Three-in-One Traversal: Best, Alternative, and Refutation in a Single Bounded Step

Mechanism

The traversal call accepts an intent, a starting anchor, a credentialed compute budget, and a governance policy. The discovery object descends through the published semantic neighborhood, accumulating cognitive state at each visited anchor: confidence updates, contradiction signals, intent refinement, and integrity tracking against the credentialed source. The traversal does not commit to a single hypothesis as it descends. It maintains three concurrent answer slots — best, alternative, and refutation — and at each anchor visit it considers how the anchor's content updates the contents of all three slots simultaneously.

The best slot holds the answer with the highest substantiated support consistent with the intent. The alternative slot holds the highest-substantiated answer that is materially distinct from the best — a competing hypothesis the traversal cannot dismiss without evidence the traversal has not yet found. The refutation slot holds the strongest counter-evidence the traversal has surfaced against the best answer: anchors whose content, if taken at face value, would render the best answer inadmissible. The three slots are populated jointly and updated jointly. An anchor that strengthens the best may simultaneously weaken the refutation, or strengthen the alternative, or all three; the traversal records each contribution into the lineage entry for the anchor visit.

Cross-validation occurs at termination. Before any answer is returned, the traversal evaluates the three slots against one another under the governance policy. The best is required to remain admissible in the presence of the surfaced refutation; if the refutation crosses a credentialed threshold the best is downgraded and the alternative is re-evaluated to see whether it survives the same refutation. The alternative is required to remain materially distinct from the best under a credentialed distinctness criterion; an alternative that has converged onto the best is rejected and the traversal either re-runs within remaining budget or returns failure. The refutation is required to be substantive: a refutation slot occupied only by weak or speculative counter-evidence is reported as such, so that the caller knows the absence of refutation rather than presence of confirmation.

The bounded compute budget is enforced uniformly across the three answers. The budget is not partitioned into thirds; the traversal is permitted to spend the entire budget on any combination of best-strengthening, alternative-surfacing, and refutation-hunting under the governance policy. Budget exhaustion before all three slots are substantiated produces a credentialed partial-answer record rather than a confident result.

Operating Parameters

The compute budget is credentialed and is specified along three axes: a wall-clock bound, an anchor-visit count bound, and a cumulative-cognitive-load bound that captures the traversal's accumulated context size and inference depth. Exhausting any axis terminates the traversal under the same partial-answer protocol.

The distinctness criterion that separates best from alternative is a credentialed parameter. It can be a semantic-distance metric in the embedding space the discovery layer was authored under, a structural-distance metric across the published semantic graph, a propositional contradiction check against the best, or a combination signed under the governance policy. The criterion is part of the input rather than a hidden default; the caller learns from the lineage entry which criterion was applied.

The refutation-substantiveness threshold is a credentialed parameter. It defines the minimum substantiated-support level a refutation must achieve before it can downgrade the best answer. A high threshold produces a primitive that defends the best aggressively; a low threshold produces a primitive that surrenders the best at the first credible counter-evidence. The threshold is a policy parameter that operators tune for their domain — clinical decision-support tunes high to avoid premature retraction; investigative discovery tunes low to surface counter-evidence early.

Lineage is sealed per traversal. The single traversal produces a single signed lineage entry that contains the visited anchors, the per-anchor contributions to all three slots, the cross-validation result, and the credentialed parameters under which the traversal ran. The lineage entry is the artifact downstream consumers of the answer reason about; the answer is not separable from the lineage that produced it.

Governance is shared across all three answers. The admissibility evaluator that authorizes the traversal authorizes it as a unit; an answer that would be admissible in isolation but inadmissible in the presence of the other two is rejected. This is the property that makes the primitive's output intrinsically cross-validated rather than independently scored.

Alternative Embodiments

The primitive admits embodiments distinguished by traversal strategy, slot-update policy, and termination condition. Traversal strategies include depth-first descent with backtracking, beam-search across a credentialed beam width, and parallel multi-frontier descent under a shared cognitive state. The choice of strategy is a parameter of the primitive rather than a structural change; all strategies populate the same three slots under the same cross-validation rule.

Slot-update policy can be eager or deferred. Eager policy updates the slots at every anchor visit and pays the cost of more frequent cross-validation passes. Deferred policy accumulates anchor contributions and updates the slots at credentialed checkpoints, paying less per-visit cost but risking later detection of contradictions. Embodiments tuned for low-latency interactive discovery favor deferred; embodiments tuned for high-stakes decision support favor eager.

Termination condition can be budget-driven, confidence-driven, or contradiction-driven. Budget-driven termination is the universal floor; confidence-driven termination ends the traversal early when the best, alternative, and refutation slots have all reached credentialed-substantiation thresholds before budget is exhausted; contradiction-driven termination ends the traversal when the refutation has crossed the downgrade threshold against every candidate best in the visited subgraph, returning a credentialed no-confident-answer result.

The primitive admits a degenerate embodiment in which the alternative and the refutation slots are configured to be the same artifact — a refutation that is itself a competing hypothesis. This embodiment collapses to a two-answer primitive (best and counter-best) under a cross-validation rule that subsumes both alternative-distinctness and refutation-substantiveness. The degenerate embodiment is disclosed for completeness; the principal embodiment maintains the slots as structurally distinct because the questions of "is there a competing answer?" and "is there evidence against this answer?" are operationally different and are tuned by different parameters.

Cross-traversal composition is supported. A higher-level discovery routine can issue multiple three-in-one traversals across distinct intents and combine their lineage entries under a credentialed combinator. The combinator itself does not collapse the three-answer structure; downstream consumers see the per-traversal best, alternative, and refutation alongside the combined verdict.

Worked Example: Clinical Decision-Support Traversal

Consider a clinical-decision-support discovery object invoked with the intent "identify the most likely diagnosis consistent with the presenting findings, surface a competing diagnosis the findings do not exclude, and surface the strongest counter-evidence against the leading diagnosis," with a starting anchor at the patient's presenting-complaint record and a credentialed compute budget of two seconds wall-clock, four hundred anchor visits, and a cumulative cognitive-load cap appropriate to the bedside response time. The traversal descends through the credentialed clinical-knowledge graph: anchors representing symptom clusters, anchors representing laboratory finding ranges, anchors representing imaging interpretations, anchors representing medication-history interactions, each authored by an attesting clinical authority and signed under that authority's key.

Early in the descent a symptom-cluster anchor strongly supports a leading diagnosis and the best slot is populated. As the traversal continues, an anchor representing a laboratory finding partly contradicts the leading diagnosis; the contribution is recorded into the refutation slot rather than discarded. A medication-history anchor surfaces a competing diagnosis whose clinical course matches the same presenting findings under a different mechanism; this enters the alternative slot. A subsequent imaging-interpretation anchor strengthens the best, weakens the refutation slightly, and does not affect the alternative. At termination the cross-validation pass evaluates the slots: the refutation has crossed the credentialed substantiveness threshold for the domain (clinical, tuned high to avoid premature retraction) but has not crossed the downgrade threshold; the alternative remains materially distinct from the best under the credentialed semantic-distance metric the discovery layer was authored under. The traversal returns the best diagnosis, the alternative diagnosis, the refutation, and a sealed lineage entry that names every visited anchor and every per-anchor contribution.

Downstream the clinical user consumes the lineage entry as a single artifact. The actuation gate — order an additional test, prescribe an empirical therapy, escalate to specialist review — is evaluated against the entire entry rather than against the best in isolation. A platform that offered only the best, with no alternative and no refutation, would not pass the credentialed gate; the gate requires the three coupled answers as inputs.

Composition with the Semantic-Discovery Stack

Three-in-one traversal composes with the credentialed-anchor authorship primitive, with the cognitive-state primitive, with sealed lineage, and with the bounded-compute-budget primitive. Each anchor visited by the traversal is itself a credentialed artifact whose published semantic neighborhood was authored under a signing key the traversal verifies. Each cognitive-state update is recorded into the lineage entry the traversal produces. The compute budget is credentialed under the same policy that authorized the traversal, and budget exhaustion is recorded as a credentialed event.

The primitive composes upward into the discovery-as-action stack. A discovery object that uses three-in-one traversal as its primitive evaluation step inherits the falsifiability property: every answer the object presents is accompanied by a surfaced alternative and a surfaced refutation, and any downstream actuation gated on the answer is gated on the entire lineage entry rather than on the best answer in isolation. The actuation governance disclosed in companion primitives consumes the lineage entry uniformly.

Prior-Art Distinction

Existing search and retrieval architectures return ranked candidate lists, with no structural commitment to surfacing alternatives that are materially distinct from the top candidate or to surfacing counter-evidence against the top candidate. Existing inference architectures return a confident answer with optional confidence scores, with no structural commitment to evaluating the answer in the presence of its strongest counter-evidence. Existing decision-support architectures bolt cross-validation on top of independent evaluators, paying the cost of running each evaluator separately and accepting that the evaluators do not share context.

The primitive disclosed here distinguishes itself by structuring all three answers as outputs of a single traversal under a single budget with shared cognitive state, and by requiring cross-validation among the three before any is returned. The shared-traversal property collapses the cost of the additional answers from N-times-single-evaluator to a small constant overhead on the single evaluator, because the same anchor visits update all three slots. The shared-budget property prevents the failure mode where a system spends its budget on confirming the best and has nothing left to surface alternatives or refutations. The structural-cross-validation property prevents the failure mode where the answers are individually plausible but jointly inconsistent.

Closest related art in argument-mining and counterfactual-explanation work surfaces refutations after the fact, against an already-committed best answer. The primitive here surfaces them concurrently, under shared context, and uses them as inputs to the commitment decision rather than as a post-hoc commentary on a decision already made.

Disclosure Scope

The disclosure covers the primitive across semantic-discovery domains where the consumer of the answer is operationally exposed to the cost of an unsurfaced alternative or an unsurfaced refutation. Disclosed deployment contexts include clinical decision-support where a competing diagnosis or a counter-indication must be surfaced rather than buried below a top result; investigative analysis where a competing narrative or exonerating evidence must be surfaced under the same budget that surfaced the leading narrative; regulatory review where a counter-finding to an inspector's leading conclusion must be substantiated within the same review interval; and any autonomy stack whose actuation is gated on a discovery answer and whose safety case requires the answer to have been evaluated in the presence of its strongest counter-evidence.

The disclosure includes the three coupled slots, the shared compute budget across the three axes, the credentialed distinctness criterion, the credentialed refutation-substantiveness threshold, the per-traversal sealed lineage, the cross-validated termination, the alternative traversal strategies and slot-update policies, and the cross-traversal composition under credentialed combinators. The disclosure does not depend on a particular embedding model, a particular semantic-graph format, a particular signing infrastructure, or a particular cognitive-state representation; the primitive is specified at a level of abstraction that admits substitution along each of those axes provided the three-coupled-answers-under-shared-budget-with-cross-validation property is preserved.