Model × harness × alignment

Shaun Smith, Hugging Face · secureai.se, Stockholm · 12 November 2026

The slides

• Model × harness × alignment: the slidesThe deck as presented, in the browser. Arrow keys to move.evalstate-presentations.static.hf.space/2026/benchmarks

Find me

• Hugging Facehuggingface.co/evalstate
• GitHubgithub.com/evalstate
• Xx.com/evalstate

Annotated traces

Public benchmark runs, trial by trial, scanned with atif-scan. Findings are review candidates, not verdicts.

• Terminal-Bench 2.1 leaderboardThe main leaderboard's rows.evalstate-tb21-annotated.static.hf.space
• Terminal-Bench 2.1 submissionsPull-request submissions.evalstate-tb21-submissions-annotated.static.hf.space
• Terminal-Bench 2.1 fast-agent runsOur own published fast-agent runs.evalstate-tb21-fast-agent-annotated.static.hf.space
• DeepSWE 1.1Datacurve's public DeepSWE runs.evalstate-deepswe-annotated.static.hf.space

Tools

• atif-scanReads agent traces offline and flags steps for review: lookups, lures, side channels, integrity gaps.github.com/huggingface/atif-scan
• fast-agentThe harness behind most of the traces in the talk. Writes ATIF traces natively.fast-agent.ai
• safeagentsA check before every tool call, between the harness and the model.github.com/huggingface/safeagents
• Harness adapters: live training leaderboardOne LoRA adapter per model and harness, with reward, pass rate and tool-call checks.harness-hub-harness-adapters-rl.hf.space