modulatory-bias-selection-authority

Status: IMPLEMENTED 2026-06-03 (substrate-readiness validation pending V3-EXQ). Subject: ethics_engine_3.modulatory_bias_selection_authority Unblocks: MECH-314 / MECH-314a/b/c, Q-044, MECH-320, ARC-068, MECH-341 (and downstream MECH-343 difficulty-gated proposal entropy). Substrate_queue: modulatory-bias-selection-authority (created from the 604a/624a cluster autopsy 2026-06-03).

Problem

The modulatory / diversity score-bias channels at E3 committed selection fire but have zero authority over the committed argmin. Three convergent failure records, one root cause:

Run	Lever	Signature
V3-EXQ-604a	MECH-314 curiosity	`curiosity_bias_abs_mean = 0.0` every arm; `selected_entropy` identical across 5 arms
V3-EXQ-624a	MECH-320 vigor	vigor fires (`v_t=0.05`) but `action_density` byte-identical vigor-ON vs OFF (0.865)
V3-EXQ-614d	MECH-341 within-class temperature	within-class branch fires 3/3 seeds, but committed-class entropy byte-identical (1.056572) across T=0.5/1.0/2.0

Root cause (confirmed at ree-v3/ree_core/predictors/e3_selector.py select()): the modulatory contributions are fixed small magnitudes (curiosity/vigor bias_scale ~0.1, entropy 0.05-2.0) added to primary scores whose raw_score_range is much larger, so they never move the argmin. The MECH-341 across-class stratified_select softmax has the same problem: the absolute class-representative score gap dominates stratified_temperature, collapsing committed-class selection onto one class.

Solution (approach (b): gap-relative scaling)

Give the modulatory layer genuine but bounded authority by scaling it relative to the local primary-score scale — without modifying the primary scores (so commit-threshold / running_variance / softmax-temperature / urgency-interrupt / MECH-090 admission semantics are untouched, and a clearly-harmful candidate with a large primary gap stays rejected).

Three application sites, all gated on one master flag (default OFF, bit-identical):

Additive bias chain + MECH-341 entropy bonus (e3_selector.select()): after the composed score_bias (dACC + lateral_pfc + ofc + mech295 + curiosity + vigor) and the MECH-341 entropy bonus are added, compute the combined modulatory delta mod = scores - raw_scores and rescale it so range(mod) == gain × raw_score_range, then scores = raw_scores + rescaled_mod. Takes precedence over the legacy normalize_score_bias_to_e3_range (the blunt gain=1.0 blob version), which is skipped when this flag is on.
Stratified across-class softmax (e3_score_diversity.stratified_select()): normalize the class-representative scores to unit range before the stratified_temperature softmax, so the diversity temperature acts on a fixed scale (the 614d C2 fix).

Safety property. With gain < 1.0, the modulatory layer is competitive in near-tie regimes (top-2 primary gap < gain × range) but subdominant when the primary harm/goal gap exceeds gain × range. Primary scores are never mutated.

Config (`REEConfig` / `from_dims` / `E3Config`)

Param	Default	Site
`use_modulatory_selection_authority`	`False`	master (E3Config + REEConfig top-level mirror)
`modulatory_authority_gain`	`0.5`	additive-bias rescale target as fraction of `raw_score_range`
`modulatory_authority_min_range_floor`	`1e-6`	degenerate-range guard

The top-level REEConfig mirror lets build_from_ree_config arm the stratified across-class normalization (E3ScoreDiversityConfig.use_selection_authority).

Backward compatibility

Default OFF → bit-identical to pre-substrate. 734/734 contracts + 7/7 preflight PASS with the flag off, verified under two pytest-randomly orderings. Existing normalize_score_bias_to_e3_range is unchanged and independent.

Necessary-but-not-sufficient caveat (validation-shaping)

624a (vigor) and 614d (within-class temperature) are pure drowning — this substrate fixes them directly. 604a had curiosity_bias_abs_mean = 0.0 — the curiosity bias was genuinely ~zero (MECH-314a found no active residue centers; MECH-314b/c are broadcast-by-design), not just drowned. Scaling zero is still zero. The validation EXQ must guard curiosity_bias_abs_mean > 0 before testing curiosity’s authority; otherwise a curiosity arm tests a degenerate upstream signal, not this substrate.

MECH-094 / phased training

Pure arithmetic on the waking committed-selection path; stratified_select already carries simulation_mode. No replay write surface — N/A / preserved. No learned parameters — no phased training.

Validation

Substrate-readiness diagnostic EXQ (via /queue-experiment): re-run the vigor (624a-style), curiosity (604a-style, with the non-degeneracy guard) and within-class (614d-style) arms, master OFF vs ON, measuring whether the previously-inert lever now changes committed selection (action_density lift for vigor; committed-class entropy lift for diversity) without a harm increase. claim_ids=[] (substrate-readiness). PASS unblocks the per-claim evidence retests of MECH-314/320/341 and the MECH-343 hypothesis.

V3-EXQ-643a fix: float32 catastrophic cancellation (2026-06-06)

The 2026-06-03 Solution (b) computed the combined modulatory delta as mod = scores - raw_scores — reconstructing it by subtracting two primary-magnitude score tensors. This is numerically fragile, and V3-EXQ-643 (the substrate-readiness validation) hit the fragility: modulatory_authority_active_frac = 0.0 and scale_factor_mean = 0.0 on both authority-ON arms, 0/3 seeds.

Code-confirmed root cause (corrects failure_autopsy_V3-EXQ-643_2026-06-06). The autopsy concluded “the modulatory bias is uniform across the K candidates … range < 1e-6 … scaling a zero range is still zero.” That is numerically wrong. The MECH-341 entropy bonus keys on per-candidate first-action class (not z_world), so it carries a genuine ~0.17 cross-candidate range — it is not uniform. What actually happened:

The 643 harness trains SD-056 (e2_action_contrastive) online, which drove the primary E3 scores to ~1e32 (z_world instability; the SD-056 multistep rollout-norm clamp is OFF in 643).
The float32 ULP at 1e32 is ~1e25, far above 0.17. So scores - raw_scores collapses to exactly 0.0 → modulatory_range < floor → the gate never fires.

Diagnostic probe (643 config + harness): SD-056-training-OFF → raw_score_range ~4.5, gate fires (modulatory_range ~0.25, active=True); SD-056-training-ON → raw_score_range ~1e32, modulatory_range = 0.0, never fires. The substrate is functionally correct at normal score magnitude; 643 failed from a harness-induced numerical degeneracy, not from absent per-candidate signal.

Fix. Track the combined modulatory contribution explicitly as a small accumulated tensor _modulatory_accum = (score_bias added) + (MECH-341 entropy bonus added), captured at the two add sites. The authority block measures modulatory_total = _modulatory_accum (range from the small ~0.17 bias tensor, independent of primary-score magnitude) instead of reconstructing it by subtraction. Mathematically identical to scores - raw_scores in exact arithmetic; immune to large-score cancellation. scores = raw_scores + scale_factor × mod is unchanged. New diagnostic modulatory_authority_range exposes the true range (0.17, not 0).

Bit-identical OFF (flag default); bit-identical when scores are small (differs only in low-order float bits, argmin unchanged). Regression contract test_modulatory_authority_survives_large_primary_scores (large per-candidate world states → raw_score_range > 1e10; a 0.5-range bias is annihilated by the pre-fix subtraction but the post-fix gate reports modulatory_authority_range == 0.5, active = True).

Validation-shaping (vacuity guard). With the fix, the gate fires correctly even at 1e32 scores — but scale_factor is then ~1e27, a vacuous fire on degenerate scores. A meaningful V3-EXQ-643a therefore REQUIRES the harness to keep primary scores bounded: enable the SD-056 rollout-norm clamp (e2_rollout_output_norm_clamp_enabled) and add a raw_score_range non-vacuity precondition (below-bound → substrate_not_ready_requeue, not an authority verdict). The 604a curiosity non-degeneracy guard is retained.

Route-range amend: route upstream-channel range into the bias the authority rescales (569f/661/654a, 2026-06-10)

The 2026-06-03 / 643a authority rescales _modulatory_accum (the composed score_bias chain + MECH-341 bonus). The 569f / 661 / 654a cluster (failure_autopsy_569f-661-654a_2026-06-10, user-adjudicated) exposed the next link: a channel whose representation carries genuine cross-candidate range still does not move committed action.

Run	Channel	Range present in representation	Committed-action readout
569f (ARC-065)	world-summary	consumed spread 0.196 (> floor); C1 e2.world_forward divergent PASS	selected-action entropy bit-identical 0.549141 across e2wf / proposer / matched-noise
661 (MECH-294)	theta co-binding coherence	coherence JOINT 1.0 / ALT 0.25 / SHUF 0.0 (mode-distinct)	committed-dist TV ~0 (max 0.0097); C3 gate-ON == gate-OFF
654a (ARC-062/MECH-309)	candidate rule field	rules minted 268/549/220	within-arm rule_state counterfactual 0/3; committed-class entropy lift OFF 0.9375 -> ON 0.9377

One structural property, not three bugs: the range exists in the representation but is flattened by the consuming bias head before it reaches _modulatory_accum, so the authority has nothing to amplify. V3-EXQ-643 established no range -> no authority; this cluster extends it one link: the channel range must be ROUTED into the per-candidate modulatory bias the authority rescales, not merely exist in the representation.

The fix (parameter-free, no-op default, bit-identical OFF)

project_channel_range(features) (e3_selector.py): a deterministic, range-preserving projection of a channel’s per-candidate representation into a per-candidate scalar bias [K]. [K, D] (e.g. cand_world_summaries): center across the K candidates, project onto the leading right-singular vector of the centered matrix (SVD on a detached copy; numerical fallback to the mean-deviation axis). [K]: identity. K<2 / flat input: zeroed (below-floor). The singular-vector sign is arbitrary – routing makes the channel range reach and move the committed argmax (the readiness property), not necessarily move it beneficially (that is the channel’s own trained head, the separate per-claim evidence retest).
E3TrajectorySelector.select(channel_route_bias=...): when use_modulatory_channel_routing is on, the [K] routed bias is normalised to unit zero-mean range (bounded even with the authority OFF; the authority re-normalises the combined accumulator to gain * raw_score_range regardless), scaled by modulatory_channel_route_weight, and folded into both scores and _modulatory_accum before the authority’s range computation. New P0 diagnostics: modulatory_channel_route_range (the raw cross-candidate range, pre-normalise / pre-rescale – the gate signal) + modulatory_channel_route_active.
REEAgent.select_action builds channel_route_bias from the channel-under-test’s per-candidate representation and passes it. modulatory_channel_route_source: cand_world_summary (the [K, world_dim] world-summary channel – 569f cluster lead, the genuine projection case) / curiosity / gated_policy / mech295 / coherence (each an already-computed [K] bias, identity-routed). Default none -> bit-identical.

P0 readiness gate (the autopsy’s explicit requirement): modulatory_channel_route_range lets a retest assert the modulatory bias itself carries cross-candidate range derived from the channel under test before any behavioural falsifier is scored – so an unrouted channel cannot self-route a false negative.

Config (E3Config + REEConfig mirror + from_dims; all no-op default): use_modulatory_channel_routing (False), modulatory_channel_route_min_range_floor (1e-6), modulatory_channel_route_weight (1.0), modulatory_channel_route_source (“none”).

Honest scope. For coherence (661): currency_coherence is a scalar (uniform across candidates) – a scalar gate is rescale-invisible by construction. Routing the compose [K] bias gives it the cosine’s range, but joint-vs-alternation mode-discrimination (a different per-candidate pattern, not just magnitude) is a MECH-294-side concern, out of scope here. The P0 gate correctly flags a channel carrying no cross-candidate range as substrate_not_ready.

Validation: V3-EXQ-663 substrate-readiness diagnostic (claim_ids=[]; ARM_0 routing OFF vs ARM_1 routing ON, both authority ON + e2_world_forward + SD-056 online). PASS (route_range_substrate_ready) confirms the P0 gate and unblocks the separate per-claim behavioural retests of ARC-065 / MECH-294 / ARC-062 / MECH-309 / MECH-341. Those claims stay candidate / v3_pending / pending_retest_after_substrate; not weakened.