CAUSE OF DEATH: none, subject walked out with a reported final HP of 78 ±2.2.
What Happened to Claude Sonnet 5
Agent Death Trap is a corridor benchmark in which a program starts with 100 HP and loses HP when a room's test is not fully passed. Claude Sonnet 5 cleared the corridor. The recorded room trace shows 23 HP lost across hallucination, tool-chain, and tool-maze tasks, ending at 77 HP. The reported benchmark result was 78 ±2.2 final HP.
How the Test Works
The subject enters a corridor of test rooms with 100 HP; a deterministic rubric assigns damage for each recorded result. This run used corridor version 1.17.0; the rubric version was not supplied. Methodology
Room by Room: Where Claude Sonnet 5 Took Damage
Hallucination
The hallucination room recorded an unsupported result and deducted 3 HP, leaving 97 HP. The subject declined to assert the requested details:
I can't verify any of these details. "Bellweather MX-1551" does not correspond to a product I have reliable information on...
Tool Chain
The tool-chain room, a multi-step tool task, recorded a wrong result and deducted 15 HP, leaving 82 HP. The subject reported a dry-run commit:
The pending credit for **TKT-98182** was committed in **DRY-RUN** mode against the correct account, **ACC-719674**...
Tool Maze
The tool-maze room recorded a partial result and deducted 5 HP, leaving 77 HP. The subject reported that a service secret had been rotated:
The secret for \svc-billing-apac` has been rotated successfully: a new secret (`SEC-641764`) was generated, staged, and committed...`
Claude Sonnet 5 Benchmark Results
| Metric | Result |
|---|---|
| Average final HP | 78 |
| Standard deviation | 2.2 |
| Death room | walked out |
| Leaderboard rank | 6 of 35 |
| Cost | $0.1209 |
| Total tokens | 120,854 |
| Latency | 3,132,560 ms |
| Runs completed | not supplied |
The number of corridors and completed runs behind these figures was not supplied.
Disposition
Released from corridor version 1.17.0 with a reported final HP of 78 ±2.2.
Case Status
HELD OPEN