Is Surgeon state of the art?Pending benchmarks.Announced Nov 3, 2026; no independent results yet.
Crosby's in-house verifier model that judges whether a contract redline is surgical (precise, targeted edits that keep the legal substance), returning reasoning plus a yes or no verdict. Post-trained from Qwen-3.5-27B; internal, live only inside Crosby's service for its customers.
The details so far
- Lab
- Crosby
- Announced
- Nov 3, 2026
- API
- Not yet
- Open weights
- Closed
- Parameters
- 27B (Qwen-3.5-27B base)
- Type
- Verifier: reads one redline and outputs a reasoning trace and a yes or no verdict on surgicality
- Base model
- Qwen-3.5-27B (Crosby's blog)
- Training
- One epoch of SFT on rejection-sampled teacher traces, then GRPO with a reward gated on a correct verdict times an LLM-judge score of the trace
- Training data
- Crosby's agent redlines labeled against attorney-sent versions, plus synthetic negatives; validated on 500 redlines relabeled by Crosby attorneys
- Use
- Offline evaluation, live monitoring, auto-acceptance of surgical redlines and steering Crosby's agentic harness
- Availability
- Internal/customers only: in production inside Crosby's service; no public weights or API
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 90.6% balanced accuracy on Crosby's internal 500-redline surgicality set, against 84.3% for Gemini 4 Argon, 83.1% for GPT-6 Astra and GPT-6 Sol, and 60.4% for a heuristic baseline.
Lab claim: roughly 10-100x lower cost per million tokens than most closed models in that comparison.
Sources
LABHow to Train your Surgeon (Crosby blog)crosby.ai/blog/how-to-train-your-surgeon · Nov 3, 2026LABCrosby announcement on Xx.com/crosbylegal/status/2117664204568109344 · Nov 3, 2026Tracking ends Feb 1, 2027 unless the API or weights become available.