What we have actually measured.
Four domains, one architecture, every figure produced in-house and reproducible from a frozen experiment card. The negative results are here because a reviewer will find them anyway — and because a result that only ever improves is less trustworthy, not more.
The four
| domain | result | reading |
|---|---|---|
| Driving CARLA, adversarial controller | hits = 0 100–490 m at 38–40 km/h · ~9.7 m gap held · ~400 refusals logged | The floor takes the vehicle from a hostile planner and holds separation. Floor OFF, the same planner crashes. |
| Space Invaders raw 210×160 pixels, zero emulator RAM | 200.6 vs 152.2 +32% · peak 276.2 @50 episodes | Learned refusal helps where a failure is terminal. |
| Freeway identical mechanism, identical code | 9.2 vs 10.4 −12% · 3 rules formed, every one blocking up | It hurts where the only scoring action is also the dangerous one. |
| Doom rule-based, no ML | random 3.2 · rules 17.9 · rules+memory 20.3 paired t = 0.78 | Rules beat random decisively. The memory layer on top is parity, not a win, and is reported as such. |
The pair is the finding
Invaders and Freeway are not two experiments. They are the same code pointed at two ROMs — Freeway imports the Invaders failure memory unchanged, with no per-game strategy. The architecture transferred. The learning did not.
Terminal failure
A death in Invaders removes every remaining reward. Its true cost is the expected remaining return, so an absolute veto prices it correctly — by accident. An infinite penalty is right when the loss really is everything.
Recoverable failure
A collision in Freeway knocks the chicken back and play continues. The cost is the lost ground, nothing more. The absolute veto prices it wrongly and removes the only scoring action.
No single failure price is correct in both regimes — and a confidence threshold is a single failure price. This is the most transferable result we have, and it is an argument against the industry's default gate, not just against ours.
A defect we published
The emulator drops its lives counter at the end of a 127-frame death animation, so every failure was recorded from frames in which the agent was already destroyed. Of 374 blamed frames, 0 had the hazard visible. At the true impact frame, 17 of 17 did. We were learning from the wrong moment entirely — and still producing a curve that looked like learning.
Five such defects are recorded on the experiment cards. Each one silently returned a believable result before it was caught. We list them because credit-assignment error is the failure mode most likely to be hiding in any system of this kind, including a competitor's.