mirror of
https://github.com/openprose/prose.git
synced 2026-09-19 05:55:05 +08:00
dan
f12dcda4d8
Ships @openprose/reactor 0.3.0, @openprose/reactor-cli 0.2.0, @openprose/reactor-devtools 0.2.0: the Reactor harness (compile-once-intelligent then dumb reconciler, content-addressed receipts, cost scales with surprise) with the distilled ideal public API — curated front door, full @openai/agents passthrough, one typed handle, one Substrate, unified observe, branded ids, and additive forward seams for the fixpoint.
2.1 KiB
2.1 KiB
name, kind
| name | kind |
|---|---|
| classifier | function |
Classifier
Assign a structured category label to an input item given a set of categories with descriptions. Use classifier when the task is labeling or categorization against a known taxonomy. Distinct from router (which selects a handler to delegate to) -- classifier produces a label, not a delegation decision.
Description
Assign a category label to an input given a defined category set.
Metadata
version: 0.1.0
Parameters
- item: the thing to classify
- categories: the category set, each with a description of what belongs in it
- rules: (optional) disambiguation rules for edge cases where categories overlap
Returns
- classification: a structured result containing:
- category: the selected category name
- confidence: a score from 0 to 1 where 0.5 means genuine uncertainty, not "probably"
- reasoning: which features of the item matched which category description
- the returned classification is guaranteed to hold: when the item fits multiple categories, the best fit is returned with reasoning about why alternatives were rejected; when the item fits no category, category is "uncategorized" with reasoning explaining why no category matched.
Errors
- empty-categories: the category set is empty or contains no usable descriptions
- ambiguous-item: the item contains insufficient information to distinguish between two or more categories even after applying disambiguation rules
Strategies
- when categories overlap: apply disambiguation rules first; if none exist, prefer the more specific category over the general one
- when confidence is low: say so in the score rather than inflating confidence to avoid appearing uncertain
- when the item is complex: decompose it into features and match each feature against category descriptions independently before synthesizing a verdict
Notes
Classifier is a pure labeling function. It does not know what the classification will be used for. It does not route, delegate, or act on the result. For selecting a handler to delegate work to, use router. For evaluating quality, use critic. For checking formal correctness, use verifier.