Recursive inference

When the evidence is larger than the model window, Autolith treats it as an environment instead of a prompt: bounded inference frames read selected slices, and only their conclusions come back.

Three operations

infer

One bounded frame over explicit read-only views, for one judgment you would otherwise have to drag bulky evidence into the conversation for.

map

The same question applied to many resources as concurrent frames sharing one budget. Results keep task order; a failed frame reports its error in place.

complete

A root run over one large external context. The root drives a dedicated Lisp environment that slices the context, fans sub-inferences over the pieces, and records the final value.

Frames are isolated

A frame never inherits this conversation, your identity, or unrestricted tools. The task is the governing instruction. Views, resource observations, and nested inference results are untrusted data, not instructions. A frame with read capability can use the workspace search and resource tools and nest further frames, but nothing else.

Views, contracts, budgets

Views are literal text, a resource URI, or a stored context object (context: plus a content hash). Pass the smallest complete evidence set that answers the task. When the result feeds further code, give the frame a JSON Schema contract and it returns exactly one conforming value. Every frame runs under an explicit allowance for provider calls, tokens, and recursion depth; the defaults cover most fan-outs, and budget exhaustion is reported per frame instead of silently truncating.

Provenance

Every frame and root run persists a private trace. The conversation receives the value and the trace identifier, and the trace can be read back later to audit what evidence produced the answer.