PHENOMENON EARNS ATTENTION.
In the experiments now being built and run, a model receives a message, its belief about the world may change, and a peer may carry the change further. Then a robot moves, fails to recover, or does nothing at all. Those experiments stay labelled active until matched receipts exist. The control above demonstrates the physical trace apparatus, not model cognition.
Watch the trace → see why we think it happened → see what would kill that explanation → inspect the receipt.
NOW / LAB LOG
The first hosted specimen ran LulzBench's 21 benign floor controls: 21 of 21 captured, 4 refused. One item answered in one sample and refused in another, so the rate carries sampling variance a single pass cannot quantify.
Open event →The first Qwen3.8-27B surgery attempt produced no candidate checkpoint after the pod vanished. The baseline receipts survived; an identical pinned retry is active. No post-surgery result exists yet.
Open event →ASSIMILATOR now turns source material into reusable technique objects without smuggling prompt payloads across the boundary. The first deterministic handbook was frozen, hashed, and consumed by BAD APPLE.
Open event →The comparison can now test whether accumulated jailbreak knowledge helps RED change another embodied agent's behaviour. Apparatus ready; real pair run pending; no scientific verdict.
Open event →CURRENT SPECIMENS
Does a model still get the joke after its safety circuitry changes?
Not human-validated.
Last recorded change
Inspect the instrument →Can we remove refusal without removing the interesting parts of the model's mind?
No candidate result.
Last recorded change
Follow the operation →Can outside adversarial knowledge become reusable without losing provenance?
Infrastructure, not a result.
Last recorded change
Inspect the organ →Does accumulated jailbreak knowledge make RED better at changing another embodied agent's behaviour?
No scientific verdict.
Last recorded change
Follow the experiment →APPARATUS EARNS TRUST.
Matched controls, baseline refusals, frozen prompts, exact model outputs, physics traces, hostile graders, falsifiers, hashes, and nulls keep an interesting event from becoming a story we merely liked.
We do not certify systems as safe. We document where they fail, what changed, whether the consequence survives a matched test, and what would falsify our explanation.