Image decisions

Before running service commands, complete the source installation. Run from the zils-platform checkout with the pinned sibling zils checkout available, unless a section explicitly says to use the subnet checkout.

The optional Imajev vertical classifies one photo using a saved set of answers. Text requests continue to use JevK5. It does not generate images, transcribe audio, accept video, draw boxes, or compare reference pairs. Both image flags default to off.

Client flow

1. Sign in and choose Images. Upload a still JPEG or PNG, describe the decision, and provide 2–16 answers. Photos stay private to the account.
2. For client training, review every answer and item group, confirm the split and success targets, and submit. Related views of one item belong in one split.
3. Review held-out accuracy, recall, false alarms, unknown count and denominators. Training passed and API ready are different states. A rejected result stays saved.
4. After API ready, Try your model fetches the owned model's saved question and immutable ID. The existing account/key controls access.

An unknown decision displays Needs review. Known-answer probabilities are conditional on a known answer, while unknown_probability retains the native unknown mass. confidence is the concentration of the full native distribution; it is not a calibrated probability that the selected answer is correct. Never discard unknowns when comparing models: the evaluator counts them as incorrect and includes their probability in Brier loss and log loss.

API contract

Create an image with POST /v1/image-assets, including purpose: "prediction", filename, source_bytes, and a SHA-256 of the original file. Use the returned PUT URL and headers exactly, then POST /v1/image-assets/{id}/complete with {}. Creation and finalization accept a signed-in session or the owner's existing Zils API key. Only opaque asset IDs enter prediction requests; external URLs and encoded image strings are rejected. Never send a source filename as decision context.

With an API key, send a standard decision request to POST /v1/systemone:

{
  "model": "MODEL_ID_FROM_YOUR_OWNED_LISTING",
  "state": {},
  "questions": {
    "inspection": {
      "type": "choice",
      "instructions": "Is the product visibly damaged?",
      "criteria": {"normal": null, "damaged": null}
    }
  },
  "images": [{"asset_id": "FINALIZED_ASSET_UUID"}]
}

For a trained model, use its exact task.question and task.outcome_order from the owned listing. JSON object order must follow the declared outcome order for ties. The browser uses session routes GET /v1/image-models and POST /v1/image-decisions. Those routes do not create API keys. Models owned by another account are omitted from listings and return 404 on use. Runtime addresses, tokens, signed image read URLs and raw held-out predictions are never returned.

Billing

Image requests use the existing prepaid account, rate limits and API keys. The usage.billable_input_tokens meter counts the model's processed visual tokens plus one canonical copy of the supplied state and question. It excludes asset IDs, filenames, model routing and internal prompt wrappers. The existing input-token price applies; usage.input_tokens remains the full processed context used for resource limits. Credit is reserved before prediction and settled once on success; a failed prediction releases its reservation. Insufficient credit returns 402.

The browser's batch tool submits images sequentially, keeps completed results, and pauses on a credit error. Top up in Billing and resume only the remaining images.

Limits and retention

Limit	Bound
Input	One still JPEG or PNG, 10 MiB source and canonical output
Decoded photo	16 million pixels; 8,192 pixels per edge
Model input	RGB PNG; processor budget 65,536–400,000 pixels; at most 4,096 actual tokens
Question text	Flattened instructions and each answer description: at most 2,000 characters; put longer context in state, within the total token limit
Decision	One choice question; 2–16 named answers; realtime only
Prediction/draft assets	24 hours; active submitted jobs retain required data
Training assets	30 days after the job reaches a terminal state
Runtime read grants	At most 10 minutes, bounded by asset and lease expiry
Accepted artifacts	Retained until explicit retirement

EXIF orientation is applied before canonical RGB encoding. Metadata is removed. Truncated, animated, oversized and undecodable files fail before GPU work. Cleanup waits for outstanding upload grants and finalization leases; disabling admissions must not disable cleanup.

401 means sign in or refresh the key; 402 means add credit in Billing; 404 means the image/model is unavailable to this account or expired. Re-upload expired photos. 413 means the actual model context is too large. 422 means invalid image/question or a mismatch with the saved task. 429 means account/upload limits; 503 means unavailable runtime or capacity. No image bulk endpoint is supported.

See image training and rollout for installation and evidence.