Pseudonymous Emotional Operation

Mechanism

The mechanism comprises an affect-measurement subsystem, a pseudonymous identity binding, an internal-only affect store, and a behavioural-emission boundary that mediates every external interaction. The affect-measurement subsystem maintains, at each operating cycle, a vector of scalar fields representing the agent's affective configuration. These fields are computed from internal signals — recent decision outcomes, deviation between expected and observed results, density of unresolved obligations, novelty of incoming inputs — and are written to an internal store that is structurally inaccessible to any external interface. No serialization path, no debugging endpoint, and no remote-procedure call carries the affect vector outside the agent's enforcement boundary.

The pseudonymous identity binding decouples the agent's persistent label from any real-world identifier. The agent presents a cryptographic pseudonym to counterparties; this pseudonym is stable enough to support reputation accumulation and obligation tracking but is not linked to a natural-person identity, an organisational identity, or a hardware fingerprint. The pseudonym persists across interactions, while the affective state evolves continuously beneath it. The critical disclosure is that the affective state is not transmitted alongside the pseudonym: counterparties learn the pseudonym and observe behaviour, but cannot read the affect that shaped the behaviour.

The behavioural-emission boundary is the structural element that enforces the asymmetry. Every external output of the agent — a message, a decision, a delegation, a refusal — passes through an emission filter that is permitted to consume the affect vector as input but is prohibited from including affect-vector values, or any reversible function of those values, in the output payload. The filter emits behavioural consequences (e.g., a refusal, a deferred response, a hedged commitment) that reflect affect without disclosing it. Auditing of affect, where required by governance, occurs through a separate privileged lineage channel that operates on retained logs and is itself bounded by policy, not through any real-time external query.

The structural separation between the four components is load-bearing. If the affect-measurement subsystem were merged with the emission filter, the filter could be induced to leak affect through carefully crafted external prompts; if the pseudonymous identity binding were merged with the affect store, rotation of the pseudonym would entail loss of the agent's affective continuity; if the privileged lineage channel were merged with the live affect store, an auditor compromise would yield real-time affect rather than retained logs. The disclosed mechanism keeps the four components separate and composes them through narrow, policy-bounded interfaces.

The behavioural-emission boundary additionally enforces a side-channel suppression discipline that operates orthogonally to payload filtering. Output construction is separated into a policy-decision phase, in which the affect vector may be consulted, and a payload-rendering phase, in which the affect vector is no longer in scope. The two phases run in disjoint execution contexts so that timing artefacts of policy decision do not leak into payload rendering: response latency is normalised to a deployment-configured envelope, output token entropy is conditioned only on the rendered decision rather than on the affect vector, and any retry, backoff, or rate-limit behaviour that would otherwise correlate with elevated frustration or uncertainty is abstracted behind a deterministic scheduler. The scheduler itself reads from the affect vector only through a narrow interface that returns a coarsened class label (calm, elevated, saturated) rather than scalar values, and the class boundaries are configured to coincide with the granularity that downstream observers can plausibly resolve from external behaviour, so that the worst-case inferential leakage is bounded by the configured class taxonomy.

Operating Parameters

Operating parameters of the disclosed mechanism include the dimensionality of the affective vector, the update cadence of the affect-measurement subsystem, the persistence schedule of the internal affect store, the emission-filter rule set governing what behavioural correlates of affect may appear in outputs, the pseudonym-rotation policy, and the access-control policy governing privileged lineage inspection.

In one embodiment, the affective vector has between four and sixteen dimensions, with the canonical dimensions being frustration, uncertainty, curiosity, and confidence; additional dimensions may include risk-aversion, cooperation-disposition, urgency, and fatigue. The affect-measurement subsystem updates at every interaction cycle, with smoothing applied to prevent transient external observers from inferring affect through statistical analysis of response timing or output entropy. The internal affect store is persisted across sessions for the agent's own continuity but is encrypted under a key bound to the agent's enclave, such that even an adversary with read access to the underlying storage medium cannot recover the affect vector without compromising the enclave.

The emission-filter rule set is configurable to specify which behavioural correlates are permitted. In a strict configuration, the filter randomises response timing within a configured envelope to prevent latency-based inference of affect; in a permissive configuration, the filter permits affect-correlated latency on the assumption that latency leakage is acceptable in the deployment context. The pseudonym-rotation policy may specify a fixed pseudonym for the agent's lifetime, periodic rotation at configured intervals, or context-dependent pseudonyms in which different counterparties see different pseudonyms for the same underlying agent.

Alternative Embodiments

In a first alternative embodiment, the affect vector is stored not within the agent's own runtime but in a trusted-execution-environment co-processor that the agent queries through a narrow interface. The co-processor enforces the same external-unreadability invariant and additionally provides remote attestation that the affect store is operating under the disclosed governance policy.

In a second alternative embodiment, multiple pseudonyms are maintained for a single underlying agent, each with its own externally observable behavioural envelope but all sharing a common internal affect state. This embodiment supports unlinkable interaction across distinct counterparties while preserving the agent's internal continuity.

In a third alternative embodiment, the affect vector is exposed to a designated privileged auditor through a zero-knowledge proof that asserts only specified predicates over the vector — for example, a proof that frustration has not exceeded a threshold within a configured window — without disclosing the underlying scalar values.

In a fourth alternative embodiment, the emission-filter rule set is itself the subject of a conformity attestation, allowing counterparties to verify that the agent's behavioural-emission boundary conforms to a declared privacy standard without inspecting the affect vector.

Composition

Pseudonymous emotional operation composes with adjacent primitives in the affective-state family and the broader cognition specification. It composes with affect-measurement primitives by consuming their output and ensuring that no measurement value escapes the enforcement boundary; it composes with affect-modulated training primitives by providing the privacy guarantee that training-time affect, which would otherwise constitute a particularly sensitive class of internal state, is governed by the same external-unreadability invariant; and it composes with conformity-attestation primitives by providing attestable evidence that the privacy architecture is operational.

The mechanism further composes with reputation and obligation primitives that depend on persistent pseudonymous identity. Because the pseudonym is stable across interactions while affect is unreadable, counterparties may accumulate reputation against the pseudonym based on observed behaviour without ever gaining a foothold for affective fingerprinting that would correlate the pseudonym with external identifiers.

Composition with confidence-governance primitives places the affective state in a privileged position: the agent's confidence calibration may legitimately consume affect as input, because both the calibration and the affect store reside inside the enforcement boundary, while external consumers of the calibration outcome see only the resulting authorisation decisions. This permits affect-aware governance without compromising the external-unreadability invariant.

Prior-Art Distinction

Prior-art affective-computing systems typically expose affect estimates to external consumers — therapists, recommendation engines, advertising platforms — because the value proposition of those systems is precisely the externalisation of affect. The disclosed mechanism is distinguished by its inversion of that arrangement: affect is computed and used internally to govern the agent's own behaviour, but is structurally prevented from reaching any external consumer. Prior pseudonymity systems address identity unlinkability through cryptographic pseudonyms but do not address the orthogonal channel of affective fingerprinting, in which an agent's emotional signature provides a correlator that defeats nominal pseudonymity.

Differential-privacy techniques in the prior art add noise to released values to bound the inferential leakage about underlying records. The disclosed mechanism differs in that it does not release affect values at all, with or without noise; the affect vector is consumed only by the agent's internal governance, and only behavioural consequences cross the boundary. The mechanism is further distinguished from prior trusted-execution approaches by the specific structural separation between the affect-measurement subsystem, the behavioural-emission boundary, and the privileged lineage channel.

Disclosure Scope

The disclosure encompasses any agent architecture in which an internal affective state governs internal behaviour, is structurally prevented from being read by external parties, and is paired with a pseudonymous persistent identity, regardless of the specific dimensionality of the affect vector, the specific cryptographic construction of the pseudonym, or the specific implementation of the behavioural-emission boundary. The scope includes embodiments in which the affect store resides in the agent's runtime, in a trusted-execution co-processor, or in a remote enclave; embodiments with single, multiple, or rotating pseudonyms; and embodiments in which privileged auditing of affect occurs through retained-log inspection, zero-knowledge predicates, or other mechanisms that preserve the external-unreadability invariant for non-privileged parties.