counselbench-100 / SCORING.md
SamuelChien821's picture
Add files using upload-large-folder tool
7fb2ebd verified
|
Raw
History Blame Contribute Delete
1.94 kB

CounselBench-100 scoring contract

CounselBench-100 v1.1 reports continuous, deterministic criterion credit and a separate strict pass. The grader makes no model, network, clock, locale, or random call.

Why v1.1 replaced the launch metric

The v1.0 baseline contained ten substantive GPT-5.6-sol executions. Every run read all 96 documents and completed 109–117 successful MCP calls, yet all ten received zero because the verifier required full JSON equality and a single all-or-nothing memo grounding gate. Some expected issue labels and remediation owners were assigned by generator order but were not recoverable from the records. That made the binary display invalid as a diagnostic model metric.

v1.1 exposes stable finding IDs, issue labels, severity, owner, and response deadline in realistic record-control metadata. The verifier scores those recoverable values individually while retaining a strict all-criteria pass.

Criteria per task

Category Criteria Weight
Review procedure 8 25%
Structured findings 152 55%
Counsel memo 22 20%
Total 182 100%

The findings category has eight document-level criteria plus nine criteria for each of 16 findings: presence, five exact identity/evidence fields, required fact anchors, source-bounded controlled facts, and required action anchors. The memo category checks five sections, one complete grounding criterion per finding, and absence of forbidden claims.

Caps and strict pass

  • Missing deliverables or outputs not written through MCP cap reward at 0.20.
  • Any other incomplete review-procedure criterion caps reward at 0.49.
  • Strict pass requires every procedure, finding, and memo criterion.
  • Exact equality with the reference answer remains a diagnostic only.

All criterion outcomes and category scores are included in the verifier report, so a leaderboard value can be reproduced without an external judge.