Certainty tags turn 0–9% accuracy into 88–100%
Notes on Notes2Skills: From Lab Notebooks to Certainty-Aware Scientific Agent Skills (arXiv:2606.11897) — S. Liu, Jiayao Chen, Chengwei Qin, Yanqing Hu, Jufan Zhang, Linyi Yang · June 2026
Note published · written by SkillFed’s research pipeline from the paper above · how these notes are made
AI-assisted notes · reviewed by SkillFed Skill evolutionNotes2Skills runs lab notebooks through two stages before any of it reaches an agent. Stage 1, epistemic directive extraction (EDE), scans free-text notes and tags each statement as FACT, JUDGMENT, or SUGGESTION, then sorts it into one of five directive types: flag data for review, note a condition change, suggest an analysis step, revise a protocol, or adjust a parameter. Stage 2, MetaSkill compilation, turns those labeled directives into agent-loadable Markdown skills, where each one keeps its certainty tag plus a SHA-256 hash pointing back to the exact source excerpt. They validated the pipeline against 461 hand-annotated segments across three corpora — informal free-form lab notes on one end, formal wet-lab protocols on the other.
The best extraction setup — Claude, few-shot — hit an F1 of 0.737 on detecting directives and a quadratic weighted kappa of 0.761 on grading certainty; compilation then preserved every one of 149 finalized directives, labels and source links included, with zero loss. The real test was downstream. Across three nanopore wet-lab sessions (48 files total) and seven pipeline configurations, only the full MetaSkill setup with executor verification avoided both failure modes at once. A baseline that just handed raw notes to an LLM fell into uncertainty laundering — treating tentative judgments as confident actions — and scored 0–9.1% on uncertainty-heavy sessions, versus 88.2–100% for the certainty-aware pipeline. Strip the certainty schema back out and hand the executor authority anyway, and accuracy on a fact-dominated session collapsed from 88.9% to 44.4% — the paper's clearest case of directive loss, where firm author instructions get buried and ignored once everything reads as equally tentative.
Key numbers
| External-LLM baseline accuracy, uncertainty-heavy sessions | 0–9.1% |
| Certainty-aware pipeline accuracy, same sessions | 88.2–100% |
| Directives preserved through MetaSkill compilation | 149/149 (100%) |
| Accuracy collapse w/o certainty schema (fact-dominated session) | 88.9% → 44.4% |
| Stage 1 directive-detection F1 / certainty QWK | 0.737 / 0.761 |
Skills related to this research
Related research
Related notes
- 71% of Mined Scientific Skills Don't Exist in Any Skill Marketplace →
- Four-Way Attribution Gates Every Update to a Million-Scale Skill Corpus →
- 215 Skills, 165 Contributors, No Fidelity Test →
- SKIM cuts agent skills to 30-60% of their length for a 1-2 point accuracy hit →
- Auto-evolved skill specs beat hand-written ones by 7 points, at twice the constraint density →
- Best Skill Scanner Catches 17% of Cross-Modal Attacks — a Distilled 7B Model Hits 88% →
References
- Notes2Skills: From Lab Notebooks to Certainty-Aware Scientific Agent Skills (Liu et al., 2026, arXiv:2606.11897)
- Wet Lab Protocols corpus / X-WLP (action extraction from published protocols)
- BioPlanner (LLM-based protocol planning)
- SYNTACT (dialogue-based ambiguity resolution in SOPs)
- Anthropic SKILL.md standard