DAWM Phase 7: Multi-Track Consolidation — Hierarchical Semantic Scoping

Status: Closed (controlled hierarchical semantic scoping mechanism). Sub-phases 7.1 through 7.6, plus an oracle diagnostic.

1. Question

A prior phase had already closed the engineered-schema relation-binding problem: a semantic state can form by composing relation-label evidence, pair geometry and direction evidence, and coordinate edge-existence evidence into relation edges. Phase 7 did not reopen that. It asked the next question:

Given a semantic state that can form, can semantic scope reduce the active candidate manifold and improve local precision -- without losing cross-scope recovery?

The core thesis: semantic precision is governed by the active semantic candidate range, not only by model capacity or flat-label accuracy. Hierarchy here is not a label taxonomy; it is candidate-space control.

All worlds are controlled engineered small-cell worlds with predefined hierarchy and observable scope routing, so that scope error stays separable from label error, binding error, and existence error. The closure target is mechanism visibility, not benchmark victory.

2. The six-stage mechanism chain

Phase 7 is a chain of sub-phase closures, each repairing the limit exposed by the prior one.

7.1 — Hard scope

A predefined hard hierarchy sharply cuts the active semantic candidate count and raises precision per active candidate; scope error becomes separable from label, binding, and existence errors. A flat vocabulary exposes on the order of dozens of active candidates; hard scope reduces this to a small handful. Limit: a wrong hard scope creates semantic tunnel vision -- facts living in a sibling scope become unreachable.

7.2 — Soft / overlapping scope

Soft scope mixtures strongly reduce hard-scope tunnel vision while keeping the effective candidate count far below flat; single-scope facts do not materially collapse. An oracle diagnostic first established the effect; a transparent learned soft router was then added. Representative learned results show the soft router roughly halving tunnel-vision rate relative to hard scope while keeping effective candidate count well under the flat baseline. Limit: always-soft pays for the extra scopes even when there is no conflict to resolve.

7.3 — Dynamic expansion

Explicit conflict-conditioned expansion keeps hard scope when it is sufficient and expands only when relation evidence demands it, beating always-soft semantic efficiency. The lesson: scope should be dynamic execution, not a fixed soft mask. Limit: the first success relied on an explicit relation-proxy hint.

7.4 — Learned conflict trigger

Under weakened local evidence (relation subtype, regime, or primary scope removed), a transparent learned trigger detects top-1 scope insufficiency, removes tunnel vision, and keeps candidate count below always-soft, with higher precision per effective candidate than the soft baseline. A no-hint control collapses back toward hard behaviour, confirming the trigger uses real evidence. Limit: risk of lexical leakage -- the trigger could be memorizing controlled toy tokens rather than detecting conflict.

7.5 — Lexical-leak falsification

A falsification campaign: aligned masked evidence still drives useful expansion; globally shuffled and no-hint controls degrade strongly; a structure-only (pure role-token) condition exposes the next wall. Interpretation: the trigger is not only memorizing family names -- but pure role-only conflict evidence is too coarse and over-expands. This splits the wall into two questions: does a conflict exist and which secondary scope to add.

7.6 — Two-stage scope expansion

Decompose into a conflict-exists stage and a which-secondary-scope stage, then recombine into a learned-conflict plus learned-secondary system. Across a multi-seed local run the two-stage system achieves the highest precision per effective candidate of the compared systems while driving tunnel vision to zero, at lower candidate cost than always-soft scope. An independent multi-seed remote replication matched these results. In this control the learned two-stage system reaches the explicit relation-proxy upper bound: conflict precision and recall both at the maximum, secondary-scope accuracy at the maximum, and wrong-scope expansion at zero.

3. Full mechanism chain (summary)

Flat vocabulary -> too many active candidates -> hard hierarchy (candidate reduction and local precision, but tunnel vision when scope is wrong) -> soft overlap (tunnel repair, extra candidate cost) -> dynamic expansion (pay for extra scope only when needed) -> two-stage expansion (separate "should expand" from "which scope", preserving tunnel repair below always-soft cost).

4. What Phase 7 proves

Controlled evidence that hierarchical semantic scoping can improve precision by reducing the active candidate range. It also establishes the observable objects that make the mechanism legible: active semantic candidate count; precision per active candidate; scope-selection accuracy; tunnel-vision rate; the scope-error / label-error / existence-error decomposition; wrong-secondary-scope expansion; and conflict-exists precision and recall. At this stage these observables matter more than raw benchmark score because they preserve mechanism-level visibility.

5. What Phase 7 does NOT prove

Phase 7 remains controlled and engineered. It does not establish ontology discovery, open-world semantic ambiguity handling, natural-language-only scope inference, multimodal semantic scoping, or stateful cursor-graph routing. The honest next transition is not to jump to broad real-world domains, but to reduce the engineered scope evidence while preserving the same error-decomposition (attribution) table that made Phase 7 interpretable.

6. Evidence level

Reproduced in the controlled setting: the headline two-stage result replicates across a multi-seed local run and an independent multi-seed remote run, with falsification controls and an oracle diagnostic bracketing the learned results.