Catalog of mechanism and boundary findings

This is a one-entry-per-finding index over a structured-state reasoning research program. Each entry carries a disclosure status: either mechanism (a working primitive, repair, or applied use) or boundary (a characterized limit, ceiling, or honest negative). Numeric configuration values, thresholds, weights, and budgets are referenced by description only and do not appear here. One finding slot is deliberately retired (see the note on retired findings below); a separate process/governance entry records an experiment-governance lesson rather than a claim-bearing technical result.

Substrate findings

Semantic-state substrate / runtime-object framing. Mechanism. A token state is treated as a typed runtime object exposing, through a read-only projection interface, a named schema of attributes (identity, entity-class, relation-type, evidence-role, contradiction-marker, uncertainty-level, provenance, depth). The schema is explicit and inspectable: admissibility decisions referencing any attribute must be reproducible from the projected view alone, without consulting model weights. The substrate is additive: it does not modify the underlying token data structure, only surfaces already-derivable signals under one named contract. Its new contribution, the depth attribute, is a closed-form (non-learned) function over already-populated fields, ranging over a small discrete set of named epistemic regimes (direct / inferred / contradiction-resolved / speculative). Boundary: small-cell symbolic microworlds only (e.g. kinship, provenance), no multimodal evidence, a small fixed set of regimes only, closed-form and non-learned, with no demonstrated universal transferability.

Controlled-cell execution scaffold / observability protocol. Mechanism. Reasoning evolution is gated by a controlled-cell scaffold in which a single shared trained cursor is evaluated across multiple admissibility-construction modes that differ only in how the candidate adjacency mask is built. The scaffold exposes per-step candidate-set size as a first-class measurable alongside accuracy, trace-match, abstain rate, and confident-wrong rate, and gates closure on a multi-tier acceptance protocol (candidate-space reduction, accuracy preservation, trace-match preservation, confident-wrong-not-increasing, and a depth-overflow-to-abstain safety tier). It decomposes into named epistemic regimes (clean, tied, resolved, multi-true, diffuse) and admits an oracle / non-oracle bisection separating the locally recoverable subset of the typing problem from the structurally external subset. A bit-exact oracle-equivalence sub-result holds on the resolved subset (oracle and support-majority heuristic produced identical accuracy and confident-wrong deltas there). Boundary: two microworlds only; the regime decomposition is not shown exhaustive; the cursor is learned but the admissibility budget is not; the second structural wall is partly a generator-construction property rather than a universal information-theoretic claim; no adversarial or open-world coverage.

Decomposition and scope findings

Edge-decode three-channel decomposition rule. Mechanism. Edge decoding factors into three separable channels (label, direction, pair-existence), enabling a structured negative-control decomposition.

Multi-class negative-control protocol. Mechanism. A structured family of structurally distinct negative controls (at least six families) isolating which channel a perturbation attacks; the controls are structural, not tied to specific perturbation magnitudes.

Two-stage scope-expansion rule. Mechanism. A progression from flat through hard / soft / dynamic to a two-stage scope-expansion regime for candidate filtering.

Cursor execution observability. Mechanism. Activation-graph cursor execution observability with budget-bounded abstain, surfaced as a set of audit equalities over the cursor's per-step behavior (candidate confidence, abstain decisions, budget-bounded wrong-route abstain).

Vocabulary-density boundary. Boundary. The multi-encoder convergence regime holds within a vocabulary-density operating cell; beyond a density limit the convergence degrades. The existence of the boundary is disclosed; the exact entity-count threshold is configuration and is not given.

Routing, slot, and economic findings

Slot-stationarity collapse boundary. Boundary (not repaired). Under permutation-invariant set matching, entity slots collapse to a stationary degenerate assignment. Sharpened by three negative repair mechanisms (slot-anchor cross-entropy, Sinkhorn-anchor, persistent-warmup), each of which induced degenerate collapse or failed persistence, all capping well below the upgrade bar, so the boundary stands with three negative mechanism results bounding it. A later slot-diversity penalty cleared the acceptance criteria on a majority but not the required supermajority of seeds, so the boundary still stands.

Routing-ambiguity prerequisite. Boundary. Learned routing requires a genuine routing ambiguity (a multi-witness construction) at the select step; without it the learned-routing primitive has nothing to resolve. Recorded as a zero-ambiguity-cell limit.

Stop/abstain instability under behavior cloning for a tiny policy. Boundary. For a small policy trained by behavior cloning, the stop/abstain decision is unstable (a set-reconstruction collapse). Mitigations framed in terms of policy capacity and loss-weight balance.

Routing repair via multi-witness; delayed-commit utility ceiling. Mechanism (with paired ceiling). A multi-witness construction repairs the routing-ambiguity prerequisite; the delayed-commit strategy (holding multiple routes under top-1 / top-2 candidate-score ambiguity) has a characterized utility ceiling.

Cost-aware learned-routing primitive. Mechanism. A learned routing policy carrying cost-aware heads (economic routing): the primitive on which budget-aware behavior is built.

Activation-reduction utility ceiling. Boundary. The activation-reduction strategy has a utility ceiling; paired with the routing-repair ceiling it characterizes the economic-optimization envelope.

Budget-tight economic pre-emption. Mechanism (applied use of cost-aware routing). Under a tight activation budget the policy can decline to start (economic pre-emption) rather than begin a rollout it cannot afford. The configured budget itself is configuration, referenced by description only.

Early-stop head not trained as a mid-rollout detector. Boundary. An auxiliary head supervised passively (as a no-op readout) is not thereby enabled as a live mid-rollout controller; auxiliary heads require active-role supervision to act as controllers. Pairs with the routing-ambiguity and stop-instability boundaries as a heads-must-be-trained-in-their-operating-role envelope.

Composition and schema-expressiveness findings

Cross-frontier composition is architecturally undefined. Boundary (cross-frontier architecture coherence). Two frontiers (a permutation-invariant graph-set encoder and a per-step cursor MLP) have no defined injection seam; multi-frontier composition requires an explicit seam to be specified first. This is a structural disclosure with no numeric thresholds involved.

Value-erased schema cannot support value-mismatch negative control. Boundary (schema expressiveness). A schema rendering that erases value channels cannot, by construction, distinguish value-matched from value-mismatched cases; the value-mismatch negative control is structurally impossible under such a schema. This bounds the rendering channel-content required for value-sensitive convergence.

Note on retired and governance-only findings

One finding slot is deliberately retired. An early run appeared to show an anti-coherence effect; a degradation diagnostic proved the apparent effect was an artifact of a wiring no-op (a schema/noisy stream silently did nothing because the training step was not extended alongside the loss computation). The anti-coherence claim was false, so no finding was minted, and the numbering skips that slot.

A separate governance entry records that wiring trap itself as an experiment-governance artifact (how a silent no-op was caught and quarantined). It is a process and decision-log artifact, not a claim-bearing technical finding.

Summary

The catalog comprises eighteen named entries. Roughly half are disclosable mechanisms (the semantic-state substrate, the controlled-cell observability scaffold, the edge-decode decomposition, the negative-control protocol, the scope-expansion rule, cursor execution observability, the multi-witness routing repair, the cost-aware routing primitive, and economic pre-emption). The remainder are boundaries, several of them honest negatives: a vocabulary-density ceiling, a slot-stationarity collapse bounded by three failed repairs, a routing-ambiguity prerequisite, stop/abstain instability under behavior cloning, an activation-reduction utility ceiling, an untrained-head limit, a cross-frontier composition gap, and a schema-expressiveness limit on value-sensitive negative controls. One slot is retired and one entry is governance-only.