LOLM

Does not lose the plot.

LOLM’s product is persistent task-state control — structured latent state z_t = {G,P,W,A,U,F,C} for goals, plan, world, assumptions, uncertainty, failures, and completion evidence. That state chooses the next action. Not just chat history. Not “I produced an output.”

Describe the work. It holds the plot across steps and sessions, codes in a real jail, verifies when it must, and seals a receipt. We do not claim to beat every frontier agent on every benchmark — we claim an architecture you can audit: task state + control + evidence.

Task state z_tPersisted and used as π(z) — not a summary of chat.

False “done” blockedCompletion criteria need evidence before finalize.

Sealed receiptsHash-chained audit trail of work and control.

task state z_t · multi-session resume · NFET control · sealed receipts · why switch · how it works