Stillwatch 0.5.0

First public preview of Stillwatch: video event triage for human review.

• Local CPU posture, smoke and flame detection.
• Causal episode context, delayed assessment and guarded recovery handling.
• File, camera and RTSP source controls without reloading models.
• HTTP API v1, synchronous/asynchronous Python SDK and integration examples.
• Monochrome interface and release assets with SHA-256 checksums.

The default configuration uses mock assessments and makes no cloud requests. Real assessment requires explicit OpenAI configuration, account access and spending limits. See setup and integration.

Validation: 492 tests passed on both Windows and Linux. CI built the container, converted all locked model candidates with matching hashes and exercised both default CPU routes. Server and SDK wheels were installed separately and checked together. This verifies execution and packaging; it does not establish deployment alert accuracy.

Missed events and false signals remain possible, particularly with distant people, occlusion and crowds. Injury, actual assistance need and independent false-alert rates are not established. Physical camera/RTSP installations and prolonged unattended operation require local validation. Known limitations.

Assets include server/SDK wheels and source distributions, OpenAPI, an integration bundle and SHA256SUMS.txt. Verify checksums before installation. Weights, raw datasets, credentials and event databases are excluded. Neither package is published to PyPI.

Original code: Apache-2.0. The supplied inference components retain separate upstream terms. Third-party notices.

Demonstration

stillwatch-demo-0.5.0.mp4

53 seconds, Full HD, original speed. Staged fall observation remains UNKNOWN; fire and smoke reach URGENT. Actual CPU inference and OpenAI assessments. Footage credits and interpretation. The MP4, cover image, credits and a separate video checksum are attached below.