Research service · 4 November 2026
Inside our private research scorer.
The private research scorer turns an image into a versioned record of problem scores and uncertainty. This reference describes our internal evaluation tool. A customer-facing scoring API is a possible future product. Our offer today is a private evaluation and a report of findings. The examples reflect code inspected on 4 November 2026.
Explore the API contractSelf-hosted research integration
An operator runs the model checkpoint through the local PyTorch or ONNX scorer and supplies the service URL and server key. Set CROP_VISION_URL to that deployment. Keep the key on your server. Customer endpoints and keys would belong to a separate future scoring API.
Send one photo
POST /score accepts multipart fields image and photo_key. The reference must contain 1–200 printable characters. The upload must declare Content-Length. JPEG, PNG, WebP, HEIC and HEIF content types are accepted; successful decoding depends on the installed image codecs.
Image limits: 15 MiB and 25 megapixels. The JSON alternative accepts photo_key and image_url, capped at 16 KiB. Remote images require an operator-approved HTTPS host and public network address; redirects are rejected.
curl "$CROP_VISION_URL/score" \
-H "Authorization: Bearer $CROP_VISION_API_KEY" \
-F "photo_key=your-photo-reference" \
-F "image=@planter.jpg;type=image/jpeg"Use these examples in your own server environment.
Preserve provenance and uncertainty
| Field | Meaning |
|---|---|
| photo_key | Your photo reference, echoed back. |
| model_id | The loaded model identifier. |
| probs | Probabilities by taxonomy kind. |
| problem_prob | The overall problem probability. |
| predicted_kind | Predicted kind, or null when abstaining. |
| abstain / abstain_reason | below_threshold or ambiguous; null reason when a prediction is returned. |
| threshold / img_size | Threshold and input image size used by the scorer. |
| scored_at | UTC scoring timestamp. |
Abstention is not a diagnosis. A below_threshold result means the model did not cross its configured alert threshold. An ambiguous result means the top problem scores are too close. Route uncertain cases for expert review.
Handle failures
- 400
- Invalid photo reference, JSON or upload fields.
- 401
- Missing or incorrect bearer token.
- 411 / 413
- Missing upload length or request too large.
- 415 / 422
- Unsupported content type or rejected image.
- 502 / 504
- Remote image fetch failed or timed out.
- 503
- Authentication or the optional backend is not configured.
What is still planned
Severity, recommended actions, customer API keys, usage billing and a public /v1/score endpoint belong to the product roadmap. They are not fields or features of this HTTP scorer. The scoring library supports optional image embeddings, but this HTTP route does not request them. We scope customer evaluations separately and keep raw field records in-house.
Source inspected: crop_vision/serve/server.py and crop_vision/serve/score.py. Source review only; no live inference was run for this page.