russell@unturf.com
83ab29734a
ticket(#000008): architectural review II + disable hierarchy
Two review passes added to ticket #000008:
§9 — first architectural review (2026-05-03). De novo synthesis
applying CTI / Merkle-AGI-DAG / Reverse-RAG / PROMETHEUS-Σ stack.
Replaces the §4 "land Option D" recommendation with a phased plan
(Phase 0 telemetry → Phase 1 classifier → Phase 2 cap → Phase 3
reminder held → Phase 4 reject held). §4 marked superseded but
preserved as design log. §8 Status updated.
§10 — second-pass refinements. Where §10 conflicts with §9, §10
wins. Substantive deltas:
- §10.1 split ALL rung into bounded / unbounded / emergent
(all-Beatles ≠ winners-of-all-major-sports)
- §10.2 entropy-pressure framing; cap is one knob among three
- §10.3 three new violation kinds, no new audit_mode token
(BROAD_QUANTIFIER_RUNAWAY/CAP_APPLIED/SCOPE_UNBOUND)
- §10.4 three-clause CTI contract (quantifier / answer /
evidence) instead of one flat clause
- §10.5 concrete output templates per policy mode
- §10.6 bench-field corrections; model_profile_id includes -fp8
- §10.7 explicit do-not list
- §10.8 quantitative decision tree (≥5pp deltas per mode)
- §10.9 five-step method walkthrough binding to CLAUDE.md
§10.11 — disable hierarchy + dry-run discipline. Six-level
disable hierarchy (per-test → per-call CLI → per-phase policy
→ per-mode → per-model-profile → master-via-governance-hash).
Dry-run discipline: Phase 1 classifier lands with
quantifier_guard_apply_caps=false; bench review of intensity
distribution catches mis-classification before any verdict is
affected. Per-phase regression fixtures named. Rollback playbook
prefers policy-flip over git-revert.
Honest expectation-setting: §5.1.1 baseline shows
NO_EVIDENCE_POINTER 3/3 dominates; cap-only (Phase 2) likely
moves needle barely; Phase 3 reminder might or might not move
the gate; Phase 4 reject mechanically works but carries false-
positive risk on bounded universals (handled by §10.1 split).
Telemetry-first stays the load-bearing recommendation.