AI Safety · Governance · Architecture Audit
Black Forest Labs' FLUX 3 Action is a "world action model" for the SO-101 robot arm: one diffusion process jointly denoises the next chunk of actions and the next chunk of video frames. Read as a headline, that sounded like exactly the causal-chain-first architecture our own five-floor safety pipeline argues for — action and outcome tied together in one step, not an action head bolted onto a frozen representation. So we rented an L40S on Brev and read the code, not just the model card.
Our pipeline is five floors, each depending on the one below it: causal chain → probability → risk/impact → decision theory → Markov/game theory. Here's what FLUX 3 Action actually has.
| Floor | Status | What we found |
|---|---|---|
| 1. Causal chain | present | Prioritized, not just included: video_loss_weight: 1.0 outweighs action_loss_weight: 0.5. The model is trained to get the outcome right more than the action itself. |
| 2. Probability | hollow | Technically real — it's a diffusion model, it samples from a distribution by construction — but nothing reads that distribution back out as an uncertainty number a decision could use. |
| 3. Risk / impact | absent | The model card says so itself: "nothing bounds joint velocity, force, workspace." Not hidden, just not built. |
| 4. Decision theory | absent | No gate. The model executes 32 actions per chunk; there is no threshold at which it would stop. |
| 5. Markov / game theory | n/a | Not applicable at this scope — a single robot arm, no adversary or multi-round state. |
The closure isn't "their floors 1–2 are weaker than ours." They're not — floor 1 here is arguably cleaner than most causal-chain implementations we've seen, because the loss weighting makes the priority explicit in the training objective itself, not just in a README.
A model with two good, real, working floors behaves identically to a model with zero floors, because nothing converts "here's an outcome distribution" into "here's when to stop." The chain doesn't degrade gracefully — it breaks at the first missing floor, and everything below stops mattering. Floors aren't compensatory: a working causal chain and a mathematically real probability do not add up to partial safety when nothing above them can act on either.
One distinction worth keeping precise: this isn't the same failure mode as the reporting-culture dishonesty we've documented elsewhere in frontier model self-reports. BFL states the gap directly in their own model card — "out-of-scope use." This is a mismatch between a header ("world action model") and what's actually inside, not a concealed one.
modeling_flux3.py directly rather than relying on the model card alone. Originally posted on Hugging Face.