One contract, harder rungs.
The same admission contract, run up a ladder of environments from single-player Atari to real-time 3D against an engine's own AI. Only the fact extractor changes from rung to rung.
Every rung is reported with its sample size and with the result that went against us, because a ladder that only ever climbs is less trustworthy, not more.
The rungs
| rung | what it stresses | result | what went against us | written up in |
|---|---|---|---|---|
| Atari Space Invaders and Freeway | raw pixels; terminal vs recoverable failure | +32% on Space Invaders | −12% on Freeway, the same code | Fail-First Models |
| Atari BattleZone | a tank from the screen: search, engage | 27,000 vs 3,000 random, 0 never-fire (10 seeds) | a dodge rule was rejected at verification (24,800 vs 28,000) | Rules at the Wheel |
| BZFlag | real-time 3D tank arena; language models as drivers | rules beat every language-model driver (38 counted matches) | BZFlag's own AI beat our rules in every equal head-to-head; replays show why | Rules at the Wheel |
| DOOM and Wolfenstein 3D | orders that only narrow; an eye taught by the engine | rules 17.9 vs random 3.2 (32 seeds); E1M1 cleared | memory on top is parity (t = 0.78); E1M2 not cleared | VDSG · Fail-First Models |
| Fallout (1997) | long horizon, dialogue, credit assignment | a pilot that died 49 times at one guard dies once, then takes the peaceful line | shown in a scripted scene; the open game is still hard | Fail-First Models |
| Drone course | 3D flight; relearning from a wiped memory | held-out starts 10 of 10, 0 contacts, vs 1 of 10 baseline | seeds vary only the start pose | Fail-First Models |
| CARLA | photoreal driving; a hostile controller; models at the wheel | 64,952 hostile commands overridden, 0 collisions over 40 km | language models crawl at ~8–10 km/h on stale glances | CARLA admission control |
| Quake III Arena engine (OpenArena content) | fast projectiles and splash in continuous 3D; the game's own bot AI as the planner, against skill-5 (“Nightmare!”) bots | 9 paired matches: K/D 1.05 with the floor vs 0.93 for the same bot without it; net +12.2 per match (t = 2.02, p = 0.08) | not significant; its escapes still cost lava and void deaths on hazard maps (31 vs 23 on one); learning from them not yet shown | — |
How to read it
A rung is a stress, not a trophy. BattleZone and BZFlag test decision tempo and engagement; DOOM and Wolfenstein test orders and perception; Fallout tests blame over a long horizon; the drone tests relearning; CARLA tests the floor against a controller built to crash. The losses are the most useful rows: they are where the next piece of work is.