MECH-353: affect.blocked_agency_control_failure_stream (z_block)

Claim ID: MECH-353 Subject: affect.blocked_agency_control_failure_stream Status: IMPLEMENTED 2026-06-06 (substrate; v3_pending until the discriminative experiment PASSes) Registered: 2026-06-05 (lit-pull targeted_review_blocked_agency_anger_stream, chip task_64c2e558) Depends on: SD-029 (agency comparator), MECH-112 (z_goal / wanting), MECH-320 (tonic vigor / assert pole), MECH-342 (decommit actuator), ARC-016 (commitment-threshold gate), SD-011 (differentiated-from: harm), SD-019b (differentiated-from: suffering). Detector discrimination additionally requires SD-056 (action-conditional world_forward) trained. Blocks: the smallest-V3 blocked-action discriminative experiment; clears its own v3_pending on PASS.

This doc records the V3 substrate implementation. The affect-register framing (the “why”, the full neighbour-differentiation grid, and the literature) lives in docs/architecture/affect_primitives.md (Extension Register, blocked_agency). This file is the implementation contract.

Problem

REE represents harm (SD-011, z_harm_*), suffering (SD-019b z_harm_a + Q-036 escapability — the capacity-collapsed withdraw pole), residue (MECH-056), and licit self-imposed commitment-holds (MECH-090). It does not represent the affect that arises when an intended, predicted-to-succeed action is repeatedly BLOCKED while the goal and the agent’s capacity-belief are retained — the energised “assert / restore” pole. The lit-pull verdict (Davis & Montag 2019 RAGE as a distinct primary-process system; Papini 2024 frustrative non-reward, zero noxious input; Bertsch 2020 reactive-aggression assert channel under prefrontal gating; Carruthers 2012 comparator model) establishes this is a distinct stream, not reducible to harm + residue + frustration. Folding it into the harm register would repeat the SD-010 -> SD-011 philosophy-right / mechanism-wrong error (feedback_biology_before_formal_definitions).

Solution

z_block is a derived affect readout — a pure-arithmetic regulator (ree_core/affect/blocked_agency.py, mirroring the MECH-313 / MECH-320 / MECH-342 pattern), NOT an encoder output. It is also surfaced as LatentState.z_block ([batch, 1]) for downstream consumers / instrumentation.

Detector / antecedent (computed in REEAgent._update_blocked_agency, called from sense() right after the MECH-095 TPJ update):

  1. Action-outcome comparator (SD-029 on the z_world channel; Carruthers 2012). zw_pred = E2.world_forward(z_world_prev, last_action); pred_mag = ||zw_pred - z_world_prev|| (the intended effect size); outcome_mismatch = ||zw_pred - z_world_now|| / (pred_mag + eps). Gated by a predicted-effect floor: when pred_mag < floor the action was not predicted to produce an effect, so there is nothing to be blocked from (outcome_mismatch = 0; this also fails safe on an untrained world_forward). Calibration with a trained, action-conditional world_forward: blocked -> z_world unchanged -> outcome_mismatch ~ 1.0; success -> z_world reaches the prediction -> outcome_mismatch ~ 0.0. This is FNR’s expected-minus-obtained (Papini 2024) generalised from reward-omission to action-effect-omission.

  2. Expectation (MECH-112). The goal-retained gate (goal_state.is_active()) supplies “there is an intended outcome” the block frustrates.

  3. Motor-agency / external-attribution comparator (z_self channel). motor_agency = 1 / (1 + ||E2.predict_next_self(z_self_prev, last_action) - z_self_now||). High = the motor command executed as predicted -> the mismatch is external, not the agent’s own motor error. (predict_next_self is E2’s primary trained objective, so this gate is reliable even when world_forward is weak.) This is the load-bearing distinction the verdict emphasises.

Integration (BlockedAgency.update). z_block accumulates the comparator mismatch over a window only when motor_agency >= attribution_motor_floor (external) and outcome_mismatch >= outcome_mismatch_floor (genuinely blocked) and the goal is retained; it leaks on success. Capacity-gated split: z_block_assert = z_block * capacity_belief, withdraw_handoff = z_block * (1 - capacity_belief), where capacity_belief = clip(1 - w * ||z_harm_a||, 0, 1) (suffering collapses capacity-belief, so the assert pole yields to the existing withdraw pole).

Consumers (in select_action):

  • ASSERT / escalate-effort / alternative-action search (the new behavioural pole REE lacks; the “raise MECH-320 vigor” effect). A self-emitted per-candidate score-bias composed into dacc_score_bias after the tonic_vigor block: NEGATIVE on action-trajectories (favour acting), POSITIVE on no-op (penalise passivity), extra POSITIVE on the just-blocked first-action class (try a different action). Scaled by z_block_assert; zero when there is no asserting block this tick (so OFF / no-block ticks add nothing — bit-identical).
  • DECOMMIT (MECH-342) gated by ARC-016. When assertion fails across the window (z_block_assert sustained above decommit_bound for decommit_consecutive_ticks), update() emits a decommit_signal. select_action consumes it — beside the MECH-342 maintenance-release site — to beta_gate.release(), gated by the ARC-016 commitment-threshold (e3.current_precision <= blocked_agency_decommit_arc016_precision_max; <= 0 disables the gate). Releases the blocked commitment rather than escalate unboundedly (Bertsch 2020 prefrontal inhibition of reactive aggression).
  • HANDOFF to suffering. As capacity_belief falls the assert share decays and withdraw_handoff rises; the existing SD-019b / z_harm_a / Q-036 withdraw machinery takes over. No forced z_harm_a write — the assert pole simply yields.

Config (all no-op default; bit-identical OFF)

REEConfig + from_dims: use_blocked_agency (master) plus blocked_agency_{accumulation_rate, leak_rate, outcome_mismatch_floor, attribution_motor_floor, capacity_collapse_weight, require_goal_active, z_block_cap, assert_action_weight, assert_passive_weight, assert_alt_action_weight, assert_bias_scale, decommit_bound, decommit_consecutive_ticks, decommit_arc016_precision_max, predicted_effect_floor, noop_class}.

Env manipulation (CausalGridWorldV2, SD-029 / SD-022 env-curriculum precedent; env-only kwargs, not surfaced through from_dims): scheduled_action_block_enabled (+ _interval, _prob). On fire, the agent’s chosen move is externally cancelled (the agent stays put) for that tick with no damage and no layout change — transition_type="action_blocked", info tags action_blocked_this_step / action_block_event_count. This is the clean external-block antecedent (motor intact -> the attribution gate reads it as external), distinct from SD-022 limb-damage movement failure (own-motor error coupled to damage).

Backward compatibility / training notes

  • use_blocked_agency=False -> agent.blocked_agency is None, LatentState.z_block stays None, no consumer fires; verified bit-identical action stream (the regulator uses no torch RNG and adds zero bias when z_block~0). 803/803 ree-v3 contracts pass with the master OFF; 9/9 new MECH-353 contracts pass.
  • No phased training (pure-arithmetic regulator; no learned parameters).
  • MECH-094: update() is a no-op under simulation_mode / hypothesis_tag (replay must not accumulate blocked-agency on imagined outcomes); z_block is left None on hypothesis-tagged latents.
  • Detector discrimination requires a trained substrate. At smoke / untrained scale, z_world deltas do not track single-cell moves (the encoder is random; realised effect is uninformative), so the action-outcome comparator is uninformative — expected. The validation experiment trains the encoder (representing scene/position in z_world) and the action-conditional world_forward (SD-056) in P0, then measures block-vs-control discrimination in P1. If a trained substrate still cannot make the comparator discriminate, that is an informative substrate-ceiling finding routing to encoder / world_forward enrichment — NOT a falsification of MECH-353’s affective claim.

What this enables

The smallest-V3 blocked-action discriminative experiment: an env that repeatedly blocks an intended, predicted-to-succeed action while harm and goal-value are held constant; measure whether z_block rises, whether it drives assert/persist (effort escalation / alternative-action search) distinct from the withdraw signature, and whether it dissociates from z_harm_a under a matched controllability manipulation. PASS clears v3_pending on MECH-353.

Scope

This is Stream A (V3-tractable, single-agent). Stream B (coercion / domination / injustice) needs a model of an other-agent coercer (psychological reactance; Steindl 2015) and is V4-social, deferred — NOT implemented here.

SD-029, MECH-112, MECH-320, MECH-342, ARC-016, SD-011, SD-019b, Q-036, MECH-056 (residue, differentiated-from), MECH-090 (commitment-hold, differentiated-from), MECH-095 (TPJ comparator, sibling agency signal on z_self), SD-056 (action-conditional world_forward, detector prerequisite), MECH-094 (simulation gate).

REE is developed by Daniel Golden (Latent Fields). Apache 2.0.