Text signal
Visible phrases in an inspectable synthetic document.
Clinical tooling · Transparent baseline
Refusing to guess.
The first document classifier used visible rules, evidence, and a deliberate UNKNOWN result—not because rules are the destination, but because the categories still needed to be challenged.
The contract
A route must carry its trail.
Filename hint
Supporting evidence that never decides alone.
Threshold
A versioned, reviewable decision boundary.
UNKNOWN
A defined handoff to human review.
Refusal
“I cannot classify this” is useful information.
Ambiguous
Conflicting or weak signals remain unresolved.
Scanned
Unavailable text is reported rather than silently sent to OCR.
Human review
The workflow makes correction explicit.
Feedback
Disagreement tests the taxonomy itself.
Evidence matrix
Tests protect the documented behavior.
| Fixture | Expected behavior |
|---|---|
| Clear synthetic text | Route with matching evidence |
| Misleading filename | Filename cannot decide alone |
| Conflicting signals | Refuse or follow the documented threshold |
| Image-only PDF | Report text unavailable; do not pretend it was empty |
| Future model | Must beat the approved baseline on an approved evaluation set |
Guardrail: These implementation tests do not establish clinical performance. Taxonomy, weights, and workflow require subject-matter validation.
Watch
Refusing to Guess
Why CareFlow used a transparent rule baseline to learn the taxonomy, refusal behavior, and evidence contract before adding machine learning.