multi-seed gate evaluation reported, at a high level: budget compliance passed on all seeds; early-stop precision was unstable (passing in only a minority of seeds); provenance stability passed on all seeds; an anti-destructive-reduction hypothesis passed on all seeds (under a redefinition described below); and the composite gate failed because it inherits the unstable early-stop axis.

Verdict: honest partial closure. Core composition and anti-destructive claims pass cleanly across all seeds; the early-stop-as-controller axis does not close.

The anti-destructive-reduction hypothesis was redefined to test reduction under early-stop-disabled path-walk mode, isolating wrapper fidelity (does the rollout loop faithfully reproduce the economic layer's decisions?) from the orthogonal question of whether the early-stop head is a trained detector. This is a scoping correction, not a loosening.

Negative controls: two of the three negative controls collapsed as required (the targeted gates are not trivially satisfiable; one collapse confirms the early-stop measurement is sound). A third negative control — witness-shuffle on the provenance register — did not collapse: provenance stability survived a corrupted witness mapping, indicating the property is broader / more robust than the control assumed, which arguably strengthens the provenance-stability finding. Reported honestly rather than buried.

Boundary (budget-tight economic pre-emption) — the rollout controller declines to start a path it predicts it cannot finish within budget; under a tight budget this predict-and-preempt behavior diverges from a single-step cursor (which would still activate one node). This is new economic behavior, not a reduction failure, and is fully observable. It pairs with the positive mechanism above as an applied use of the cost head.

Boundary (early-stop head is not a trained mid-rollout goal detector) — because it was a no-op during the training/eval loop that produced it, it received no supervision toward a step-wise stop decision; when first activated as a live signal it fires trivially at the first step and its precision is seed-unstable (passing in only a minority of seeds). Enablement lesson: an economic policy's auxiliary heads require active-role supervision to function as downstream control signals. (An incidental observation — that first-step firing sometimes commits the correct gold answer, a possible commit-leakage artifact — is flagged for separate future work, not analyzed here.)

4. Net status

State formation, scoping, rendering invariance, and state-trajectory execution with economic routing are all positively validated at controlled-cell level. Phase 9 contributes one positive mechanism and three boundaries. Loss weights, budget targets, ratio thresholds, register capacity, and the focal parameter are held out of all claim language and described only generically (a configured loss-weight balance, a configured activation budget, a configured register capacity).