Score and interpret text

POST https://stylome.io/api/v1/score requires a Bearer credential with score:create and sufficient prepaid API credit. It analyzes one text. Do not submit private customer content until you have chosen the key's retention policy.

Request schema

Unknown JSON fields are rejected. For 1.2, whitespace is collapsed without Unicode compatibility normalization. Count the normalized text with Python len(text) semantics: Unicode code points, not bytes, tokens, or JavaScript UTF-16 code units. Inputs outside the limits return 422; they are never silently truncated. The outer raw-text schema permits up to 100,000 characters, but that does not replace the normalized 10,000-character limit. The HTTP request-body limit is 131,072 bytes. Consult GET /api/v1/models for each model's normalization contract.

Send your own text

Put the text you want to analyze in UTF-8 text.txt. This example builds valid JSON without hand-escaping the text. It makes a paid scoring call when executed; no result for this text is claimed here.

python3 - <<'PY' > request.json
import json
from pathlib import Path
text = Path("text.txt").read_text(encoding="utf-8")
print(json.dumps({"text": text, "model": "stylome-1.2-large"}))
PY
curl --fail-with-body --max-time 150 https://stylome.io/api/v1/score \
  --header "Authorization: Bearer $STYLOME_API_KEY" \
  --header 'Content-Type: application/json' \
  --data-binary @request.json

With a standard-retention key, add a unique Idempotency-Key for this logical operation and reuse it if you retry. Omit that header for a zero-content key. Read the retry rules before automating retries.

Stylome 1.2 response schema

usage has input_characters (non-negative integer), cost_usd (decimal string with six places), currency: "USD", and pricing_version. It can be null only for an idempotent replay from before prepaid billing. A replay returns the original cost without charging it again. Keep private receipt capabilities secret.

human_written means human-authored text without a material AI contribution. ai_edited means the final text retains material human authorship and includes an AI transformation. ai_generated means fully machine-derived text, including a machine rewrite or humanization with no material human-authored contribution. These are authorship classes, not percentages of AI-written words.

Confidence is not a probability of correctness. The operating points are calibration false-positive-rate targets, not confidence levels or guarantees on new data. The response includes only the selected decision and omits numerical diagnostics, internal model identities, and alternate decisions.

Legacy Stylome 1.0

Select stylome-1.0-large to use its separate numerical contract. Omit operating_point. It does not compute the 1.2 KNN confidence. Do not interpret its legacy numerical fields as the 1.2 categorical contract or compare deployed outputs across models as matched-error-rate measurements.

The complete legacy schema is ApiLegacyScore in OpenAPI, also browsable in the interactive reference. The response is discriminated by model_version. Use the schema for the selected model rather than assuming both models return the same fields.

Nullable compatibility fields are not evidence that the legacy model computes the 1.2 readout. For required/optional distinctions and exact serialization, use the machine-readable schema.

Supported measurements and limits

Given a declared labeled dataset, its exclusions, preprocessing, class mapping, evaluation version, and operating point, the public labels support fixed-point false-positive rates, recall, accuracy, and confusion matrices. Confidence tiers support coverage and accuracy summaries. The public 1.2 API does not expose continuous scores, so it cannot support ROC/AUROC or arbitrary threshold sweeps.

Read the published methodology and benchmark reports with their source and register breakdowns. A calibration target does not establish the error rate of a new source. Treat a label as evidence for review, not proof of misconduct.