agentleFS
Sign inStart free
IgorGanapolskyIgorGanapolsky/ThumbGate/.agents/skills/decisive-step-evaluator/SKILL.md

decisive-step-evaluator

Skill29 starsChanged 7 months agoMITView on GitHub

NVIDIA arXiv:2610.10478 decisive-step evaluation: isolate the 1st code edit that flips a repository's tests from failing to passing. Score models only on this decisive action (Spearman rho = 0.964 vs SWE-bench Verified) rather than diluting scores across all exploration steps (rho = 0.915).

Create a free account to mount
Files
SKILL.md

Decisive-Step Evaluation (NVIDIA arXiv:2610.10478 FORMAT Steal)

Goal

Apply the empirical finding from Before They Can Solve: Predicting Post-Training Coding-Agent Performance from Base Models (NVIDIA Research, arXiv:2610.10478):

  1. Isolate the Decisive Step in multi-turn trajectories: the first code edit that flips tests from fail to pass.
  2. Score candidate models only on the decisive action (via Decisive BPB and Patch MCQ contrastive discrimination) rather than averaging across every noisy exploration step.
  3. Select models for agent workflows and post-training based on decisive-step signal ($\rho = 0.964$).

Constraints

NEVERALWAYS
Score models by averaging across all turns (exploration, ls, grep)Isolate the decisive step where test status flipped from fail to pass
Evaluate base models from cold-start multi-step runs (mis-ranks on tool syntax)Evaluate prefix-conditioned at the decisive step
Accept trajectories without a verified test flipReject trajectories that never failed or never passed (hasDecisiveStep: false)
Select candidate models based on parameter count aloneSelect based on decisive-action discrimination rate ($\ge 60%$)

Commands

# Doctor status
npx thumbgate decisive-step --json

# Rail map
npx thumbgate decisive-step --map-only --json

# Test suite
node --test tests/decisive-step.test.js

Reference

  • Paper: arXiv:2610.10478 (NVIDIA Research, Oct 2026)
  • Synthesizer: Rohan Paul (@rohanpaul_ai, Oct 12, 2026)
  • Evaluator: scripts/decisive-step-evaluator.js
  • Doctor: scripts/decisive-step-doctor.js
  • Documentation: docs/agents/decisive-step-evaluator.md

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free