The Clipped-ReLU DEL Authorization Function

Status: PROPOSED CANON v1.0 (2026-08-11) — governance surface under CONSTITUTION.md Article VI; this document becomes canon when the owner merges its review PR, and its per-action-class parameterization remains [Roadmap] until built. Owner decision 11 (clipped-ReLU DEL as canon) is decided by that merge. Sources: docs/product/SIGNAL_SHOPIFY_MASTER_v4.md §2.6, §3.2; docs/product/SIGNAL_OVERVIEW_v5.md §2.6; CONSTITUTION.md II.1/III.6; OPERATING_SYSTEM.md Art. 2.4/4; the shipped demo gate # MIZ OKI 3.5/mizoki_runtime/demo_signal.py (documented read-only — demo behavior is NOT modified by this document; demo surfaces ship only via the owner-approved site dispatch).

1. The function

For every action class c:

authority_c = min(cap_c, max(0, DEL_score − threshold_c))

• Flat zero below threshold — deterministic denial. No partial execution, no probabilistic leakage. Agents (however they negotiate) only propose; the deterministic policy service authorizes. A denial routes the proposal to a human with its reasoning path and a smaller alternative — never silently dropped.
• Margin-proportional authority above threshold. A barely-cleared score earns only the smallest reversible version of the action (the ACT-991 canon, §5).
• Saturation at cap_c. The covenant cap bounds authority regardless of margin.
• Adaptive guardrail envelopes are threshold shifts under volatility — the envelope may move threshold_c up (more conservative) in turbulent regimes; it never lowers a floor.

2. Threshold ownership — floors are not configurable downward

• The platform sets per-class floor thresholds. Merchants/customers may raise thresholds via covenant; they may never lower them below platform floors (master v4.0 finding N3 — High/safety). Safety is not merchant-configurable downward.
• The platform-wide floor that exists today: DEL ≥ 80.0 at the Decision Control Plane (CONSTITUTION II.1; OPERATING_SYSTEM 2.4). Domains may raise their bar via policy.yaml (per-domain 90s exist); none may set it below 80.0 — check_conformance() enforces the alignment. The clipped-ReLU parameterization inherits this: no threshold_c may sit below the DCP floor for its domain.
• Media-domain bounds that exist today and bind every media action class: the $25,000 autonomous ceiling (above it → approval-required to a named human), MDES bands with a hard operator gate on expansion, and experiment-required for media claims without incrementality evidence.

3. Mechanical demotion — never discretionary

• Per-class demotion on sustained near-zero margins: an action class whose authority hovers at zero is demoted automatically (master §2.6).
• Covenant proximity auto-drops an L5 class to L3 (master §3.2). Every demotion path is arithmetic against declared constraints; no human judgment is required to demote, only to re-promote.
• Autonomy levels are held per account × action class (L4 on budget while L2 on bids is normal); every level carries a one-tap kill switch, immutable journal, weekly plain-language digest.

4. The shipped gate, documented from source — [Validated] at miniature scale

# MIZ OKI 3.5/mizoki_runtime/demo_signal.py implements this function's shape at demo scale, deterministically (seed-replayable; same seed → byte-identical run). Documented values, read from the module on 2026-08-11:

ReLU gate (ReLUGate) — signal admission before planning:

score = max(0, uplift) × confidence × ln(1 + sample_size)

with three floors, each emitting a visible failure reason:

Guardrail battery (GuardrailSet.evaluate) — validation of each planned action, checks in source order:

1. budget_swing_cap — budget actions capped at ±20% (BUDGET_SWING_CAP_PCT)
2. bid_swing_cap — bid actions capped at ±30% (BID_SWING_CAP_PCT)
3. confidence_floor — decision confidence ≥ 0.70 (CONFIDENCE_FLOOR)
4. sample_floor — supporting conversions ≥ 15 (SAMPLE_FLOOR)
5. rollback_ready — rollback token minted before execution

Any failed check blocks the action before execution (blocked_by names the rule); each scenario seeds one deliberate block so the veto is always visible; surviving actions rank by expected_value × confidence; executions run dry_run with rollback tokens; a pure-veto run holds ("No move earned execution this run — the desk held"). The margin-proportional clip and per-class cap_c/threshold_c registry are not in the demo — the demo is the gate + guardrails miniature, [Validated]; per-class parameterization is [Roadmap].

5. ACT-991 — the canonical example

The governed-decision class this function generalizes, proven end to end on live Cloud Run (run 2ebf83c1, 24/24 steps — verified result 2026-07-27; the proposal itself is an illustrative scenario): DEL score 84 against threshold 80 → margin 4 — a barely-cleared score; the $45k proposal exceeded the $25k media autonomous ceiling → routed approval-required to a named human; the authorization issued was signed, expiring, single-use, bound to the exact proposed action (tampering → 422, second redeem → 409); Stage-3 recommend-only recorded intent without executing. Under the clipped-ReLU reading: small margin ⇒ smallest reversible authority; ceilings and gates bind regardless of score.

6. Intent-model promotion gates (F8.1) — canonical thresholds

The gates that decide whether an intent model may leave observe-only are canonical and identical everywhere (CONSTITUTION III.6; overview §2.6; capabilities doc Stage 2):

• Brier ≤ 0.20 · AUC ≥ 0.72 · stable lift across ≥ 2 purchase cycles, and promotion happens only via human approval.
• A model below the bar is advisory-only (observe-only) regardless of the merchant's autonomy level — merchant L-level never overrides a model gate.
• Brier degradation > 0.20 after promotion auto-demotes the model to observe-only, demotes the dependent action classes, and opens an incident.
• Live demonstration that the gate holds: the LII offline backtest recorded AUC 0.6884 (honest, leakage-gated) against the 0.72 line — a documented NO-GO; the production scoring path stays observe-only (open-work register item 17).

7. Build state — what exists vs. what this canonizes

No claim in this document upgrades any capability's label; the platform ceiling remains built, pre-benchmark.