RT-Safe
Harbor: pendingA simulated urban benchmark with 36 routes across 5 maps evaluating embodied agent safety under real-time constraints using VLMs.
- Samples
- 36
- Released
- 2026
- Scores
- 8
SciΠvot
SciΠvot: Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents
A 248-case benchmark evaluating scientific judgment by asking models to choose between original LLM plans and human-corrected alternatives using project records.
How to run
ltb · scipivot
$ npm i -g lettokenburn-cli$ ltb bench info scipivot
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 entriesEvery 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
SciΠvot comprises 248 cases from 26 PhD-level projects across six fields, pairing original agent plans with human-corrected alternatives based on intervention records. Models select the better plan using pre-intervention context.
Scoring measures accuracy against human corrections. Frontier models perform below chance (max 63.5%) compared to fresh researchers (73.4%), often preferring flawed original plans.