APEX-Harm
Harbor: pendingA benchmark evaluating prompt and script injection safety for long-horizon agents in shared professional workspaces across finance, consulting, and legal tasks.
- Samples
- 180
- Released
- 2026
- Scores
- 10
MechHypoBench
Evaluates AI agents' ability to generate open-form mechanistic hypotheses from 17.98 million empirical records across 14 scientific fields by testing implied consequences.
How to run
ltb · mechhypobench
$ npm i -g lettokenburn-cli$ ltb bench info mechhypobench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
5 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.6 Sol | 64.9 | codex | Independent | medium | Oct 4, 2026 | arxiv.org |
| AI-Scientist-v2 | 59.2 | ai scientists | Independent | medium | Oct 4, 2026 | arxiv.org |
| DeepSeek V4 Flash | 49.2 | codex | Independent | medium | Oct 4, 2026 | arxiv.org |
| Claude Sonnet 5 | 43.7 | claude code | Independent | medium | Oct 4, 2026 | arxiv.org |
| GLM-5.2 | 42.5 | codex | Independent | medium | Oct 4, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
Combines paper-derived mechanisms from 14 scientific fields with real-world datasets containing 17.98 million records. It preserves observational complexity while providing specified underlying mechanisms for hypothesis generation.
Assesses open-form hypotheses by evaluating their consequences under withheld conditions using ground truth outcomes. Experiments reveal a substantial gap between generated hypotheses and underlying mechanisms.