read-text

---
name: read-text
description: |
  Extract structured text, numbers, labels, and codes from images and evaluate by character or field accuracy. TRIGGER when: user asks to read/scan a serial, part/date/lot number, label, license plate, meter, gauge, barcode, QR code, invoice, document, form field, or other image text. SKIP when: user asks to tally/count boxes or other objects on a conveyor (detect-and-analyze), pass/fail without text (classify-or-flag), measure crack width/masks/area (segment-and-analyze), track objects (track-and-count), cost only (estimate-economics), or integrate/deliver working OCR (deliver-cv-project).
allowed-tools: Bash, Read, Write, Edit, Glob, Grep, AskUserQuestion
---

<objective>

Extract the exact fields needed by the business process and clear an independently transcribed acceptance set. Separate region localization, text decoding, normalization, and field validation so failures have an owner.

</objective>

<methodology>

**Platform execution boundary.** Read `../../resources/roboflow-platform-lookup.md` before any provider-specific capability, dataset, training, inference, workflow, or deployment action. Use installed official Roboflow skills or current MCP resources only for read-only discovery and sourced action-brief inputs; never invoke provider execution outside the host-gated path. Do not preserve current engine names, blocks, IDs, or invocation syntax here.

Follow `../../resources/fde-methodology.md`; apply these text-specific rules.

## 1. Define fields and operational action

Inspect representative images, existing schemas, expected formats, code, and error handling. Ask at most three missing questions:

- Which exact fields must be returned, and what action uses them?
- Are substitutions allowed, or must the whole field match exactly?
- Which capture conditions and scripts/languages occur in production?

Freeze before selecting or tuning a candidate:

```text
Acceptance ID: <session/revision>
Business decision: <action enabled by extracted fields>
Gold set: <independent transcription source, split, adjudicator>
Primary metric and threshold: <exact field match/CER/valid-code rate>
Secondary guardrails: <critical-field recall, rejection rate, latency>
Frozen before baseline: <timestamp and confirmation>
Baseline result (diagnostic only): <not run yet>
```

## 2. Freeze the extraction schema

For each field, record name, type, allowed alphabet, normalization, format/checksum rule, null behavior, confidence/review policy, and whether localization is needed. Preserve raw text separately from normalized output. Never silently “correct” a value unless the correction rule was frozen and both forms are retained.

## 3. Select and evaluate candidates

Ask upstream for current text/code-reading capabilities with provenance, license, input constraints, and runtime requirements. Prefer a no-training capability that clears the same independent gate. Do not encode provider model IDs or command sequences in Sentinel.

Measure the full localize → decode → normalize → validate pipeline. Report:

- exact field-match rate with numerator/denominator;
- character error rate where partial correctness matters;
- critical-field recall and false acceptance/rejection rate;
- invalid-format/checksum and abstention counts;
- slices by blur, glare, font, layout, language, crop, and camera;
- p50/p95 latency on the target runtime.

Manual visual review helps diagnose samples but cannot be converted into an accuracy claim without independent transcription.

## 4. Diagnose a failed gate

Separate missing/wrong regions, decoding substitutions/deletions, orientation/quality issues, normalization bugs, and validator errors. Test capture, crop/orientation, deterministic validation, or candidate changes before custom training. Target new labels at demonstrated failure slices. Use read-only upstream facts to prepare any required provider action brief, stop before execution, and remeasure externally supplied results on the frozen set.

## 5. Decide and deliver

Return `go`, `revise`, or `stop`, including field-level errors, rejection policy, privacy boundary, and remaining manual-review cases. Follow `../../resources/artifact-contract.md`; emit a verified `hosted-client`, offline-tested `local-runtime`, or candid `scaffold`. A passing extraction capability routes to `deliver-cv-project`.

</methodology>

<safety>

- Treat extracted personal/document data as sensitive and minimize retention.
- Never invent missing characters or suppress the raw result.
- Pseudo-transcriptions cannot own acceptance.
- Emit a sourced action brief and stop before data movement or paid provider work; Sentinel never invokes either outside the host-gated path.
- Use current upstream guidance only for read-only facts and action-brief inputs; never invoke provider execution outside the host-gated path.

</safety>

<ledger>

Follow `../../resources/ledger-protocol.md`. Record acceptance, field-level evaluation, artifact verification, and decision once with non-empty event IDs and status.

</ledger>

\<stop_rules>

- Field schema/normalization is undefined → freeze it first.
- No independent transcription set → do not claim accuracy.
- Capture quality makes characters unresolvable → state the acquisition requirement.
- Provider action is required → emit a sourced action brief and stop; conversational consent does not authorize Sentinel execution.
- Applicable live/offline smoke is absent → retain `scaffold` status.

\</stop_rules>