Paracelsus for Machines: Dose, warning and the way back in learned systems
Year: 2026
Type: preprint
Abstract: Models are edited, steered and fine-tuned after training, yet the side effects of these interventions are rarely measured in a standard way. We borrow four instruments from clinical pharmacology and test each on two kinds of learned system: a dose–response law, an assay floor, a monitored dose escalation, and the cost of withdrawal. In four open language models (1.5 to 7 billion parameters), the divergence caused by a residual-stream patch grew as the square of its dose, with exponent 2.00 in float32. In bfloat16 a rounding floor added to that divergence and hid the law. A certify-then-monitor rule kept 100% of held-out checks within budget. In a neural cellular automaton trained to grow and regenerate a one-headed or twoheaded worm, a 20-step pulse to an internal voltage channel wrote a second head. The new form survived removal of all written tissue in 99–100% of 128 fresh worms, across three independently trained rules. Whole-body divergence accelerated before the first change of form in 6 of 6 rule-and-direction pairs, and in 192 of 192 individual worms. Undoing the write cost 2 to 8 times less while tissue was regrowing than after it had regrown. Every prediction was fixed before any of its results existed. The three language-model tests failed as registered, and 7 of 20 automaton predictions were spoiled by our own instruments; all are reported. The automaton is a toy, the record was not timestamped by a third party before the runs, nothing has been replicated independently, and no claim here is clinical.
Author: Castro, César
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: dose escalation, fine-tuning, model monitoring
Open Access status: green
OpenAlex ID: W7234225617
