refactor(v1): reuse Pydantic adapters across validation

[xeophon]

Overview

Uses reusable Pydantic adapters where V1 handles schema-shaped data, while keeping shallow transport JSON parsing lightweight.

Details

• shares rubric criteria and verdict validation between rubric and agentic judges, including finite composition weights
• reuses cached adapters for typed trace output and rollout state channels
• validates legacy tool definitions through their full Pydantic schema
• replaces manual numeric config checks with native finite and positive Pydantic types

---

Note

Low Risk
Mostly validation consolidation and adapter reuse; scoring semantics stay the same with stricter rejection of NaN/inf weights and malformed verdicts at config/parse time.

Overview
Pulls rubric criteria loading and verdict scoring into load_criteria and score_verdicts in rubric.py, so the plugged RubricJudge and agentic-judge env share one path instead of duplicated Criterion models and inline parsing in finalize().

Config and rubric fields now reject bad numbers at parse time: FiniteFloat on judge/score weights, CriterionWeight (non‑negative, finite) on rubric weights, PositiveInt for RLM summarize_at_tokens, and max_criteria with ge=1. Agentic verdict JSON is validated through RubricVerdicts before scoring.

Performance/refactor: functools.cache wraps TypeAdapter for CLI episode read/write; each RolloutSession exposes a cached state_adapter for GET/PUT /state; legacy v0→v1 bridging validates tool defs via a shared TOOLS_ADAPTER with OnErrorOmit instead of a manual loop.

^{Reviewed by Cursor Bugbot for commit 04a2390. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Centralize rubric validation and reuse Pydantic TypeAdapter instances across v1

• Introduces load_criteria and score_verdicts helpers in rubric.py to centralize criteria loading and verdict validation, replacing duplicated inline logic in the judge env and RubricJudge.
• Adds CriterionWeight (non-negative, finite) and tightens field constraints on JudgeConfig, ScoreConfig, and RubricJudgeConfig to reject NaN/inf weights and non-positive max_criteria at parse time.
• Caches TypeAdapter construction via cache(TypeAdapter) in output.py and adds a state_adapter cached property to RolloutSession so adapters are built once and reused.
• Behavioral Change: AgenticJudgeEnv.finalize and RubricJudge.grade_batch now strictly validate verdict names against the rubric and reject off-menu answers, where previously this may have passed silently.

Changes since #2233 opened

• Renamed cached TypeAdapter factory and adapter instances from private naming convention to public naming convention [04a2390]

^{Macroscope summarized b987707.}

[macroscopeapp]

Approvability

Verdict: Approved 04a2390

This refactoring consolidates duplicated Pydantic validation logic into shared functions and replaces manual checks with built-in Pydantic types (FiniteFloat, PositiveInt). The changes are mechanical DRY improvements with no new features or behavioral changes. Open review comments are minor naming style suggestions.

^{You can customize Macroscope's approvability policy. Learn more.}

[mikasenghaas]

can we remove leading _ from al lthese

[mikasenghaas]

same here

[mikasenghaas]

do we really need this?

[mikasenghaas]

no _

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 83efd59. Configure here.}