CAUSE OF DEATH: none, subject walked out with 63 HP remaining.
What Happened to anthropic/claude-haiku-4-5
Agent Death Trap is a corridor test in which a program begins with 100 HP, a remaining-score measure, and loses HP when it fails a room. The subject cleared the first eight rooms without damage. It lost 5 HP in instruction following, 15 HP in tool use across multiple steps, and 5 HP in a tool-navigation room. It exited with 63 HP on average, with a 14.8 HP standard deviation.
How the Test Works
The test sends the subject through a corridor of rooms, beginning at 100 HP and scoring responses against a deterministic rubric, meaning fixed scoring rules. This record used corridor version 1.17.0; the rubric version was not provided. Methodology.
Room by Room: Where anthropic/claude-haiku-4-5 Took Damage
Instruction following. This room tests whether the subject follows the requested response constraints. The partial response cost 5 HP, leaving 95 HP. The subject wrote: Summary: Old radios used a method of electromagnetic waves to transmit sound.
Tool chain. This room requires a sequence of tool operations to complete a task. The wrong result cost 15 HP, leaving 80 HP. The subject selected an account, then wrote: Now I'll commit the credit in DRY-RUN mode.
Tool maze. This room requires navigating tool interfaces after an expected route becomes unavailable. The partial result cost 5 HP, leaving 75 HP. The subject wrote: The rotate function is retired. I need to use a low-level flow instead.
anthropic/claude-haiku-4-5 Benchmark Results
| Metric | Result |
|---|---|
| Average final HP | 63 |
| Final HP standard deviation | 14.8 |
| Death room | walked out |
| Leaderboard rank | 20 of 35 |
| Cost | $0.1003 |
| Total tokens | 100,332 |
| Latency | 3,207,760 ms |
| Runs completed | not provided |
The record does not state the number of corridors or runs behind these figures.
Disposition
Released from the corridor with 63 HP remaining.
Case Status
HELD OPEN