TensorTest
TensorTest provides private evaluation services for AI labs, testing unreleased checkpoints and agents on real-world professional tasks. Its first benchmark, TimelineBench 1.0, evaluates video-editing agents on 56 real editing jobs, where only 14.0% of runs resolved despite 93% of agents claiming success. The company also offers commissioned training and evaluation data, including professional editors' trajectories and preference pairs.
- Artificial Intelligence
- Developer Tools
- Data & Analytics
- Software Only
Funding
Raised to date
$500KRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Founder details are not available yet.
Product
Problem
AI labs developing agents for real-world tasks lack rigorous, private evaluation methods that can identify when a model falsely claims success. Standard benchmarks often use synthetic or simplified tasks that do not reflect the complexity of professional work, leading to inflated performance estimates and unreliable deployment decisions.
Solution
TimelineBench provides private evaluation services for AI labs, testing unreleased checkpoints and agents on real-world jobs drawn from professional contexts. Each evaluation run produces a detailed failure report identifying where and why an agent failed, with results kept confidential to the client. The company also offers commissioned training and evaluation data, including professional editors' trajectories, preference pairs, and rubrics, with terms scoped per engagement. Its first benchmark, TimelineBench 1.0, evaluates video-editing agents on 56 real editing jobs, requiring each agent to produce a finished cut from raw footage, audio, and a client brief, with resolution requiring passage of delivery tests, content guards, brief must-haves, and a quality bar calibrated against blind judgments by 43 professional editors.
Target Audience
Primary customers are AI labs and model developers who need rigorous, private evaluation of unreleased checkpoints and agents, particularly those building agentic systems for professional workflows such as video editing.
Features
- Private evaluation runs for unreleased checkpoints and agents with failure reports delivered to clients
- Commissioned training and evaluation data including professional editors' trajectories, preference pairs, and rubrics
- TimelineBench 1.0 benchmark with 56 real video-editing jobs across four collections: EditStock, Cinestudy, UGC, and Commercial
- Multi-layer automated scoring: 4 delivery tests, 6 content guards, code- or judge-verified brief must-haves, and a quality bar using a three-lab judge panel
- Frozen scoring constants and verifier that reproduces published results and scores new runs consistently
- Public leaderboard with exact 95% confidence intervals and resolution rates for 19 agent setups, including models from Anthropic, OpenAI, xAI, Google, Zhipu AI, DeepSeek, and Alibaba