DAWM Phase 10: controlled small-cell integration coherence (preliminary)

Status: Published (review artifact). Evidence level: experimental. Honest framing: This documents a preliminary integration smoke harness, NOT a completed Phase 10 closure. An earlier internal draft overstated this as a controlled closure; that framing was corrected. Correct status: a controlled-integration contract was drafted, an initial smoke-harness pass succeeded, and full Phase 10 closure has NOT been reached.

1. What this is

Phase 10 asks whether the controlled small-cell primitives developed across earlier phases can coexist in a single integration pass without any one of them regressing the observability of the others. The harness runs the controlled probes for the predecessor phases together and emits a single per-component attribution summary.

Integration chain under test (high level):

• Factorized semantic-state formation
• Scoped semantic candidate control (a learned two-stage dynamic scope expansion)
• Rendering-invariant semantic state plus cross-rendering contradiction localization
• Persistent activation-graph cursor execution with a recoverable cursor path

2. What was built

A controlled-integration design contract, an integration-summary routine, and an accompanying test suite. The integration summary executes, in one combined run, the learned two-stage dynamic scope expansion, the controlled rendering-state convergence with contradiction localization, and the activation-graph cursor execution, then reports each component's metric in a single table.

3. Evidence (environment A)

A single combined invocation over a controlled small-cell domain produced an overall pass. Qualitatively:

• The learned scope expansion showed a positive scope-overlap delta and a reduced candidate ratio relative to the flat baseline.
• Rendering-state convergence and contradiction localization were fully satisfied on the controlled cell.
• Cursor answer accuracy and cursor-path exact match were fully satisfied; the cursor touched roughly half the nodes of the dense baseline (an efficiency ratio of about 2x against dense traversal); the unsupported-answer rate was zero.

All eight per-component checks passed together: scope-overlap delta, candidate ratio, rendering-state convergence, contradiction localization, cursor accuracy, cursor-path match, cursor efficiency, and unsupported-answer suppression.

4. Cross-environment verification

The same harness was re-run on a second, independent computing environment with a different seed set, again yielding an overall pass with the same qualitative pattern: a positive learned scope-overlap delta, a reduced candidate ratio, fully satisfied rendering convergence and contradiction localization, full cursor accuracy and path match, an approximately 2x cursor-efficiency ratio against dense traversal, and a zero unsupported-answer rate. The full component test suite passed.

The component-pass pattern therefore reproduces across two distinct seed sets and across two independent environments.

5. Interpretation / finding

The value of this pass is NOT scale. It is that the core measured objects coexist without losing observability: state, scope, rendering, cursor path, activation budget, and error attribution all remain individually measurable in a single combined run, and no small-cell component regresses when the components are integrated. This is the controlled small-cell integration layer closing — the predecessor primitives are drop-in compatible within one harness.

6. Limits (honest negatives)

This does NOT claim: full real-world Phase 10 validation; a large multimodal architecture; learned ontology; learned cursor routing; or production memory. The metrics are on an engineered controlled small-cell domain; several component scores saturate (perfect values), which is expected for the controlled cell and is precisely why this is not real-world evidence.

The full-roadmap Phase 10 remains open until the same per-component attribution summary passes on a real-world or substantially less-engineered, larger domain. The next stage is to carry the identical per-component attribution summary onto less-engineered evidence while preserving observability of every component.

7. Reproducibility

The harness reproduces across both environments with the same per-component pass pattern; fixed-seed runs are bit-stable under the program's reproducibility binding. Exact seeds, configuration identifiers, and implementation details are withheld from this disclosure.