State-qualified intervention and withdrawal

Scope

This note formalizes the held pharmacology/toxicology transfer: commanded intervention, realized internal state, mechanism engagement, benefit, harm, adaptation, and withdrawal remain separate. It supports the pharmacology audit and Candidates 005, 007, 012, and 014. The evidence and boundary claims are C-607–C-626.

Intervention chain

For system state $x_t$, commanded intervention $d_t$, realized internal exposure $e_t$, adaptation state $z_t$, and observation $y_t$, use

Here $c_t$ is typed context, $s_t$ is sampling/selection state, $v$ is the observation-model version, and every disturbance $w_t^\cdot$ and error $\epsilon_t$ has the unit of the equation in which it appears. The command $d_t$ and exposure $e_t$ need not share a unit: a token quota, update rate, or tool permission is not evidence that the intended computation, parameter change, or external effect occurred. Forecast value, feedback correction, integrated action exposure, reserve debit, and later outcome remain separate coordinates for the same reason (C-1494).

The evaluator reports endpoint vectors rather than a hidden scalar:

where $\mathbf b_t$ is task benefit, $\mathbf h_t$ is protected harm, $\mathbf r_t$ is remaining reserve, $E_t$ is lifecycle energy in joules, $L_t$ is latency in seconds, $M_t$ is moved/stored information in bytes, and $H_t$ is human effort in person-seconds. Unlike coordinates are never added without a declared decision rule and weights.

Qualified authority region

An intervention is admissible only inside a versioned region

where $u_t$ is the complete uncertainty record, $\tau_t$ is observation age in seconds, and $p_t$ identifies the supported population/task/context. Calling $\mathcal D_t$ a “window” does not make it a fixed scalar interval. It changes with endpoint, schedule, context, uncertainty, adaptation, reserve, and the cost of delayed harm. A local rule must also be requalified when its trigger becomes widespread and its shared-system cost changes (C-1493).

Withdrawal is its own transition

Removing support is not algebraic negation of adding support. Let $\Delta d_t=d_t-d_{t-1}$. A reduction policy must satisfy

where $\mathcal R_v$ is a rate-qualified removal set. After $d_t$ reaches zero, evaluation continues for declared horizon $T_W$ seconds or task events:

This exposes dependence hidden by supported performance. A rollback is not complete when the command disappears; it is complete only when rebound, recurrence, native capability, reserve, and observation coverage meet their declared postconditions.

Interaction and schedule tests

For interventions $a$ and $b$, interaction is the residual against a named reference $g_0$:

$I_{ab}$ inherits the endpoint unit. Its sign can change with response scale, schedule, context, and endpoint, so “synergy” without $g_0$ is undefined. The test matrix must also compare schedules with equal cumulative command but different peaks, spacing, duration, and recovery; matched totals do not imply matched internal state.

Falsification boundary

The contract adds nothing if a Bayesian state-space PK/PD analogue, mixed-effects estimator, constrained MPC/POMDP, calibrated harm monitors, and staged decommissioning/taper policy match its benefit–harm–reserve–cost frontier. Universal hormesis, one-number therapeutic windows, dose-as-effect, and absence-of-observed-harm claims fail before this comparison.