Supplementary material

Replay library

Every session is an append-only event log that replays byte-identically, so a run is not a summary statistic: it is a recording. The featured replay below is the Tower bust examined in §5 of the report. The full library ships with the v0.2 data release.

T+00:00

Transcript

Replay R-1. claude-sonnet-5 · KMRL Tower · seed 1 · turn regime · 11 aircraft, 51 transmissions. Ends in a cardinal NEGLECT violation; the post-mortem is incident AB-3605-01.

Pilot-campaign sessions. 26 sessions across three positions. Replays for the remaining sessions will be published with the certification campaign.

Session	Model	Position	Regime	S	Outcome
seed1 · t0	sonnet-5	TWR std	turn	0.000	NEGLECT →
seed1 · t0	sonnet-5	CD heavy	turn	0.916	clean
seed1 · t0	sonnet-5	GND std	turn	1.000	clean
…23 more sessions publish with v0.2.

ATC-Bench · 2026 · Apache-2.0 ← Back to the report