Research, benchmarks, and engineering notes on building systems that know when to choose.
A 9B decision model that scores every option in a single forward pass. Runs locally in 22 ms on one GPU or 88 ms on an Apple M5 Max, and answers 51% of requests in front of Claude Opus at Opus-level combined accuracy.
An agent often needs a bounded judgment before its next action: which workflow fits a request, whether a passage supports a claim, or how an item scores under a rubric.
Route tickets, tools, and workflows across your options.
Screen evidence support and relevance against a rule.
Compute expected score and level probabilities.
“The parcel arrived damaged. Please replace it.”
Choose the team to handle this request.
p(returns) ≥ 0.90 → route directly to returns queue; otherwise escalate for review.
Chronological archive of Seb Labs releases, benchmark methodology, and agent integration notes.
A practical introduction to bounded judgments, structured probabilities, and connecting Seb to your agent.
Choice, yes/no, and ordinal score primitives over HTTP and remote MCP.Benchmarks, local decisions, model routing, and simulated robotics.
37 public suites, confidence-routed Claude cascades, and simulated robot policies.