SOTA: progress-guided stepwise model routing for long-horizon workflows

## Finding

ProgRouter, arXiv:2608.25998, submitted 2026-08-26 and accepted to EMNLP 2026 Findings, replaces one-shot query routing with online stepwise routing based on evolving workflow progress, remaining difficulty, time budget, and long-term operating cost. The originating team reports lower operating cost than key baselines while preserving strong performance across HumanEval+, MBPP, MATH-500, and ASQA.

Evidence class: originating-team conference paper; exact RuV benefit is unreplicated.

Weighted Opportunity Score: 4.39/5.

## Existing RuV capability

Ruflo already has model routing, Q-learning routing, cost tracking, task orchestration, and learned priors. The missing primitive is a provider-neutral `ProgressReceipt` that measures marginal progress after each workflow step and makes that signal available to the router.

## Proposed reusable primitive

```ts
interface ProgressReceipt {
workflowId: string;
stepId: string;
completedSubtasks: number;
remainingSubtasks?: number;
stateQuality: number;
progressDelta: number;
confidence: number;
elapsedMs: number;
tokens: number;
costUsd?: number;
evidenceDigest: string;
}
```

At each routing point, estimate expected progress gain for each eligible model or execution tier. Route according to expected marginal value subject to hard budget and authority constraints.

A first implementation should use existing deterministic workflow signals before adding another learned model: tests passed, files changed as intended, subtasks closed, verifier score, unresolved blockers, and repeated no-progress steps.

## Policy

Never escalate solely because a workflow is long. Escalate when the expected marginal progress per unit cost is superior and the action remains inside the existing capability envelope. Provider/model choice cannot modify the evaluator, budget, or authority boundary.

## Benchmark

Compare fixed strong model, current Ruflo router, and progress-guided router on matched long-horizon code, math, retrieval, and tool workflows. Report success, cost, tokens, p50/p95 latency, model mix, progress prediction calibration, stall recovery, and per-step routing decisions.

## Falsification

Reject if simple current routing is within variance, progress scoring costs more than it saves, the router oscillates between tiers, or progress proxies can be gamed without improving held-out task success.

## Acceptance

At equal or better task success, require at least 15% lower model cost or 15% fewer frontier-model calls on held-out workflows, no protected-slice regression above 2 points, complete routing receipts, and no authority or evaluator mutation.

Start by locating Ruflo's existing model routing, Q-learning routing, cost tracking, task orchestration, and learned-prior entry points. Compare those paths with the proposed ProgressReceipt and benchmark requirements; done means held-out workflows meet the stated cost or call reduction, regression, receipt, and authority constraints.

Written by the indexing model from the issue text.