ticket #000037: David review applied — bicameral substrate amendments
Apply David's 2026-05-10 review of Prometheus-Σ doc-only Phase 0:
- Rename #000028 to multi-witness canonical agreement (witness vs
carrier distinction)
- Correct bicameral speed claim — kernel cheap inside its domain,
LLM is the scarce resource (whole Kelly argument depends on it)
- Add DEFERRED output label distinct from MARGINAL
- Move sweep-state schema to Phase 1 design choice; recommend
document_sweep_state sidecar over documents.last_swept_at column
- Broaden Target A sweep beyond previously-STRICT rows to include
CANONICAL_PROJECTION, EVIDENCE-WARRANTED, MARGINAL, ANCHOR-WARRANTED
- Numerically stable softmax + normalized entropy H/log(n)
- Kelly safety guards (b_i ≤ 0, vetoed, B = 0, sum-zero allocations)
- EMA-smoothed difficulty update law
- Extended utility with SelfModelCalibrationGain, WarrantPromotionGain,
MemoryInvalidationRisk, separate SecurityRisk
- 8-class hard-veto order; soft-hash veto integration with explicit
replay-window-unbounded → ESCALATE rule
- Gödel framing strengthened — UNKNOWN/ESCALATE are visible outputs
of the discipline, not the side-step itself
- Phase trigger exception handling (division-by-zero, small-sample)
- Controller proposes MemoryRoot/SelfModel updates, never mutates
- New §14 exception-handling matrix
- New §15 weight profiles (safe / conservative / exploratory)
- New §16 Phase 1 dataclass API + test list
- New §17 adjacent integrations (warrant-promotion, soft-hash,
spatial-temporal future)
- §19 open questions answered concretely
- §21 review log preserves provenance
Module paths updated: arborist/v9/prometheus.py →
arborist/substrate/prometheus.py per the v7+v8 → substrate
consolidation that landed earlier today.
Two doctrinal lines reaffirmed verbatim:
The controller does not defeat Gödel.
LLM is witness, never authority.
russell@unturf.com
7e81425d49