Affect-Modulated Inference Admissibility

Mechanism

Affect-admissibility sits at the intake of the agent's inference pipeline, between the perception layer and the semantic admissibility gate. Each candidate input, whether it originates as a user message, a tool result, an internal recall, or a peer agent's output, is annotated with two scalar affect descriptors before it reaches the gate. Valence captures the polarity of the input on a positive-to-negative axis; arousal captures its activation intensity on a low-to-high axis. The descriptors are computed by a deterministic estimator that operates on the surface form and metadata of the input, not on the agent's introspective state, so that an audit can reproduce the descriptors exactly from the raw input.

A two-dimensional admissibility map then selects an operating regime for the gate. Inputs that fall in the low-arousal, neutral-or-positive-valence quadrant are routed through the standard regime, in which the gate applies the agent's default fan-out, default depth, and default audit verbosity. Inputs that fall in the high-arousal, low-valence quadrant, which corresponds structurally to threat, distress, hostility, or urgency signals, are routed through an elevated-audit regime. This regime narrows the fan-out so that the agent considers fewer parallel continuations at a time, reduces the maximum depth of any single inference branch, lowers the rate at which the agent commits transitions, and raises the granularity of lineage capture so that every gate decision in the cycle is recorded with its full input context.

Intermediate quadrants are handled by interpolation between the standard and elevated regimes. The interpolation is monotonic in arousal and inverse-monotonic in valence, which is to say that increasing arousal at fixed valence never relaxes the gate, and increasing valence at fixed arousal never tightens it. This monotonicity property is structurally enforced and is independently auditable: a regression test can sweep the descriptor space and confirm that no operating point produces a regime laxer than its less-arousing neighbour.

Critically, affect-admissibility does not override the agent's hard governance criteria. The standard semantic admissibility gate, the integrity constraints, the trust slope, and the policy reference all continue to apply with their normal authority. Affect-admissibility tunes how the gate looks at an input; it cannot make the gate accept something the policy forbids, and it cannot make the gate reject something the policy permits at the standard regime. The mechanism modulates stringency within a permitted band; it does not replace the deterministic floor.

Operating Parameters

The estimator surface declares which features of an input contribute to its valence and arousal scores and with what weights. Operators may swap estimator implementations through policy without changing the rest of the pipeline; the contract is the descriptor pair, not the algorithm that produces it. Each estimator must be deterministic with respect to its declared input set and must publish a calibration table that maps representative inputs to their expected scores so that drift can be detected.

The regime map declares the fan-out multiplier, depth multiplier, commit-rate multiplier, and audit-verbosity level for each cell of the descriptor space. A typical configuration uses a coarse partition such as a three-by-three grid, but the mechanism does not constrain the granularity. The map must be monotonic in the directions described above; a policy linter rejects any candidate map that violates monotonicity before it can be deployed.

The reaction-cadence parameter governs the minimum interval between successive committed transitions in the elevated regime. By stretching this interval, the operator buys reaction-time headroom that downstream supervisors or human reviewers can use to interrupt or shape the agent's response. A configurable hysteresis prevents oscillation between regimes when an input sits near a regime boundary; once an input has placed the gate into elevated audit, the gate remains elevated for a declared dwell time before it can return to standard.

The final parameter is the override channel: a structurally distinct path by which an authorized supervisor can force the gate into elevated audit irrespective of the descriptor scores. Override events are themselves logged and are subject to the agent's lineage rules, so that an operator who applies an override is accountable in the same audit framework that holds the agent accountable for its inferences.

Alternative Embodiments

A conversational-agent embodiment uses linguistic affect estimation: lexical, syntactic, and prosodic features feed the descriptor estimator, and the elevated regime narrows the agent's response generation to deliberate, audited completions when a user expresses distress or hostility. The agent does not refuse to engage; it engages with greater care and with a shorter forward horizon.

A perception-driven robotics embodiment uses sensor-derived affect proxies: rapid scene change, proximity events, and acoustic salience contribute to the arousal score, while context categorization contributes to the valence score. The elevated regime narrows the action fan-out so that the robot considers fewer simultaneous candidate actions and applies a slower commit cadence, providing more time for human override in high-stakes moments.

A multi-agent embodiment propagates affect descriptors across the trust boundary between cooperating agents. A peer agent's output that arrives with elevated arousal and depressed valence raises the receiving agent's regime even if the local estimator would have classified the content as neutral, ensuring that affective state is not laundered through indirection.

A degenerate embodiment, useful for compliance demonstration, maps the entire descriptor space to the elevated regime, producing an agent that always operates with narrow fan-out and elevated audit. This embodiment is structurally identical to the general mechanism with its modulation surface collapsed to a constant, and is admissible under the disclosure when the operator's risk posture warrants it.

Regime Propagation and Decay

Affect-admissibility is not a one-shot decision per input but a regime that persists across a window of inference cycles. When the gate enters the elevated regime in response to a high-arousal low-valence input, the runtime sets a regime-active flag with a declared dwell time. Subsequent inputs that arrive during the dwell are evaluated under the elevated regime regardless of their own descriptors, which prevents an adversary from interleaving neutral filler with hostile content to wash out the regime. The dwell time is policy-declared and is itself logged so that operators can audit how long the agent spent in elevated audit and what triggered each entry.

When the dwell elapses, the runtime applies a decay schedule rather than a hard transition back to standard. The decay schedule reduces the elevated multipliers along a declared curve over a declared interval, so that the agent does not snap from cautious to nominal in a single cycle. Operators may choose linear, exponential, or staircase decay, and the choice is part of the policy artefact. During decay, any new high-arousal low-valence input resets the dwell timer to its full value, ensuring that sustained stress keeps the agent in its careful posture without operator intervention.

The mechanism also exposes a manual de-escalation channel by which a supervisor can shorten the dwell or accelerate the decay when out-of-band context indicates that the affective signal was a false positive. De-escalation events are logged with the supervisor's identity and the asserted justification, so that an audit can distinguish automatic regime transitions from supervised ones.

A complementary escalation channel permits a supervisor or peer agent to force the gate into elevated audit even when descriptor scores would not warrant it, with the dwell timer set to the supervisor's declared interval. Escalation and de-escalation are symmetric in their auditability: every manual transition is recorded with the same fields and the same retention requirements as every automatic transition, ensuring that the regime history is fully reconstructible whether the agent shifted itself or was shifted from the outside by an authorized principal.

Composition

Affect-admissibility composes with the semantic budget mechanism by sharing the fan-out and depth multipliers: the elevated regime spends from the same budget as the standard regime, and a high-arousal low-valence cycle therefore exhausts the budget faster, which itself enforces a slower cadence even before the explicit reaction-cadence parameter is applied. The two mechanisms reinforce each other rather than overlapping.

It composes with the agent's lineage and audit infrastructure by labelling each gate decision with the active regime, so that any post-hoc review can distinguish decisions that were made under standard conditions from decisions that were made under elevated stringency. It composes with the dream state by carrying the same descriptor estimation into off-line forecasting: dreams of high-arousal scenarios are themselves rehearsed under the elevated regime, ensuring that dreamed plans for stressful situations inherit the same caution that waking inference would apply.

Prior-Art Distinction

Sentiment classification, content moderation, and emotion-aware dialog systems are established techniques. Affect-admissibility is not a claim over sentiment scoring or over post-generation filtering. It is a claim over a structurally located gate whose stringency parameters, but not its acceptance criteria, are modulated by deterministic affect descriptors at intake, before the inference pipeline expands.

Conventional emotion-aware systems either route inputs to different model heads, alter the agent's policy at runtime in ways that an auditor cannot easily reconstruct, or apply moderation only to outputs after they have been generated. Affect-admissibility modulates only the working parameters of the existing gate, preserves the deterministic governance floor, enforces monotonicity across the descriptor space, and produces lineage-grade audit records of regime selection. This combination distinguishes the mechanism from prior approaches that conflate affect-awareness with policy mutation or with output-only filtering.

Disclosure Scope

This article discloses affect-admissibility as a structural intake mechanism of the cognition patent's inference-control surface. The disclosure is independent of the specific affect estimator employed and applies to any deterministic descriptor producer that yields a valence-arousal pair from declared input features. The disclosure covers conversational, perception-driven robotics, multi-agent propagating, and degenerate-constant embodiments, and the regime-map shape described is illustrative rather than enumerative.

Implementations that allow affect to override the deterministic governance floor, that violate the monotonicity property across the descriptor space, or that omit lineage capture of regime selection fall outside the scope of the mechanism as disclosed. Implementations that preserve the floor, the monotonicity, and the audit are within scope regardless of how their estimators or regime maps are realized.