DeepFakeShield API

Always 200 while this process answers requests. Touches nothing: no database, no detector backend, no FFmpeg. This is the endpoint a container health check or a reverse proxy should use, because /health deliberately returns 503 when a dependency is unusable and a proxy that is told 503 stops routing to a container that is perfectly alive. Use /health on a dashboard or a probe, where the dependency detail is the point.

Returns 200 when the deployment can do its job: the database answers, FFmpeg is present, every modality it serves has a selected engine that is available, and — only in FORENSICS_MODE=remote, which promises to fail rather than substitute — the neural sidecar is up. 503 otherwise. The forensics check reports the optional neural backend and carries required: false in auto and builtin modes, where a missing sidecar is a recorded fallback rather than an outage. Safe to poll; use /live for liveness.

Lists every engine that can answer for a modality, which one is selected, whether it is available, and — for engines with weights — the recorded SHA-256 and its integrity state.

Accepts multipart/form-data with a file part plus form fields, or application/json with a chunkedUploadId obtained from the stream endpoint.

The media is stored under a randomised key, hashed with SHA-256, analysed, and then deleted unless retainMedia=true. The response returns immediately with a queued analysis; poll it or read it later.

For sources that cannot produce a multipart body: send the bytes as the request body with Content-Type set to the media type and X-Filename. Returns a chunkedUploadId to submit with POST /analyses.

The durable record: media SHA-256, engine names, versions and weights checksums, configuration version, measurements, decision and its rule, and the declared limitations. ?format=txt renders the same record as a report; ?download=true sets a file name.

summary is written by code from the evidence record and is present whether or not any language model ran: the final decision, the score and how it relates to the thresholds, each channel, the input quality, the reason codes, every limitation that applies, and the calibration caveat. narrative is a model-written note and is present only if it passed the vocabulary guard; it is checked for forbidden claims, not for accuracy. A note that was discarded leaves status failed with errorCode VERIFICATION_FAILED and no narrative.

The same projection the dashboard screen is rendered from, so a figure here and the figure on screen cannot disagree. decisions holds counts; decisionLabels holds the display strings for the four states.

The token is returned exactly once and is not recoverable afterwards.

Refuses to remove the last active administrator.

Checks configuration, storage, FFmpeg, the database, engine availability and model integrity, and returns a per-check report.

Verifies Sensifai-Signature (HMAC-SHA256 over <t>.<raw body>) with a five-minute tolerance, deduplicates on the event id, and projects subscription and balance changes. Always answers 2xx for a verified event, including a duplicate.