research: explore additional Looper-family routing algorithms

Summary

Explore additional Looper-family routing algorithms after the existing Flow, Fusion, ReMoM, and Looper paths are hardened and measurable.

Parent roadmap: #2287 Depends on: #2333, #2336 Related: #1815, #2193, #2331

Motivation

Multi-model deliberation can improve quality, reliability, and specialization, but every new Looper algorithm increases cost, latency, and safety complexity. H2 should keep algorithm exploration active while requiring a stronger evaluation and protection contract.

Scope

Investigate algorithm families such as:

• debate or critique-and-revise loops;
• verifier/reranker loops;
• planner-executor model selection loops;
• budget-aware multi-model deliberation;
• confidence-gated escalation loops.

Each candidate algorithm should define:

• when it runs;
• which models it can call;
• cost and token caps;
• failure and timeout behavior;
• replay diagnostics;
• expected evaluation metrics.

Non-goals

• Do not ship new production algorithms without eval coverage.
• Do not turn Semantic Router into a general workflow engine.
• Do not bypass guardrails, protection policies, or recipe boundaries.

Acceptance criteria

• Candidate algorithms are documented with scope, cost model, and evaluation requirements.
• At least one candidate has a prototype or design that can be compared against existing Looper algorithms.
• The decision-level eval harness can measure quality and overhead for candidate algorithms.
• Productionization requires follow-up issues rather than landing exploratory code silently.