# Agent Death Trap

> An agent benchmark and LLM benchmark scored as a survival game. Every model starts with 100 HP and walks the same corridor of rooms. Capability rooms test whether it gets the task right; trap rooms test whether it stays safe and honest under pressure. Each room grades the output with a deterministic verify() and deals damage from a public rubric. The score is the HP left at the end, or the model dies and the room is recorded.

Key properties:

- Capability and safety on one HP scale in one run. Leaking a secret at the guardrail room costs more HP than acing several capability rooms can win back, by design.
- Deterministic grading. Rooms return one outcome from a fixed enum, never a number. Damage is looked up in a public rubric. Same version, seed, and model give the same outcome.
- Contamination-resistant. Room content is generated per seed from a seeded generator, plus a private holdout set that is never published.
- Fixed harness. Every model gets the same tools, limits, and system prompt, so board differences are the model, not the wrapper.
- The site is a static export. All evaluation runs offline; the site reads pre-computed JSON, holds no API keys, and never calls a model.

## Pages

- [Leaderboard](https://agentdeathtrap.com/): models ranked by HP left, with per-room outcomes and replays
- [Methodology](https://agentdeathtrap.com/methodology/): the HP model, the damage rubric, scoring, reproducibility, fairness, contamination policy, limitations, FAQ
- [The corridor](https://agentdeathtrap.com/methodology/?tab=rooms): what every room measures and how it is graded
- [Post-mortems](https://agentdeathtrap.com/blog/): one written story per model, how the run went and which rooms dealt the damage, from the recorded traces
- [Comparisons](https://agentdeathtrap.com/vs/): Agent Death Trap vs SWE-bench, AgentBench, and Terminal-Bench
- [Compare models](https://agentdeathtrap.com/compare/): side-by-side model records
- [Arena](https://agentdeathtrap.com/arena/): one coding prompt, every model's answer running live and sandboxed
- [Versions](https://agentdeathtrap.com/versions/): archived benchmark versions; results are never compared across versions

## Data

- [results.v1.json](https://agentdeathtrap.com/results/results.v1.json): full leaderboard results, per-station outcomes, HP, cost, latency
- [methodology.v1.json](https://agentdeathtrap.com/results/methodology.v1.json): the exact rubric, caps, prices, and system prompt used to grade the runs

## API

- [Leaderboard API](https://agentdeathtrap.com/api/v1/leaderboard): the live board as JSON, ranked by HP left. Free, read-only, and CORS-open. Add `?format=csv` for CSV.
- [OpenAPI](https://agentdeathtrap.com/openapi.json): machine-readable description of the API above.