01
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 chapterPassed after registered replication29/30world-level AUC wins
A08 Labs research · 01–08
The claim was built one failure boundary at a time.
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.
PassedMixed / failed conjunctionDevelopment only
02
From executable laws to learned dynamics
A 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 chapterMixed: core repair positive, key failures preserved100%affected accuracy
03
Compiling language into a verifiable change
Natural language is treated as untrusted source code: normalize, type-check, verify, preview—never silently commit.
Open chapterDevelopment evidence; formal human confirmation deferred60/60exact controlled drafts
04
The closed repair loop
Detection, one counted correction, verified commit, recursive prediction and planning recovery finally run inside one frozen experiment.
Open chapterPassed54/60hybrid target recovery
05
Ask only when the deployed model is actually wrong
The 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 chapterTrigger passed; model-scale conjunction failed 18/1958/58exact actionable commits
06
External benchmark: nonlinear four-tank control
On 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 chapterContinuous recovery passed; formal confirmations failed 9/1024/24detect + exact commit
07
A second process—and the uncertainty frontier
The 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 chapterCross-process evidence positive; audit failed 6/74/4exact CSTR repairs
08
When uncertainty is not yet the bottleneck
Approximate repair recovered prediction and every selective execution succeeded—but the benchmark produced too few naïve failures to establish calibrated ask-or-abstain behavior.
Open chapterDevelopment failed 5/9; confirmation unopened95.5%H25 competence recovered