core/docs/adr/ADR-0254-grounded-open-hedge-arm.md

ADR-0254: Grounded-Open Hedge Arm for the Shadow Coherence Gate

Status: Proposed (predicate pre-authorized by Joshua Shay, 2026-07-22 weekly-audit ruling T13 decision 2; ratify on PR #103 merge) Date: 2026-07-23 Deciders: Joshua Shay (ruling authority) + Claude (implementation) Companion docs: ADR-0036-safety-refusal-policy.md, ADR-0037-per-predicate-ethics-refusal.md, ADR-0038-hedge-injection.md, ADR-0252-problem-solving-paradigm-consolidation.md Preserves (governing, unchanged): wrong=0, refuse-don't-guess, decode-don't-generate, fail-closed (INV-34), no lexical cue-tables (ADR-0252).

---

1. Context

The dual-competing Shadow Coherence Gate (core/cognition/surface_resolution.py::resolve_surface, introduced PR #96) abstains — emits a typed CoherenceRefusal — whenever the conjugate competitor (geometric residual contract) is open. The conjugate is open when the field's versor_condition ≥ 1e-6 (→ missing_bindings) or the GoldTether residual R_GT > 1e-6 (→ unresolved_hazards), as built by _geometry_contract_assessment.

The weekly audit (2026-07-22) surfaced an over-refusal in the warmed_session_consistency lane. Once the wave field is perturbed by a prior turn, a subsequent pack-grounded definitional surface — e.g. "What is doubt?" → "To doubt means to think maybe not. pack-grounded (en_core_meta_v1)." — hits an open goldtether_residual and is hard-refused with "I cannot certify an answer: the geometric coherence contract is open (goldtether_residual)."

This is a false refusal. The pack surface never claimed geometric certification. It stands on its curated textual provenance; the open geometric residual says only that the field did not close, not that the pack answer is wrong.

A tempting fix — a "definitional/epistemic" query-type classifier that bypasses the gate on a lexical cue — is rejected: it fails a geometric coherence gate open on a surface cue (fail-open cue-table; violates ADR-0252 and INV-34).

2. Decision

Route open-geometry-but-pack-grounded surfaces to a hedge arm — serve the pack surface non-authoritatively (authoritative=False) — instead of hard-refusing, discriminated purely by grounding provenance (structural), never by question type.

Authorized predicate:

open_conjugate ∧ pack_grounded ∧ ¬hazard → hedge_arm

Realized structurally inside the gate (single source of truth):

• pack_grounded — grounding_provenance ∈ {pack, teaching} (mirrors the existing chat.runtime "grounded" test). vault/oov/partial/none never hedge.
• ¬hazard — enforced by a residual allowlist, not question typing. Hedge iff every open token ∈ {versor_condition, goldtether_residual} (geometric-coherence residuals the pack surface never certified). Any unrecognized open token — where a genuine safety/harm/imperative hazard would land — fails the allowlist and takes the unchanged hard refusal. missing_wave_field is deliberately excluded (absence of any field is a harder failure than an open-but-computed residual).

The hedge outcome: the pack surface with a deterministic "Grounded but not geometrically certified —" marker, authoritative=False, hedged=True, authority="grounded_open_hedge". Walk/compose folds are suppressed — a hedge never accretes deterministic inference authority.

3. Why this is fail-closed, not fail-open

• The only ContractAssessment reaching resolve_surface is the shadow-gate geometry assessment (pipeline.py:416), whose tokens are ⊆ {versor_condition, missing_wave_field} (bindings) ∪ {goldtether_residual} (hazards). Genuine safety/harm/imperative hazards ride separate axes — SafetyVerdict → typed refusal (ADR-0036) and the logos-morph override (pipeline.py:513) — both of which supersede this surface downstream. The hedge cannot leak a harmful surface: safety and morph refusals still win.
• The predicate is structural (provenance + a closed allowlist of geometric tokens). It reads no question text. An unknown token defaults to refusal (INV-34).
• require_closed_geometry callers that omit grounding_provenance get the historical hard refusal unchanged (no accidental hedge).

4. Consequences

• The over-refusal is fixed: the warmed "What is doubt?" case now serves the hedged pack definition (verified end-to-end).
• New auditable authority tag grounded_open_hedge flows to TurnEvent.authority_source; audit distinguishes hedge from both certified answers and refusals.
• Not conflated with ADR-0038. The ADR-0038 hedge is ethics-commitment-driven (hedge_commitments + manifold preferred_hedge_soft); this hedge is coherence-geometry-driven. They are independent mechanisms. Unifying the grounded-open marker with the manifold's preferred_hedge_soft vocabulary is a deferred follow-up (defer substrate-vocab commitment) and intentionally out of scope here.
• authoritative=False lives in the gate; downstream observes it as authority_source="grounded_open_hedge" (consistent with how refusals already surface — no new result plumbing).

5. Validation

• Unit (RED→GREEN): tests/test_grounded_open_hedge_arm.py — hedge activates on pack + geometric-residual openness; fails closed on non-pack grounding, unknown/mixed hazard tokens, missing_wave_field, and None assessment; closed geometry stays authoritative; omitted provenance stays refusing.
• Real-data (GSM8K-style): the warmed mixed-intent sequence exercises the arm on real field dynamics; the warmed_session_consistency lane never hard-refuses a pack surface, with telemetry_consistency_rate and no_placeholder_rate held at 1.0.
• Regression: architectural invariants (incl. INV-34) green; GSM8K wrong=0 preserved (the closed-geometry certified math path is untouched).