Self-Esteem as Internal Validator

Mechanism

The self-esteem-class field is computed by a deterministic two-track comparator embedded in Chapter 3 of the cognition patent. The first track, the reported-confidence track, ingests the confidence values that the agent attaches to its own outputs at the moment of action — these are the same confidence values exposed to downstream consumers and recorded in the lineage. The second track, the observed-competence track, ingests outcome data from prior actions: each prior action's reported confidence is paired with its eventual outcome relative to the action's declared objective, and the pair is reduced to a competence sample. The comparator maintains a sliding statistical structure over the competence samples and computes, at each evaluation step, the gap between the rolling reported-confidence distribution and the rolling observed-competence distribution.

When reported confidence consistently exceeds observed competence, the field enters an over-confidence state. The over-confidence state does not directly modify the agent's outputs; instead, it is published to the integrity field as a self-esteem residual, where it is consumed by the three-domain envelope as a multiplier on the lineage-adherence floor. The effect is that an over-confident agent finds the lineage floor tightening on it, requiring stronger evidence of continuity with prior commitments before it can act, until reported and observed values reconverge. When observed competence consistently exceeds reported confidence, the field enters an under-confidence state, in which the self-esteem residual rate-limits the scope of action the agent will accept: the agent declines to expand into novel operational territory faster than its self-assessment justifies, even when external authorization would permit a broader scope.

The comparator is purely deterministic. The same lineage and the same reported-confidence stream produce the same self-esteem state on every replay, on every substrate. The sliding-window parameters, the gap-detection thresholds, and the residual mapping function are all declared in the policy reference and versioned in the lineage. The field exposes not only the over- or under-confidence flag but also the magnitude of the gap, allowing downstream primitives to react proportionally.

Operating Parameters

The field is parameterized by several axes, each policy-governed. The window length governs how many prior actions are considered in the rolling competence statistics; short windows make the field highly responsive to recent performance shifts, while long windows make the field stable against transient noise. The confidence-binning resolution governs how reported-confidence values are discretized for comparison against outcomes; finer bins detect calibration error in narrow confidence bands at the cost of sample sufficiency per bin. The gap-detection threshold governs the minimum magnitude of reported-versus-observed divergence that will trigger a state transition; below the threshold, the field reports calibrated and the residual is zero.

The scope rate-limit governs the speed at which the agent will accept new operational territory when the field is in the under-confidence state; it is expressed as a maximum admissible expansion rate per unit lineage. The over-confidence multiplier governs the strength with which the residual tightens the lineage-adherence floor in the integrity envelope; it permits deployments to choose how aggressively over-confidence is penalized. All parameters are recorded with the lineage on every evaluation, so that an audit can reconstruct not only what the field decided but the parameter regime under which it decided.

The field is also parameterized by an outcome-attestation source, which declares which lineage entries count as competence evidence. Some deployments treat all completed actions as evidence; others restrict evidence to externally attested outcomes; still others weight evidence by the strength of the attestation. The policy reference governs the choice, and the field's behavior shifts accordingly without any architectural change.

Alternative Embodiments

The two-track comparator is the primary embodiment, but several alternatives are contemplated. A multi-track embodiment maintains separate comparators per action class, so that the agent's self-esteem in, say, diagnostic reasoning is tracked separately from its self-esteem in procedural execution; the integrity envelope then receives a vector of class-resolved residuals rather than a single residual. A peer-relative embodiment ingests competence samples not only from the agent's own lineage but from the lineages of peer agents operating under the same policy, producing a self-esteem field that is calibrated against the population rather than against the agent's individual history; this is appropriate for fleet deployments in which a single agent's history is too sparse to support stable statistics.

A coupled-prior embodiment weights observed-competence samples by a recency-decayed kernel rather than a hard window, producing a smoothed competence signal that is robust to bursty outcome streams. An adversarial-resistant embodiment hardens the comparator against lineage-tampering by attesting each competence sample with a non-forgeable lineage signature, so that an attacker who attempts to inflate the agent's apparent competence by injecting synthetic outcomes is detected at the comparator boundary. Each embodiment retains the core property that reported confidence is structurally compared to observed competence and that the resulting residual is published to the integrity field rather than directly modifying outputs.

Composition with Other Primitives

The self-esteem-class field is downstream of the lineage primitive and of the confidence-governance primitive, and upstream of the three-domain integrity envelope and the discovery-traversal primitive. Lineage supplies the outcome data; confidence governance supplies the reported-confidence stream; the integrity envelope consumes the residual as a floor multiplier; discovery traversal consumes the residual as a scope rate-limit. The closed loop runs as follows: the agent reports confidence; the action executes; the outcome enters the lineage; the comparator updates; the residual is published; the integrity envelope and discovery traversal adjust; the agent's next reported-confidence stream is produced under the adjusted regime. No step in the loop is heuristic; each is a pure function of declared inputs and policy parameters.

The field is also coupled to the deviation-function primitive: when a deviation is detected by the function, the deviation event is itself a competence sample, and the comparator updates accordingly. This coupling produces a self-correcting agent: an agent that has deviated and recognized the deviation experiences a reduction in its self-esteem field, which in turn tightens the integrity floors and rate-limits the scope, reducing the probability of further deviation in the immediate term. The mechanism is not punitive; it is calibrative.

A symmetric coupling holds for sustained competence: an agent whose observed outcomes consistently meet or exceed the confidence with which they were declared experiences a gradual relaxation of the under-confidence rate-limit, expanding admissible scope on a schedule governed by policy rather than by self-assertion. The architecture thereby distinguishes earned scope from claimed scope, and the lineage record of self-esteem residuals over time is itself an auditable history of how the agent's authority was earned, granted, and exercised. This auditability is the practical payoff of the structural separation: a regulator examining an incident can read the residual history and determine whether the agent's confidence at the moment of the incident was within or outside its calibrated range.

Distinction Over Prior Art

Prior approaches to confidence calibration in autonomous systems are predominantly statistical post-hoc adjustments — temperature scaling, isotonic regression, Platt scaling — applied to the output distribution of a learned classifier. These approaches calibrate the classifier's outputs to match historical accuracy, but they do not produce a structural field that is consumed by an integrity architecture, do not distinguish over- from under-confidence as separate operational states, and do not couple the calibration result to the agent's admissible scope of action. They are output transformations, not structural validators.

Self-monitoring approaches in autonomous agents typically use uncertainty estimates to throttle action; they conflate the question of how confident the agent is with the question of how calibrated that confidence has been historically. The self-esteem-class field separates these explicitly: reported confidence is one input, historical calibration is another, and the residual between them is the operational signal. No prior approach in the available literature exposes a deterministic, policy-versioned, lineage-recorded comparator that distinguishes over-confidence from under-confidence and routes each to a distinct downstream consumer in the integrity architecture.

Disclosure Scope

The disclosure covers the two-track comparator, the directional residual that distinguishes over- from under-confidence, the routing of the residual to the integrity envelope as a floor multiplier and to the discovery-traversal primitive as a scope rate-limit, the policy-versioned parameter regime including window length, binning resolution, gap threshold, scope rate-limit, and over-confidence multiplier, and the alternative embodiments enumerated above. The disclosure further covers the closed-loop coupling to the deviation-function primitive in which deviation events function as competence samples, and the lineage-attested competence-sample variant that resists adversarial tampering. Implementers operating a self-esteem field that compares reported confidence to observed competence, distinguishes over- from under-confidence as separate operational states, and routes the resulting residual to an integrity architecture for floor or scope modulation are operating within the scope of the disclosure regardless of the specific statistical structure used.

The disclosure extends to deployments in which the self-esteem-class field is exposed to peer agents as a compact attestation of calibration, allowing peer agents to weight collaborative inputs by the calibration of their source. It extends to multi-jurisdictional deployments in which different jurisdictions impose different floor and ceiling regimes on the residual; the field is constructed once and the policy reference selects the active regime by jurisdiction tag at evaluation time. It extends to long-horizon agents whose lineage exceeds the evaluation window by orders of magnitude, in which case the field exposes both a current-window residual and an archival residual computed over the full lineage, allowing downstream consumers to choose which timescale governs their response. It extends to ensemble deployments in which several self-esteem comparators run with different parameter regimes in parallel and a meta-comparator selects or blends their outputs under policy control. None of these extensions alters the core structural property: that the field is a deterministic, lineage-recorded comparison of reported confidence to observed competence whose directional residual is routed structurally to the integrity architecture.