Learned cursor routing under activation-graph cells: a negative-then-repair study of multi-candidate disambiguation, abstain stability, and deferred commitment

Overview

This study reports a two-step campaign on learned cursor routing within an activation-graph workspace model. The first step is a decisive negative closure that surfaced two structural boundary findings. The second step applied targeted cell-, policy-, and loss-level fixes, achieved a multi-seed success on the load-bearing hypotheses, and produced a third boundary finding. Together the three findings form a coherent disclosure cluster on the operating conditions of a learned-routing primitive.

Step one: decisive negative

A single-seed full training run used a small behavior-cloned policy (on the order of a few thousand parameters), trained for several thousand steps on a few thousand oracle traces, in well under a minute of wall-clock time.

Final validation showed select accuracy at ceiling but trivially so (see below), while abstain accuracy landed near chance (~0.49), failing the strict abstain bar decisively.

Per-hypothesis verdict

• Trace equivalence (strict bar): FAILED, with a combined estimate dominated by abstain inconsistency.
• Budget-bounded abstain (strict bar, zero hallucination): DECISIVELY FAILED. Abstain accuracy on the held-out set was worse than an always-abstain baseline given the roughly 79% select / 21% abstain mix.
• Dense/cursor reduction ratio: NOT MEANINGFULLY TESTED; it passed only by inheritance from the parent cell's already-validated baseline, i.e. "the policy did not break the cursor," not learned-routing evidence.
• Wrong-route abstain (strict bar): FAILED on the same abstain population.

Boundary finding: routing-ambiguity prerequisite

The decisive structural discovery was that at every select step in every generated trace, exactly one candidate matched the (current-entity, requested-relation) filter. This was confirmed across several distractor counts, with a multi-candidate rate of zero in all conditions. The parent cell's generator never created two fact-nodes with the same (source, relation) but a different destination within one situation, and the distractor relations were deliberately distinct from the chain relations, so distractors added inspection cost but not candidate count. The "pick the right candidate" task was therefore trivially solvable because there was nothing to pick between: learned routing was structurally equivalent to a fixed lowest-id rule when there are only zero or one matches.

Finding statement: A demonstration of learned cursor routing requires an activation-graph cell with empirically multi-candidate routing decisions at the per-step matching layer. A cell construction in which target relations are distinct from distractor relations produces zero multi-candidate steps and makes learned routing structurally equivalent to rule-based routing.

Boundary finding: abstain instability under behavior cloning for a tiny policy

The abstain task itself is non-trivial, but the policy never converged on it: abstain accuracy oscillated across training and landed near chance, despite an up-weighted abstain loss. The cross-entropy signal on the stop output was sparse and conflicted with the implicit "always select when candidates exist" signal from the select task. Diagnosis: the small cursor-state feature representation was too weak to separate "valid path forward" from "should stop" stably.

Finding statement: A behavior-cloned cursor policy with a small cursor-state feature dimension and a stop output trained via cross-entropy exhibits unstable abstain convergence when (a) all select steps have single candidates and (b) abstain examples are a training-data minority.

Pre-declared scope (step one)

DOES claim: zero routing ambiguity at the select step of this cell; learned routing cannot be validated against the rule baseline on this cell; abstain learning unstable on this tiny policy under behavior cloning with the current features. Does NOT claim: that learned cursor routing is fundamentally impossible; that larger policies or different features fail similarly (untested); that the underlying research vertical is falsified.

Step two: success with a new boundary

The successor attacked both negative findings with structural fixes and re-ran across three seeds (several thousand steps over a couple thousand items, on the order of a few tens of thousands of policy parameters, a couple of minutes wall-clock per seed).

The three load-bearing fixes (high-level)

• Cell-level: a multi-witness cell creates multiple fact-nodes with identical (source, relation) but distinct destination, forcing the cursor to disambiguate on a small set of per-witness fields. The multi-candidate rate moved from zero to a substantial fraction of select steps.
• Policy-level: a modestly larger policy adds dedicated heads for per-candidate confidence, a separated abstain-confidence decision, a deferred-commitment signal, and a hold-count distribution, together with an enlarged cursor-state representation and an explicit candidate-count feature. Abstain accuracy moved from near-chance instability to ceiling and was stable across seeds.
• Loss-level: the objective is decomposed into several components rather than two: behavior cloning, candidate-confidence supervision, a focal abstain term, a stop term, a deferred-commitment term, and a hold-count term. Candidate-confidence supervision uses a per-witness support signal as target so the policy surfaces its own uncertainty, and the abstain term uses a focal objective to handle the minority-class imbalance.

Load-bearing positive: observability preserved

Trace equivalence passed unconditionally: all trace-audit metrics (cursor-state schema completeness, trace-path equivalence, routing-trace completeness, provenance-trace consistency, stop-state consistency, provenance consistency) hit ceiling across all three seeds. This is the central observability claim: learned routing does not break observability when the policy is wrapped in a cursor execution function that preserves the trace schema by construction.

Caveat on multi-select accuracy: disambiguation is partly driven by exposing a per-witness support signal as a feature. A policy trained without that signal would have to disambiguate purely structurally, an explicit open follow-up that is out of scope for this version.

Boundary finding: delayed-commit utility ceiling

The delayed-commit hypothesis failed its uplift threshold with no observed benefit, but the failure is structural rather than a mechanism failure: greedy per-step commit already reaches ceiling answer accuracy, so there is no headroom for a deferred-commitment primitive to improve on. The hold mechanism itself executes correctly; the cell is simply too easy to benefit from holding.

Finding statement: A delayed-commit routing primitive (deferred commitment via multi-route holding when candidate confidence is split) provides measurable answer-accuracy benefit only in cells where the greedy-commit baseline is below the answer-accuracy ceiling; the achievable benefit is bounded above by the gap between ceiling and greedy accuracy. To make this finding testable, an evaluation cell must attenuate the strongest disambiguation signal for some witnesses, introduce structural multi-path ambiguity, or force longer chains where early-commit errors compound.

Pre-declared scope (step two)

DOES claim: multi-witness routing decisions can be learned with preserved trace equivalence (3-seed stable); the uncertainty representation plus a focal abstain term stabilizes a decision boundary that behavior cloning alone could not; cell-, policy-, and loss-level fixes are each load-bearing; the delayed-commit ceiling is a new operating-condition disclosure. Does NOT claim: that learned routing replaces all rule-based routing; that a general vertical crossing is achieved (one controlled cell, engineered features); that delayed commitment is necessary in this workspace model (conditional on the greedy ceiling); and no AGI, autonomous-reasoning, or transformer-replacement claims.

Net result

The campaign is a clean negative-then-repair arc. The negative precisely characterized why the first attempt could not produce learned-routing evidence; the repair validated three structural fixes and preserved observability across three seeds; and the residual negative bounds the operating envelope of the delayed-commit primitive. The three findings (routing-ambiguity prerequisite, abstain stability conditions, and delayed-commit utility ceiling) form a single operating-condition disclosure cluster for a learned-routing primitive.