Discuss adaptive model capability and reasoning-effort routing

[Kevinjohn]

The review workflow does not currently steer different phases toward higher or lower model capability or reasoning effort, leaving hosts unable to balance review depth, cost, and throughput consistently.

Discussion

Some review stages involve substantially different reasoning demands. Architecture mapping, critical invariant analysis, cross-component reconciliation, candidate validation, and the independent final audit may benefit from the highest available capability. Mechanical inventory, normalization, report generation, and bounded low-risk inspection may not need the same allocation.

It may be useful to add host-neutral adaptive routing guidance that can:

• increase capability when risk, ambiguity, or cross-boundary reasoning demands it;
• reduce capability for bounded or mechanical work;
• select between models, reasoning-effort levels, or a combination of both;
• escalate dynamically when a task uncovers unexpected complexity or a potentially material finding;
• fall back safely when a host cannot select models or effort levels.

Possible routing dimensions

• Review phase: architecture, semantic review, validation, reconciliation, tail review, candidate validation, or final audit.
• Work-unit risk: Tier A, B, or C, without treating risk tier as a complete proxy for reasoning difficulty.
• Review angle: security, concurrency, recovery, distributed behavior, compatibility, documentation, and so on.
• Task shape: integrative reasoning, bounded semantic inspection, or mechanical/schema-constrained processing.
• Evidence discovered during execution: ambiguity, cross-component invariants, potential P0/P1 findings, or conflicting evidence.

Illustrative policy

• Highest available capability or effort: architecture and unit design, Tier A review, independent second review, cross-component reconciliation, candidate validation, and final audit.
• Standard capability or effort: bounded Tier B/C semantic review and focused validation.
• Lower capability or effort: path inventory, metadata extraction, normalization, report generation, and mechanical consistency checks.
• Escalation: move work upward when ambiguity, security, concurrency, recovery, trust-boundary, or material-severity concerns appear.

Design questions

1. Should routing be advisory guidance in the Skill, a canonical field in assignment manifests, or both?
2. Should the vocabulary describe abstract capability classes such as highest_available, standard, and mechanical rather than specific products or model names?
3. How should model selection and reasoning effort interact when a host supports both?
4. Which phases must never be downgraded, and which can be downgraded only with later higher-capability verification?
5. Should Tier A automatically establish a minimum capability, or should routing depend on both risk and task shape?
6. What evidence should trigger escalation or permit de-escalation during an active attempt?
7. How should assignments remain portable across hosts that expose different controls—or none at all?
8. What state should be recorded so the final audit can evaluate whether routing decisions were appropriate?

The goal would be to improve review quality where deeper reasoning matters while avoiding unnecessary expense and latency for work that can be completed reliably with lower capability.