Expand description
Turn understanding as small verified model tasks.
An Understander reads one turn through a fixed set of narrow tasks run by the
turnframe-tasks engine, then assembles their answers into one
Understanding by code. Models judge
language; code checks structure: a task points at the user’s words instead of
quoting them, gives dates and amounts as expressions code evaluates, and chooses only
from closed sets built for its call. A verifier can take an act away, never add one.
| Module | What it holds |
|---|---|
input | what a turn’s understanding may see |
words | the user’s words, numbered, and pointers into them |
tasks | the task kinds, their prompts, schemas and checks |
values | from an extraction to the operation’s values |
check | the domain’s own check of an act |
progress | the steps a turn’s understanding publishes as it runs |
pipeline | the order tasks run in, and the settings |
Re-exports§
pub use check::ActChecker;pub use check::NoChecks;pub use input::CardOption;pub use input::Expectation;pub use input::OpenCard;pub use input::PendingAct;pub use input::PreviousReceipt;pub use input::RecordBrief;pub use input::Speaker;pub use input::TranscriptMessage;pub use input::UnderstandingInput;pub use input::WorkflowBrief;pub use pipeline::Settings;pub use pipeline::TurnUnderstander;pub use pipeline::Understander;pub use pipeline::VerifyPolicy;pub use progress::ChannelSteps;pub use progress::NoSteps;pub use progress::RecordedSteps;pub use progress::Step;pub use progress::StepSink;pub use words::Span;pub use words::Words;
Modules§
- check
- The domain’s own check of an understood act, run before the plan exists.
- input
- Everything one turn’s understanding may see, built by the runtime from stores and views.
- pipeline
- The understanding pipeline: segment; then route, frame and coverage side by side; then one chain per act, all concurrently; then assembly by code.
- progress
- The steps of an understanding as it runs, for a consumer that wants to show them.
- tasks
- The understanding task kinds. Each renders its context, states its answer schema with the closed sets of this call, and checks answers structurally.
- values
- From an extraction to the operation’s values: slicing words, evaluating dates and amounts, resolving handles. Every failure is structural and worded for a repair.
- words
- The user’s words, numbered, so a task can point at them.