Inference Transition as Mutation

Mechanism

Each inference model registered with the agent advances through a declared lifecycle of phases. Cold-start is the phase in which the model has been instantiated but has not yet accumulated enough live evidence to operate at full weight. Warm-up is the phase in which the model is producing outputs that are admissible but bounded, used for calibration and for accumulation of trust slope. Steady is the phase in which the model operates at its declared nominal weight under its declared competence envelope. Fade is the phase in which the model is being retired, its weight monotonically declining toward zero while it continues to publish outputs for lineage and disagreement measurement.

A transition is the discrete event of advancing a model from one phase to the next. Each transition is materialized as a structured transition record that names the source phase, the destination phase, the triggering condition, the policy version under which the transition was admitted, and a snapshot of the model's parameters and competence envelope at the moment of the transition. The record is committed to lineage before the transition takes effect on subsequent cycles, so the transition itself is reproducible from recorded inputs.

Mutations across transitions are audited. A mutation, in this context, is any change to the model's parameters, weights, competence envelope, or output schema that occurs as part of the transition. The auditor compares the pre-transition snapshot against the post-transition snapshot under a declared diff function, classifies the mutation under a declared taxonomy of permissible changes, and rejects transitions whose mutations fall outside the taxonomy. This prevents a model from acquiring new capabilities or shedding declared constraints through an unaudited phase change.

Runtime behavior is phase-bounded. The set of admissible outputs, the maximum weight the arbiter may assign, and the workload classes for which the model may be invoked are all functions of the model's current phase. A cold-start model may produce outputs but cannot dominate the arbiter; a fade-phase model may produce outputs but cannot enter the arbiter's contributing set above a declining cap. These bounds are enforced by the arbiter and the transition governor jointly, and the bounds in force at any cycle are recorded in lineage with the cycle's outputs.

Operating Parameters

Each phase carries a declared dwell minimum, a declared dwell maximum, and a declared advancement predicate. The dwell minimum prevents premature advancement from cold-start to steady before sufficient calibration has accumulated. The dwell maximum prevents indefinite warm-up by forcing either advancement to steady or transition to fade once the maximum is reached. The advancement predicate is a structured expression over the model's accumulated trust slope, the disagreement field measured against peer models, and any active confidence governance attenuation; advancement occurs when and only when the predicate evaluates to true under recorded inputs.

The mutation taxonomy is itself a declared parameter of the auditor. Permitted mutation classes typically include parameter refinement under a bounded norm, expansion of competence envelope only when accompanied by a recorded calibration record, contraction of competence envelope at any time, and weight rescaling within declared bounds. Forbidden mutation classes include silent change of output schema, expansion of admissible workload classes without calibration, and any mutation that would invalidate the lineage of prior outputs.

Phase-bounded weights are parameterized as monotone functions of dwell and trust slope. The cold-start weight rises from zero to the warm-up cap as the dwell minimum elapses; the warm-up cap is held until the advancement predicate fires; the steady weight is the model's declared nominal; the fade weight declines monotonically from steady to zero across the fade dwell. The weight schedule is recorded in policy and is reproducible from declared parameters.

Alternative Embodiments

The lifecycle admits richer phase enumerations in specific embodiments. A diagnostic embodiment introduces a probe phase between warm-up and steady in which the model is invoked but its outputs are recorded for analysis only and contribute zero weight to the arbiter. A quarantine embodiment introduces a hold phase reachable from any other phase by a governance trigger; in hold, the model continues to be invoked for lineage but contributes zero weight pending external review. A snapshot-rollback embodiment preserves the pre-transition snapshot indefinitely so that a transition can be reversed by re-installing the snapshot under a declared rollback record.

The mutation auditor admits alternative diff functions. A structural diff compares declared schema and competence envelope only and ignores parameter values, suitable for models whose parameters are large but whose declared structure is small. A weighted-norm diff compares parameter vectors under a declared norm and is suitable for models whose parameters are the primary surface of mutation. A behavioral diff replays a declared probe set against pre-transition and post-transition snapshots and compares outputs, capturing functional mutations that may not be visible in either structure or parameters.

Phase-bounding admits domain-specific extensions. In a vehicular embodiment, fade phase forbids any contribution to safety-critical workload classes regardless of declining weight. In a clinical embodiment, cold-start phase forbids any contribution to workload classes that produce externally communicated recommendations, restricting cold-start outputs to internal lineage. In a trading embodiment, transition events themselves are subject to a market-hours predicate that forbids advancement during declared volatile windows.

Composition

Transition mutation composes with multi-model arbitration through the phase-bounded weight schedule: the arbiter consumes each model's current weight cap and respects it during aggregation, so the arbiter automatically reflects the lifecycle of its constituent models. It composes with confidence governance through the advancement predicate, which consumes the trust slope and any active attenuation, so a model under attenuation cannot advance to steady until the attenuation is cleared. It composes with the differential alarm subsystem because each transition emits a structured event that the alarm subsystem treats as a known disturbance, suppressing spurious alarms that would otherwise be caused by the transition itself.

The mutation auditor composes with the lineage writer by emitting its diff records as structured fields rather than free-form logs. Auditors at any future time can reconstruct the full sequence of mutations a model has undergone, the policy under which each mutation was admitted, and the bounds in force at every cycle the model contributed to an output. This makes the cognitive system reproducibly accountable for the evolution of its inference machinery, not only for the outputs the machinery produced.

Prior-Art Distinction

Conventional model deployment treats lifecycle as an operational concern external to the inference loop: models are loaded, served, and unloaded by infrastructure code, and the inference runtime is unaware of their phase. Conventional online-learning systems update parameters continuously without materializing the updates as discrete, audited transitions. Transition mutation is distinguished from both by its first-class phase enumeration, its structured transition records, its declared mutation taxonomy enforced by an auditor, and its phase-bounded runtime behavior that ties admissible outputs and arbiter weights directly to the recorded phase of each contributing model.

Implementation Considerations

Transition timing is a structural concern. A transition that takes effect mid-cycle would produce ambiguity about which phase governed which output, defeating reproducibility. The disclosure contemplates that all transitions take effect at cycle boundaries: a transition record is committed during a cycle but the phase change applies only to subsequent cycles, and the cycle in which the record was committed is annotated with both the prior and the pending phase so that lineage is unambiguous.

Snapshot management for the mutation auditor requires care. Pre-transition snapshots must be captured before any change is applied, must be persisted under the same lineage discipline as ordinary outputs, and must be retained for at least the audit window declared in policy. The disclosure contemplates a snapshot identifier scheme under which each snapshot is content-addressed so that diffs can be reproduced exactly even after long retention windows. The auditor consumes snapshot identifiers rather than snapshot contents, which permits the auditor to be replayed against historical transitions without rehydrating large parameter sets in the common case.

Concurrent transitions across multiple models require sequencing under a declared total order. Two models advancing to steady on the same cycle must each see a consistent set of peer phases when their advancement predicates are evaluated. The disclosure contemplates a transition scheduler that serializes advancement decisions per cycle, evaluates predicates against the pre-cycle snapshot of peer phases, and commits transition records under the serialized order. This guarantees that no transition is admitted on the basis of a peer phase that was itself only conditionally present.

Recovery from a rejected transition is itself a structural concern. When the auditor rejects a proposed transition, the model's prior phase remains in force and the rejection is committed to lineage with the diff record that produced it, the policy version under which the rejection was issued, and the snapshot identifiers that were compared. The disclosure contemplates a rejection escalation path under which a model that accumulates repeated rejections within a declared window is automatically advanced into the hold phase pending external review, rather than continuing to operate at its prior phase indefinitely. This prevents a model that is repeatedly attempting forbidden mutations from operating as if nothing had been attempted, while still permitting recovery through an explicit, auditable channel.

The phase enumeration and the mutation taxonomy are held stable across implementations and are the canonical interface between the transition-mutation subsystem and adjacent subsystems that bound their behavior on phase. A change to either is itself a structural mutation and is subject to audit under the cognition architecture's broader mutation discipline. This is what permits the lifecycle machinery to be replaced or extended without disturbing the arbiter, the confidence governance subsystem, or the lineage writer that consume its output.

Disclosure Scope

The disclosure covers the declared phase enumeration, the structured transition record, the dwell and advancement parameterization, the mutation taxonomy and auditor, the phase-bounded weight schedule, the snapshot management discipline, the transition scheduling discipline, the alternative embodiments enumerated above, and any system that reproduces the structural relationship between declared model phase, audited mutation, and phase-bounded runtime behavior. Implementations that reproduce this structural relationship fall within scope regardless of the specific phase enumeration adopted, the specific diff function selected by the auditor, or the specific weight schedule declared. The disclosure further covers any system that materializes lifecycle phase and audited mutation as first-class fields consumed by adjacent inference and governance subsystems, regardless of whether transition mutation is implemented as a discrete component or as a function distributed across the cognitive architecture.