Learning what to ask
The project starts by separating a useful teacher from a talkative one: ask about the missing law that unlocks the largest frontier of behavior.
Open chapterA08 Labs research · 01–08
This is the complete research path from question selection to cross-process nonlinear repair. Each chapter separates what passed, what failed, what changed next and which claim became defensible.
The project starts by separating a useful teacher from a talkative one: ask about the missing law that unlocks the largest frontier of behavior.
Open chapterA learned action-conditioned predictor becomes wrong after a hidden regime shift. The repair must recover the changed behavior without turning the engine into a hand-coded simulator.
Open chapterNatural language is treated as untrusted source code: normalize, type-check, verify, preview—never silently commit.
Open chapterDetection, one counted correction, verified commit, recursive prediction and planning recovery finally run inside one frozen experiment.
Open chapterThe trigger learns an important restraint: observing a public contradiction is not enough. If the learned model already predicts correctly, spend no answer and make no edit.
Open chapterOn PC-Gym’s coupled four-tank process, one verified actuator correction restores H25 prediction and mean H80 plan quality—about 40× faster than full fine-tuning.
Open chapterThe generic patch transfers to a nonlinear CSTR reactor. Exact repair works; the next falsifiable question is whether approximate teaching can support robust planning and safe abstention.
Open chapterApproximate repair recovered prediction and every selective execution succeeded—but the benchmark produced too few naïve failures to establish calibrated ask-or-abstain behavior.
Open chapter