Glossary

Public Terms

Term	Definition
batch	An IR transform that vectorizes execution over selected input leaves.
checkpoint	A primitive that labels an intermediate value with a key and collection. It is transparent unless collect or inject is active.
collect	A context manager that captures checkpointed values during IR execution.
Collection	A namespace used by checkpoints, collect, and inject to decide which values belong together.
Cotangent	Feedback flowing backward through a pullback. In autoform, cotangents are usually text feedback.
Custom rule	A rule registered on a custom traceable function boundary to override pushforward, pullback, or batch behavior.
dce	Dead-code elimination, an IR transform that removes equations not needed by selected outputs.
Doc annotation	A schema description attached with field @ af.Doc("...").
Dynamic argument	An input leaf represented by a placeholder during tracing and provided at execution time.
Execute	The phase that runs an IR with concrete inputs through .call(...) or .acall(...).
factor	A primitive that multiplies the current path weight. It is neutral during ordinary execution and contributes to weighted results.
fold	A context manager that evaluates foldable primitive calls immediately during tracing and embeds the result as a literal.
inject	A context manager that substitutes checkpointed values from a provided dictionary during execution.
Instance-first DSL	The schema style where the schema is a value-shaped Python instance, not a separate output class declaration.
Intercept	A runtime hook around checkpointed values. collect captures intercepted values; inject replaces them.
IR	The intermediate representation produced by trace; it contains input variables, equations, and outputs.
lm_client	A context manager that changes the active LM client for lm_call and lm_schema_call.
memoize	A context manager that caches primitive results within its block. During tracing, it can deduplicate identical primitive calls.
pullback	An IR transform that propagates output cotangents backward to input cotangents.
pushforward	An IR transform that propagates input tangents forward to output tangents.
Pytree	A nested container/leaf structure that autoform can walk. Registered dataclasses can be pytrees.
PYTREE_NAMESPACE	The optree namespace reserved by autoform for user pytree registration.
sched	An IR transform that groups independent equations for concurrent async execution.
Schema	A pytree of schema leaves such as Str, Float, and Enum, used by lm_schema_call for structured output.
Static argument	An input leaf fixed at trace time by trace. The static value is a bool pytree matching positional input structure.
tag value	A hashable metadata value attached to equations during tracing.
tag	A context manager that activates one or more tag values for equations created in its block.
Trace	The phase that runs a Python function once with placeholders and records autoform primitive calls as IR equations.
Transform	A function that consumes an IR and returns another IR. Current IR transforms include batch, pushforward, pullback, sched, dce, and weighted.
weighted	An IR transform that returns (output, path_weight) for one concrete path.

Internal IR Machinery

These names are useful when reading internals or debugging a transform. They are not part of the everyday user surface.

Term	Definition
Boxing	The internal technique interpreters use to wrap values with transform-specific metadata. Users normally see the result only through public transforms.
IREqn	One recorded primitive application in an IR.
IRVar	A typed placeholder for a runtime value inside an IR.
Prim	A named primitive operation used as the dispatch key for execution and transform rules.
TraceBox	The internal wrapper used by the trace interpreter to carry an IRVar through Python code.
Tracer	The trace-time interpreter machinery that records primitive calls instead of executing them normally.
walk	The manual IR stepping interface used by execution internals and advanced debugging code. See Walk.