Confidence-Modulated Discovery Traversal

Mechanism

The discovery traversal mechanism described in Chapter 5 of the cognition patent treats each transition between anchors in the semantic index as a discrete evaluation event. When the traversal advances from anchor A to anchor B along a candidate edge, the system computes a per-step confidence value c_i bounded to the closed interval [0, 1]. This value is not a similarity score in the conventional embedding-distance sense; it is a structured quantity derived from a deterministic function whose inputs include the edge weight, the anchor coverage of the originating step, the residual uncertainty inherited from the prior step, and a domain-configurable evidence factor drawn from the policy reference.

Per-step confidence values are written to the traversal lineage as the walk proceeds. Each entry records the evaluating function identifier, the input fields consumed, the resulting c_i, and the cycle index at which the evaluation occurred. Lineage entries are append-only and content-addressed, so any later auditor can replay the traversal from canonical inputs and arrive at the same c_i sequence. There is no implicit state carried in the evaluator; all dependencies are explicit fields of the cognitive frame.

The overall traversal confidence C is computed as a bounded aggregation across the c_i sequence. The aggregation function is required to satisfy three structural properties: monotonicity (adding a low-confidence step cannot increase C), bounded codomain (C remains in [0, 1] regardless of path length), and locality (C depends only on the c_i values and a small number of policy-defined aggregation parameters, not on the absolute identities of the anchors visited). The canonical aggregator is a length-normalized geometric mean modified by a tail-sensitive minimum, so that a single near-zero c_i suppresses C even when the average is high; alternative aggregators are admitted as alternative embodiments.

Downstream policy is keyed on the confidence vector ⟨C, min(c_i), var(c_i), depth⟩ rather than on a single scalar. The vector is exposed to the agent's gating layer, where it modulates whether the discovery result is admitted as evidence, admitted with a caveat, deferred for further traversal, or rejected. Because the gating layer reads structured fields, policies can be written that distinguish between a uniformly mediocre traversal and a traversal with a single catastrophic step, even when both have similar mean confidence.

Operating Parameters

The mechanism exposes a defined set of operating parameters through the policy reference, each of which is bounded and validated at policy load time. The aggregation exponent governs how strongly the geometric mean responds to outliers and is constrained to the interval [0.5, 4.0]; values outside this range are rejected by the policy validator. The tail-sensitivity coefficient blends the aggregate against the per-step minimum and is constrained to [0, 1], with a default near 0.3 that balances mean-driven admission against single-step suppression.

Per-step confidence floors and ceilings are declared per anchor class. An anchor class corresponding to a high-authority canonical source may declare a floor at 0.4, preventing the per-step contribution from collapsing entirely on a single weak edge; an anchor class corresponding to speculative or user-generated content may declare a ceiling at 0.8, preventing apparent certainty from being inherited from sources whose authority does not warrant it. These floors and ceilings are applied before aggregation, ensuring that the bounded codomain property holds even when the underlying evidence factor is degenerate.

Traversal depth limits, evaluation cycle counts, and recomputation cadence are also exposed as parameters. A traversal may be configured to recompute confidence on every step, on every k-th step, or only at admission boundaries; the choice trades audit granularity against computational cost. The hysteretic margin around the admission threshold is parameterized to prevent oscillation when a traversal hovers near the boundary across successive cycles, and the margin is required to be strictly positive at policy load.

All parameter changes are versioned. When a policy reference is updated, the prior parameter set remains addressable in the lineage, so any historical traversal can be reinterpreted under either the policy in force at the time of the walk or any subsequent policy revision. Parameter rollouts can therefore be staged, A/B compared on replay, and rolled back without re-executing the underlying traversals.

Alternative Embodiments

The canonical embodiment uses a length-normalized geometric mean with a tail-sensitive minimum, but the structural requirements admit several alternatives. A first alternative replaces the geometric mean with a Bayesian update across per-step likelihoods, treating each c_i as the likelihood of a hypothesis "this step preserves semantic coherence" and updating a beta-distributed prior. The aggregate C is the posterior mean. This embodiment is preferred when the discovery is being conducted against a corpus with known prior distributions over anchor reliability.

A second alternative uses a min-aggregator with a smoothing tail, taking C = min(c_i) + epsilon * (mean(c_i) - min(c_i)). This embodiment is preferred in safety-critical domains where a single weak step must dominate the aggregate. A third alternative uses an interval-valued confidence, where each c_i is a pair ⟨lower, upper⟩ and the aggregate is also a pair; downstream policy can then key on the width of the interval as a proxy for epistemic uncertainty distinct from aleatoric uncertainty.

A fourth alternative permits per-class aggregators: high-authority anchor classes are aggregated geometrically, while speculative classes are aggregated with min-dominance, and the overall C is a policy-defined combination of the per-class sub-aggregates. A fifth alternative integrates the aggregator with the forecasting engine, so that c_i values along predicted paths are weighted by the prior probability of the prediction; this couples discovery confidence with forecasting confidence and is described in companion disclosures.

Composition With Adjacent Mechanisms

Confidence-modulated discovery traversal composes with the canonical-fields layer, the policy reference, the gating subsystem, and the multi-agent confidence propagation mechanism. The canonical-fields layer supplies the structured inputs to the per-step evaluator; without canonical fields, the evaluator would be reading from unstructured state and could not satisfy the determinism requirement. The policy reference supplies the parameters and the aggregator selection; without policy-driven configuration, the mechanism would not be portable across domains.

The gating subsystem consumes the confidence vector and applies admission rules. The composition is strict: the gating layer never reads raw retrieval scores from the discovery subsystem, only the structured confidence vector. This separation ensures that any change to the underlying retrieval substrate (vector store, graph engine, hybrid index) does not require corresponding changes in the gating policy, because the contract between the layers is the confidence vector schema rather than the retrieval implementation.

When the traversal spans multiple agents, the multi-agent confidence propagation mechanism takes over at the boundary, transforming the confidence vector under the bounded propagation rule before passing it to the downstream agent. The composition guarantees that a traversal initiated in one agent and continued in another produces an aggregate confidence that is no greater than the aggregate that would be computed if the entire traversal had been local; cross-boundary traversal cannot manufacture confidence.

Prior-Art Distinction

Conventional retrieval systems return a ranked list with similarity scores; they do not compute a per-step confidence along a multi-step traversal, and they do not aggregate such values into a bounded scalar usable by a policy layer. Graph traversal systems used in knowledge-graph question answering compute path scores, but those scores are typically unbounded sums or products of edge weights with no structural guarantee that they remain in a comparable range across paths of different lengths. The bounded aggregation property is absent.

Probabilistic logic programming systems compute proof confidences but do so under closed-world assumptions that are inappropriate for open-domain semantic discovery. Bayesian network inference produces posterior distributions but requires the network structure to be specified in advance; discovery traversal operates over an index whose anchor structure is not enumerated as a Bayesian network. The mechanism disclosed here differs in that the aggregation is path-defined rather than structure-defined, and the policy keying is on a structured vector rather than on a single posterior.

Retrieval-augmented generation systems that use confidence thresholds typically apply a single threshold to a single retrieval score, with no notion of per-step confidence across a multi-hop traversal and no bounded aggregation. The mechanism disclosed here is structurally distinct in that it separates the evaluation function, the aggregation function, the parameter set, and the gating policy into independently auditable components.

Disclosure Scope

This disclosure covers the structural mechanism by which per-step confidence is computed during discovery traversal, the bounded aggregation across a traversal sequence, the structured confidence vector exposed to downstream policy, and the parameterization through a versioned policy reference. The disclosure includes the canonical embodiment with a tail-sensitive geometric mean and the enumerated alternative embodiments. The disclosure extends to compositions with the canonical-fields layer, the gating subsystem, and the multi-agent propagation mechanism.

The disclosure does not cover the choice of underlying retrieval substrate, the specific embedding model used to produce edge weights, or the user-facing presentation of confidence values. Implementations using vector stores, graph databases, hybrid indices, or future retrieval substrates are within scope provided they expose per-step values to the deterministic evaluator. The disclosure is intended to be substrate-independent so that the structural guarantees survive substrate evolution.

The disclosure further covers the lineage encoding of per-step confidence, the policy-validation rules applied to aggregator selection at policy load time, and the schema of the confidence vector exposed to downstream consumers. The disclosure includes embodiments in which per-step confidence is stored as a fixed-point quantity for deterministic replay across heterogeneous floating-point implementations, and embodiments in which the confidence vector is extended with additional structured fields such as anchor-class composition, traversal entropy, or path-length penalty without departing from the bounded-aggregation guarantee. Reasonable extensions of the vector schema that preserve monotonicity, codomain bounding, and locality are within the scope of the disclosure, regardless of the specific names assigned to additional fields or the order in which they are serialized in the canonical envelope.

Equivalents are intended to be covered. A field operator that satisfies monotonicity, bounded codomain, and locality, even when expressed in formal notation different from the geometric mean disclosed as canonical, is structurally equivalent for the purposes of this disclosure. Likewise, a confidence vector exposing the same downstream-policy keying surface under alternative field names is within scope. The mechanism is defined by its structural properties rather than by the lexical form of any particular implementation.