Brain: confidence-gated escalation ladder (schema-repair + tier bump) to replace zero-retry parse→fallback (#2419 family)

[rysweet]

Parent / context: #2431 (research findings from BGML / TPAMI 2024). Directly addresses the #2419 family.

Idea (from BGML's progress-aware module, §IV-C)

BGML spends extra compute only where coarse confidence is low: a fast coarse judgment is escalated to finer processing only for the weakest cases, instead of failing or defaulting. Map this onto Simard's brain decision path.

Problem in Simard (grounded)

All three brains parse LLM output via parse_decision_from_response() — find first { / last }, then serde_json — with zero retry; any error maps immediately to a deterministic fallback:

• src/ooda_brain/rustyclawd.rs:85 (decide_engineer_lifecycle / act)
• src/ooda_brain/decide.rs:179 (judge_decision)
• src/ooda_brain/orient.rs:215 (judge_orientation)
• fallbacks: src/ooda_brain/fallback.rs:14

Only OrientJudgment has a native confidence: f64 + validate() (src/ooda_brain/orient.rs:70,85). DecideJudgment (src/ooda_brain/decide.rs:57) and EngineerLifecycleDecision (src/ooda_brain/mod.rs:86) expose only rationale — no confidence, no validation. This is the mechanism behind #2419 (decide_engineer_lifecycle defaulting to ContinueSkipping ~99.6%).

Proposal (bounded)

Replace the binary parse-or-fallback with a confidence-gated escalation ladder:

1. Cheap parse (as today).
2. On parse-failure or low self-reported confidence: escalate, bounded to 1–2 attempts:
   • schema-repair re-prompt — feed the malformed payload back asking for valid JSON against the known schema; and/or
   • a higher-effort model/prompt tier.
3. Only after escalation is exhausted, fall to the deterministic default.

Supporting changes:

• Add a native confidence field + a validate() to DecideJudgment and EngineerLifecycleDecision (parity with OrientJudgment).
• Instrument per the Brain: decide_engineer_lifecycle defaults to continue_skipping ~99.6% of invocations (instrument + fix) #2419 plan: emit brain_lifecycle_decision via self_metrics::record_metric with outcome ∈ {parsed, default_empty, default_malformed, repaired, escalated, error_*}. BrainJudgmentRecord already carries confidence+fallback (src/ooda_brain/judgment_record.rs:41).

Acceptance

• Parse-failure/default rate for decide_engineer_lifecycle (and the Decide/Orient phases) drops substantially from the ~99.6% baseline in Brain: decide_engineer_lifecycle defaults to continue_skipping ~99.6% of invocations (instrument + fix) #2419.
• Escalation is bounded (no unbounded retry / latency blow-up).
• brain_decision_parse_failure_rate metric exists and is gated in CI/self-metrics.

Effort / Impact

Effort M (few PRs). Impact H — restores the core reasoning function the brain is supposed to provide. Related: #2419, #2421, #2429, #2430, #2046, #1748.

[rysweet]

⚠️ Grounding correction before implementing (validation pass)

This issue's "Problem in Simard" section is partly stale — verified against worktrees/main @ 795c4ea5:

• parse_decision_from_response() in rustyclawd.rs no longer parses JSON. It now requires a DECISION: marker and rejects everything else — "JSON fallback removed — it was the source of fallback storms" (test-coverage: cover operator_commands_ooda daemon lifecycle + de-flake operator_commands_meeting goal_curation slow tests (child of #1977) #1980). So the engineer-lifecycle brain is not "first { / last } → serde_json". Only orient.rs (parse_judgment_from_response) still uses that pattern. The escalation ladder should build on the test-coverage: cover operator_commands_ooda daemon lifecycle + de-flake operator_commands_meeting goal_curation slow tests (child of #1977) #1980 marker parser, not reintroduce a JSON parser.
• DecideJudgment (decide.rs:53) and EngineerLifecycleDecision (mod.rs:116) are enums, not structs — "add a confidence field" needs reframing (per-variant, a wrapper struct, or reuse BrainJudgmentRecord.confidence, which already exists and records 0.5 on fallback).
• Line numbers drifted; rely on file paths + symbol names, not the cited lines.

The escalation-ladder direction (schema-repair re-prompt / tier bump before deterministic fallback, plus measurable parse-failure rate) remains valid and high-value. See validation summary on parent #2431.