Computable Therapeutic Dosing for Cognitive Disruption

Mechanism

A therapeutic dosing course begins with a diagnosis: the disruption-modeling layer identifies a cognitive parameter or parameter family whose current value lies outside the regime appropriate to the agent's environment. The diagnosis names the parameters to be adjusted, the direction of adjustment, and the magnitude of the gap between the present regime and the target regime. From the diagnosis, the dosing primitive constructs a treatment plan: a sequence of governance-bounded sessions, each of which applies a bounded adjustment to the identified parameters and records the post-session response.

Each session is itself a governed interaction. The session is invoked under a credential that names the prescribing authority, the plan it implements, and the session's position within that plan. The applied dose — the magnitude of the parameter adjustment — is bounded both by the plan's per-session ceiling and by an absolute safety ceiling declared at plan instantiation. After the adjustment is applied, the session records the agent's measured response across the planning, behavioral, and affective channels that the plan designates as relevant. The response record is appended to the treatment lineage and becomes input to the next session's dose calculation.

Tapering is not optional. Every plan declares a withdrawal schedule by which the per-session dose is reduced toward zero before the plan terminates. The withdrawal schedule prevents rebound disruption — the phenomenon in which abrupt cessation of intervention permits the original disruption pattern to reassert itself, often more sharply than before treatment began. The withdrawal phase is a governed phase like any other: each tapering session is logged, its response is measured, and the plan can be re-extended if the measured response indicates that further consolidation is required.

Operating Parameters

The principal operating parameters of a dosing plan are dose magnitude, session frequency, plan duration, withdrawal schedule, and contraindication set. Dose magnitude is expressed in the units appropriate to the parameter under intervention — a probability shift, a weight delta, a confidence-interval width — and is bounded above by both per-session and lifetime ceilings. Session frequency is declared as a minimum interval between sessions, ensuring that the agent has the opportunity to express the cumulative response before the next adjustment is applied. Plan duration is declared as a maximum total course length and as a minimum stability interval before the plan can be considered complete.

The withdrawal schedule is parameterized by a tapering function — typically linear, exponential, or step-wise — and by a withdrawal-completion criterion that determines when the plan is fully concluded. The contraindication set declares affective states, environmental conditions, and concurrent-intervention combinations under which a scheduled session must be deferred or escalated to supervisory review. Each parameter is signed at plan instantiation and any mid-course modification is itself a credentialed event admitted against the plan's governance rules.

Alternative Embodiments

A first embodiment applies the dosing primitive to a single cognitive parameter — for example, the prior probability the agent assigns to a particular environmental hypothesis — with monotonic adjustment toward the target regime. A second embodiment applies it to a parameter family with coupled adjustments, where each session distributes a budgeted total dose across several related parameters according to a declared coupling matrix. A third embodiment supports parallel plans operating on disjoint parameter sets, with a contraindication rule that prevents two plans from converging on the same parameter family within a declared interval.

Response measurement also admits embodiments. A pure-planning embodiment measures only the agent's subsequent planning behavior. A behavioral embodiment additionally measures the agent's externalized actions in its environment. An affective-monitoring embodiment, contemplated by composition with the affective-state monitoring primitive, additionally measures the agent's reported or inferred affective state and gates further dosing on the affective response.

Composition

Therapeutic dosing composes most directly with affective-state monitoring. The affective-monitoring primitive supplies a parallel signal that the dosing plan can reference both as a contraindication input and as a response measure. A session whose contraindication set names a specific affective state will defer execution if monitoring reports that state at the scheduled time; a plan whose response criterion includes affective stabilization will not declare completion until the affective channel reports within-bounds for a declared stability interval.

Dosing also composes with the forecasting engine that governs all agent planning. The same forecasting machinery that the agent uses to project outcomes of its own actions is used by the dosing primitive to project the trajectory of the treatment course: given the current response history, the engine forecasts the dose schedule most likely to produce the target regime within the declared duration, and the plan is adjusted accordingly. This unification means that the treatment course is governed by the same evidentiary discipline as the agent's substantive planning, rather than being a privileged side channel.

Prior Art

Reinforcement-learning literature treats parameter adjustment as a continuous optimization driven by a reward signal, without governance bounds, without explicit tapering, and without an audit lineage suitable for clinical-grade review. Online-learning frameworks similarly adjust parameters at every step but do not distinguish between substantive and therapeutic adjustments and do not preserve the diagnostic-to-response chain. Pharmaceutical dosing protocols supply the structural vocabulary — dose, frequency, duration, withdrawal, contraindication — but apply to biological rather than computational systems; their direct application to cognitive intervention has not, to the applicant's knowledge, been disclosed in the cognitive-architecture literature.

The disclosed primitive imports the structural vocabulary of pharmaceutical dosing into the governance discipline of computable agency, producing a treatment course that is bounded, auditable, and composable with affective-state monitoring and with the agent's native forecasting engine.

Implementation Notes

Several implementation considerations bear on the realization of the primitive. The dose-magnitude ceiling must be expressed in the same units as the parameter under intervention, and the architecture must enforce that ceiling at the moment of session execution rather than at the moment of plan instantiation; a session whose computed dose exceeds the lifetime ceiling because of cumulative prior adjustments must be either deferred, escalated, or trimmed to the remaining headroom, and the chosen disposition must itself be a credentialed plan parameter. The contraindication evaluation must be performed at session start rather than at plan instantiation, since the contraindicated affective or environmental conditions are by definition not known in advance.

Response measurement is a non-trivial engineering concern. The measurement window — the interval over which post-session response is integrated — must be long enough to capture the agent's settled response and short enough that the next session can be informed by it. Different parameter families require different measurement windows; the plan declares the window per session type. The architecture exposes the measurement-window parameter and the response-aggregation function as tunable plan attributes, signed at instantiation and modifiable only through credentialed plan amendment.

Forecasting-engine integration requires that the engine be able to project not only the agent's substantive planning trajectory but also the agent's response trajectory under the proposed dose schedule. This requires that the engine's state representation include the cognitive parameters under intervention as first-class state variables, and that the engine's transition model include the dose-response relationship for those parameters. Where the dose-response relationship is poorly characterized, the engine operates with elevated uncertainty and the plan is correspondingly conservative; as the response lineage accumulates, the relationship is refined and the engine's uncertainty contracts.

Adversarial Considerations

The credentialing discipline applied to plan instantiation and session execution addresses several adversarial scenarios. A rogue-prescriber scenario, in which an unauthorized actor attempts to instantiate a treatment course, fails the plan-instantiation credential check and produces a non-admission record naming the failed credential. A dose-injection scenario, in which an adversary attempts to apply an out-of-plan adjustment, fails the per-session credential check that binds each session to its parent plan. A response-falsification scenario, in which an adversary attempts to write a fabricated response record into the lineage, is bounded by the response-record signature, which is produced by the architecture's own measurement subsystem rather than by the prescribing authority.

Operational Scenarios

A representative scenario clarifies the primitive's operation. An autonomous planning agent operating in a logistics environment has, over an extended period, accumulated a strong prior toward a particular routing strategy that no longer matches current conditions; the disruption-modeling layer diagnoses the prior as the parameter requiring intervention. The dosing primitive instantiates a plan that adjusts the prior by a small fraction per session, with sessions occurring no more frequently than once per planning cycle. Each session records the adjustment, the agent's subsequent planning behavior, and the affective-state signal — in this context, the agent's confidence and frustration metrics — observed during the cycle. The plan declares a tapering schedule that halves the per-session adjustment over the final third of the course and a contraindication that defers any session occurring during a high-frustration interval.

Across the course, the lineage accumulates a complete record: each session's credential, dose, response, and affective context. A subsequent reviewer — whether a supervisory authority, a clinical-grade auditor, or the agent's principal — can reconstruct the full treatment, compare its trajectory against contemporaneous courses applied to similar agents, and either ratify the course's outcome or, if the response record indicates an inadequate or harmful trajectory, intervene with a corrective plan. The audit-grade lineage is what distinguishes therapeutic dosing from ad-hoc parameter tuning: the same adjustment, applied without the credentialed plan structure, produces no comparable evidentiary record and no comparable basis for review.

Disclosure Scope

The disclosure covers the representation of cognitive intervention as a parameterized treatment course, the per-session governance discipline applied to each adjustment, the response-recording lineage produced by the course, the withdrawal schedule that terminates the course, and the contraindication set that gates session execution. The disclosure further covers composition with affective-state monitoring and with the agent's forecasting engine, and the application of the primitive to operational planning agents whose treatment courses must be reconstructable for clinical-grade review. Specific tapering functions, response-measurement modalities, and parameter-family coupling matrices are implementation choices within the disclosed framework.