Context

Deep dive on "SRLM: Self-Reflective Program Search" (arXiv:2603.15653). Core finding: trajectory selection quality matters more than recursion depth (+22% over RLMs, no retraining).

Three Uncertainty Signals

1. Self-consistency filter: Sample K candidates, majority-vote filter. Discard candidates disagreeing with majority output.
2. Verbalized confidence (VC): Per-step structured confidence value (0-100). Log-space aggregation: VC(p) = sum_t log(nu_t / 100). Single low-confidence step penalizes entire candidate.
3. Trace length (Len): Total output tokens across steps. Shorter = more confident. Already tracked via TurnRecord.output_tokens -- zero new data collection.

Joint scoring: s(p) = VC(p) * Len(p) -- least-negative wins (most confident + most concise).

Implementation Blueprint

1. TrajectoryConfig dataclass: enabled: bool = False, k_candidates: int = 2, complexity_gate: tuple[str, ...] = ("complex", "epic")
2. New engine/trajectory_scorer.py: CandidateStep model, _filter_consistent(), _score(), select_best()
3. Verbalized confidence elicitation: structured {"confidence": <0-100>} JSON block in step generation prompt
4. hybrid_loop.py integration: asyncio.TaskGroup for K parallel candidates when enabled + budget allows
5. Budget guard (mandatory): If K-candidate sampling would exceed remaining task budget_limit, fall back to single-candidate

Design Decisions

• K=2-3 initially (not K=8 from paper -- 8x cost multiplier)
• Complex/epic tasks only -- never activate for ReAct (Loop 1) or simple/medium
• Off by default (enabled: False)
• Complements stagnation detection: SRLM is forward-looking (pre-commit selection), stagnation is backward-looking (past repetition correction)

Risks

• Cost multiplier: K=8 = 8x provider calls. Budget gate is not optional.
• Verbalized confidence calibration: Many models are poorly calibrated (DCPO research). Scoring degrades gracefully to Len-only.
• Self-consistency comparator: Step-type-aware comparison needed (tool-call intent, not raw string equality).
• Confident mistakes: If all K agree but all wrong, no protection. Stagnation detection + review gates remain backstop.

References

• arXiv:2603.15653
• Related: research: evaluate agent-controlled compaction for engine hybrid loop #687 (agent-controlled compaction -- both touch HybridLoop decision points)
• Related: feat: step-level quality signals + accuracy-effort observability metric #697 (step-level quality signals -- complementary observability)