research(routing): RAPS Bayesian reputation scoring for AgentRouter robustness

[bug-ops]

Summary

Bayesian per-agent reputation scoring for robust multi-agent coordination, derived from the RAPS architecture. Directly extends Zeph's Thompson Sampling AgentRouter with a principled malice/degradation detection layer.

Source: arXiv 2602.08009 — "Towards Adaptive, Scalable, and Robust Coordination of LLM Agents: A Dynamic Ad-Hoc Networking Perspective" (Li et al., 2026)

Technique (reputation component only)

Each agent/model maintains a Beta-distribution reputation score per peer based on observed behavior (task success, output quality, response time vs. expectation). Agents with degraded reputation are automatically down-weighted in routing without central coordination. Reputation updates are Bayesian: success → α += 1, failure → β += 1 on a Beta(α, β) prior.

This is structurally identical to Thompson Sampling's Beta distribution already used in zeph-llm's router — it is essentially adding a quality/reliability dimension alongside the latency EMA.

Applicability to Zeph

MEDIUM-HIGH. Zeph already uses Thompson Sampling (Beta distribution) in AgentRouter for model selection exploration. The Bayesian reputation layer is an extension of the same math applied to output quality:

• Extend ModelStats in the router with a reputation: Beta field tracking quality outcomes (not just latency)
• Feed tool execution errors, LLM parse failures, and plan failures back as reputation signals
• Route away from degraded models/agents proportionally to reputation score decay
• Integrates with #1841 (Agent Stability Index) — ASI provides the coherence signal; reputation tracks the cumulative outcome history

Implementation sketch

• Extend ModelStats with quality_alpha: f64, quality_beta: f64 (Beta parameters)
• On tool execution / plan step: record success/failure → update quality params
• Routing score: combine EMA latency weight with reputation sample Beta(α,β).sample()
• New config key: [routing.reputation] enabled = false, decay_factor = 0.95