ControlVector Logging (four-signal control adjudication, recommendation B)

Subject: telemetry.control_vector_logging Status: IMPLEMENTED Registered: 2026-06-07 Depends on: MECH-320 (tonic vigor), MECH-313 (noise floor), SD-032b dACC (EVC effort term), E3 selector Blocks (informs): ARC-068 first-class opportunity-cost separation (post-green-board)

Problem

A four-signal control adjudication (2026-06-07) asked whether REE-v3 realises four separable control signals:

  1. Value (V_outcome) – is this outcome worth obtaining/avoiding?
  2. Effort cost (C_effort) – how much work/control must be paid?
  3. Opportunity cost of time (C_time) – what is lost by not acting / by no-op delay?
  4. Vigor / activation (G_vigor) – how forcefully/quickly to act?

Code adjudication found that opportunity-cost-of-time is not an independent control signal. It is collapsed into MECH-320’s single average-reward-rate scalar v_t: tonic_vigor.compute_score_bias applies -w_action*v_t to action candidates (vigor) and +w_passive*v_t to no-op candidates (opportunity cost). Both halves are w * v_t for the same v_t (see ree_core/policy/tonic_vigor.py docstring: “mathematically equivalent to a single signed scalar applied to the action-vs-no-op contrast”). ARC-068 (opportunity_cost_no_op_penalty) is a registered claim with no implementation – its lit-pull is still pending. So the two signals cannot dissociate in direction.

The remaining signals: V_outcome is the primary E3 score axis (fully realised); C_effort is present as the Shenhav EVC form in the dACC bundle (mode_ev = payoff - control_required * candidate_effort) but with a thin horizon-length effort proxy; G_vigor is the action half of MECH-320 plus the MECH-313 noise-floor temperature lift.

Solution (recommendation B: logging only, no causal change)

A read-only, default-OFF telemetry struct that makes the four signals separately inspectable and exposes the C_time<->G_vigor collapse by logging both halves alongside their shared v_t. Pure telemetry – no change to scoring or selection.

  • Config: REEConfig.use_control_vector_logging (default False; also a from_dims kwarg). Bit-identical when off (the assembly block is skipped).
  • dACC bundle (ree_core/cingulate/dacc.py): two additive keys – control_required (float) and effort_term (control_required * candidate_effort, [K]) – so C_effort is readable without re-deriving it from payoff - mode_ev. No existing consumer reads them.
  • E3 selector (ree_core/predictors/e3_selector.py): self.last_raw_scores – the pre-bias per-candidate scores (the value axis). Stored alongside last_scores.
  • Agent (ree_core/agent.py): the MECH-320 tv_bias action/no-op split + v_t / w_action / w_passive are cached in the tonic-vigor block; after e3.select returns, REEAgent._assemble_control_vector() writes self._last_control_vector. Read directly by experiment scripts (the V3-EXQ-571 _last_score_bias_decomp pattern).

_last_control_vector schema 

{
  "step": int,
  "V_outcome": {mean, range, std, value_mean(=-mean), present},     # E3 pre-bias scores
  "C_effort":  {mean, range, std, control_required, present},        # dACC EVC effort term
  "C_time":    {potential(=w_passive*v_t), realised_mean, n_noop_candidates, present},
  "G_vigor":   {potential(=w_action*v_t), realised_mean,
                noise_floor_temp_lift, n_action_candidates, present},
  "shared":    {tonic_vigor_v_t, tonic_vigor_v_raw, w_action, w_passive, collapse_note},
  "authority": {modulatory_authority_active, modulatory_authority_scale_factor,
                e3_raw_score_range_mean},
}

shared.tonic_vigor_v_t is logged explicitly so that C_time.potential and G_vigor.potential are both computable as w * v_t for one scalar – the collapse made inspectable. authority records whether the modulatory bias actually had selection authority this tick (modulatory-bias-selection-authority substrate), because MECH-320’s bias has near-zero authority until V3-EXQ-643a validates.

Validation

  • Smoke + 4 contract tests (tests/contracts/test_control_vector_logging.py): C1 default-OFF no population; C2 ON populates the four signals; C3 collapse relationship (C_time/G_vigor == w*v_t, same v_t); C4 bit-identical OFF-vs-ON action stream.
  • Stage-B diagnostic experiment: re-run a MECH-320-ON config with logging enabled and measure the rank-correlation between logged C_time and G_vigor across ticks. Pre-registered prediction: rho ~ 1.0 (they are w*v_t for one v_t). This converts the collapse from a code-reading inference into a measured fact – the evidence governance needs before deciding whether ARC-068 should get its own scalar.

Deferred (post-green-board roadmap)

  • C – a first-class opportunity-cost signal causally modulating commit pressure. Blocked on (a) the pending ARC-068 lit-pull and (b) MECH-320 regaining selection authority (V3-EXQ-643a / SD-056 / ARC-065 GAP-A). The MVT patch-leaving paradigm is the validating falsifier.
  • D – a full four-axis causal ControlVector. Larger; deferred further.

MECH-320 (tonic vigor / the collapse site), ARC-066 (parent family), ARC-068 (the unbuilt opportunity-cost claim), MECH-313 (noise floor / G_vigor temperature lift), SD-032b dACC (C_effort source), modulatory-bias-selection-authority (authority context), V3-EXQ-571 (_last_score_bias_decomp pattern this extends), V3-EXQ-624a (the documented MECH-320 v_t sign/scale issue this telemetry surfaces).

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