Dual-channel reasoning state and a semantic-depth bridge for contradiction handling

Dual-channel reasoning state and a semantic-depth bridge for contradiction handling

Track DAWM. Evidence level: experimental — small-cell symbolic microworlds (a kinship world and a provenance world). Closure status: positive-with-attribution on the semantic-depth bridge; structural ceilings honestly recorded. This object structures a two-phase result set into two logically distinct contributions: a semantic-state substrate and a controlled-cell observability protocol.

1. Problem and arc

An earlier line tried to teach a reasoning cursor to abstain — to refuse rather than guess when contradictory evidence is peaked-but-split. Three prior abstain mechanisms (a closed-form rule, a learned post-hoc head, and a cursor retrain) each hit the same Pareto wall on the same regime. Two follow-on phases asked different questions: (a) is the wall architectural — can arbitration and calibration not share an output — and (b) is contradiction better modelled as a depth regime than as a probability conflict?

2. Phase one — dual-channel cursor (architectural success, feature wall persists)

A parallel sigmoid abstain channel was added after the cursor's bounded walk, gradient-detached from the arbitration weights and composed with the existing abstain via a max rule. The architectural goal — cleanly separable arbitration and calibration computations in one model — was achieved:

• Resolved-contradiction arbitration accuracy preservation was identical across every sweep cell and both seeds, with cross-config standard deviation of exactly zero. The abstain channel had no measurable effect on the regime it was designed not to touch; that zero-variance is the empirical signature of the gradient-isolation guarantee.
• The diffuse / out-of-distribution regime continued to abstain at roughly nineteen-in- twenty via the pre-existing entropy gate, untouched by the new channel.
• Clean-block reasoning accuracy and clean false-abstain stayed within prior noise.

But the calibration noise floor held. On peaked-but-split (tied) contradictions the tied abstain rate stayed in the low-thirties percent (below the roughly one-half bar), and worst-case confident-wrong on tied stayed in the mid-thirties percent (above its bar). Head-to-head against the three prior abstain attempts at each method's best operating point, the dual-channel approach was mid-pack on tied-abstain but strictly best on clean reasoning, on resolved-contradiction preservation, and on clean regime separation.

Finding (feature-side wall): structural separation of arbitration and calibration outputs resolves the "one softmax cannot carry two signals on one scale" problem but does NOT bridge the calibration noise floor. The residual bottleneck is feature-side: the post-hoc inputs inherit a noise floor equal to the cursor's own internal mass variance, regardless of where the calibration computation physically lives. A learned head over those same inputs reaches the same ceiling whether it shares an output with arbitration or has its own channel.

3. Phase two — the semantic-depth bridge (the decisive cell)

A pure-function depth attribute was added to each candidate edge, ranging over a small discrete set of named epistemic regimes (direct / inferred / contradiction- resolved / speculative). At a configured maximum admissible depth, edges deeper than the budget are zeroed at evaluation time. Same cursor weights, two modes (bridge-only vs. bridge+depth); the only difference is the adjacency-mask construction.

The structural finding: depth filtering is regime-targeted, not uniform compression.

• On peaked-but-split (tied) contradictions: roughly four-fifths of the candidate space pruned (clearing the candidate-reduction bar), confident-wrong dropping by about thirty-four points, and a depth-overflow→abstain rate above four-fifths. This was the first time the system structurally separated "I should refuse" from "I have a preference" in the same regime.
• On resolved contradictions (majority-supported): the filter removes the refuted losing-side edge, so the cursor's mass concentrates on the correct destination — arbitration accuracy improved by about eighteen-and-a-half points (into the low- nineties), with confident-wrong falling about twenty-one points. This was the first milestone in the line to simultaneously improve arbitration AND calibration through one mechanism; every prior abstain attempt traded them off.
• On clean data: an accuracy drop of about fourteen-and-a-half points — initially read as a falsification.

The aggregate reduction and the clean-drop bar both "failed" at aggregate granularity, but the per-regime decomposition shows the bars were at the wrong granularity. The reduction percentage correlates with epistemic status: deep pruning of unresolved- contradiction regions, light pruning of resolved / direct regions — exactly the thesis that contradiction is a semantic depth overflow, not a probability conflict.

3.1 The clean drop is an attributable artifact, not a falsification

In the kinship world a child legitimately has two parents; both edges are TRUE. The depth function, lacking a notion of relation cardinality, counts the second valid destination as a refutation, lands both edges at the tied-contradiction depth, and the depth budget filters away both real parents. The clean drop is therefore a depth-scheme imprecision specific to a set-valued (multi-true) relation ontology — not a failure of the bounded-evolution hypothesis. It is also a pre-committed prediction for the provenance world, whose relations are single-valued by construction and so should NOT exhibit the clean drop.

Finding (regime-selective compression): depth-based candidate filtering is a regime-selective compression mechanism. Its correctness depends on the depth function distinguishing multi-true (set-valued) relations from tied (single-valued) contradictions; any depth-bounded filter must specify how it handles multi-true relations or demonstrate the target world has only single-valued slots.

4. Phase two follow-on — oracle-free typing (the regime-segmented ceiling)

Replacing the world-graph oracle with a purely local support-majority heuristic ("in-world iff local support count is at least the refute count") produced a sharply bimodal result across both worlds:

• Resolved contradictions: the heuristic recovers the oracle's Pareto-positive win EXACTLY — accuracy delta, confident-wrong, and abstain match to three decimal places across both worlds (the resolved rows are bit-stable identical between oracle and heuristic). The majority signal genuinely lives in local persistent state.
• Tied contradictions: the heuristic collapses cleanly to the bridge-only baseline — zero recovery. The information distinguishing a true assertion from an injected fake is NOT in local state; the oracle was reading the out-of-band world graph.

Finding (regime-segmented typing ceiling): oracle-free typing recovery is regime- segmented. It succeeds on exactly the cases where local evidence multiplicity carries the typing signal and fails on exactly the cases where the signal lives in the world graph. The heuristic does not partially solve the typing problem — it perfectly solves the locally-recoverable subset and is structurally blind to the rest. No local- feature heuristic or learned head over the same local features can exceed this ceiling; the wall is in the information content of local state, not in model capacity. No refinement iteration was run because each candidate tiebreaker re-pays one wall by opening another (recency helps one tied family and hurts its mirror; strict-majority reintroduces the multi-true clean regression; cardinality is degenerate within budget).

5. The two contributions this object structures

The evidence establishes two logically distinct contributions.

Contribution A — semantic-state substrate / runtime-object framing. A token state is treated as a typed runtime object exposing, through a read-only projection interface, a named schema of attributes — at least {identity, entity-class, relation- type, evidence-role, contradiction-marker, uncertainty-level, provenance, depth}. The schema is explicit (each attribute has a named type and a deterministic derivation rule from existing persistent state) and inspectable (any admissibility decision must be reproducible by hand from the projected view alone, without consulting model weights). The substrate is additive — it adds an accessor layer over the existing token data structure rather than modifying it. Depth, the substrate's new contribution, is a pure function over fields already populated by the contradiction-feature pipeline, ranging over a small discrete set of named regimes.

Contribution B — controlled-cell execution scaffold / observability protocol. A shared trained cursor is evaluated across at least three admissibility-construction modes that differ only in the candidate adjacency mask. The scaffold exposes per-step candidate-set size as a first-class measurable alongside accuracy, trace-match, abstain, and confident-wrong; gates closure on a multi-tier acceptance protocol (candidate-space reduction / accuracy preservation / trace-match preservation / confident-wrong-not-up / depth-overflow→abstain safety tier); decomposes results into named epistemic regimes (clean / tied / resolved / multi-true / diffuse); and admits an oracle-vs-heuristic bisection that separates the locally-recoverable subset of the typing problem from the structurally-external subset.

6. Boundaries — what this work does NOT establish

• No deployable production runtime; all evidence is small-cell symbolic microworld (a kinship world and a provenance world). No native multimodal evidence.
• No generalisation beyond the four named regimes; continuous-depth or more-than-four- level substrates are explicitly out of scope.
• No learned-component substrate variant; the substrate and admissibility budget are pure-function closed-form (only the cursor is learned).
• No universal substrate transferability — only two microworlds carry substrate-level evidence.
• The structurally-external (tied) typing wall is partly a property of the experimental- world generator; the locally-recoverable (resolved) result carries the structural claim, and the absence-of-out-of-band-signal observation is supplementary.

7. Reusable assets / provenance

• Phase one: a dual-channel cursor subclass (a disable flag reproduces the base model bit-identically), a two-phase training script, and a freeze-arbitration helper, with per-seed result records retained.
• Phase two: a pure-function depth module, a multi-mode decisive evaluation driver, and candidate-set-size metric infrastructure, with per-cell result records retained.
• Findings recorded as: feature-side wall; regime-selective compression with multi-true discrimination; and the regime-segmented oracle-free typing ceiling — structured into the two contributions above.