v0.3.0 — the world-model release

pip install -U trio-retina

The state layer of the world-model stack — now with shipped latent producers and a front-to-back demo, a CLI, a typed package, and a pre-launch hardening pass.

Added

• Latent producers ship — DinoV2Embedder (per-object entity.vec) and VJepa2Embedder (scene-level ws.scene): bring any encoder, out comes one standardized WorldState. pip install 'trio-retina[dino]' / [vjepa].
• The world-model stack, end to end — swap the encoder (YOLO · DINOv2 · V-JEPA 2 → same schema), a small transformer imagines the future off that state, front+back compose through one standard (examples/world_model/, BENCHMARK.md). Flagship soccer demo: real footage → WorldState → predicted player runs.
• retina CLI — retina demo / run / validate / bench (stdlib-only).
• Typed package (py.typed) + concise __repr__s; sample assets (retina.sample_events() / sample_video()); deeper docs (concepts · cookbook · cli · extend · faq).

Fixed (pre-launch hardening)

• WorldState.from_frame no longer crashes on extra vec/scene keys; LineRule pending-state leak + debounce/direction edge cases; schema.validate rejects bool-as-number / bad box / non-object input; Detection.from_supervision short-class_name IndexError; WebhookSink rejects non-http(s) schemes.

Changed

• [all] now includes matplotlib; honesty/wording fixes across docs.

Full notes: CHANGELOG.md

v0.2.1

Polish + edge-hardening release. pip install trio-retina

Changed

• RTSP/edge hardening — video_frames no longer dies on a transient live (rtsp:// / live=True / webcam) read failure: it reconnects with exponential backoff and drops to the latest frame under back-pressure. Finite files unchanged (real EOF still ends the stream, every frame delivered).
• README/docs headline quickstart now runs on a bare pip install trio-retina (numpy only, no model/GPU/video) — the YOLO + video form moved to a labeled secondary block.
• CountRule(3) works (threshold is now positional); CountRule(threshold=3) unchanged.

Added

• examples/bench_overhead.py — honest numpy-only micro-benchmark of the Retina-layer overhead (tracker + rules + event build, detector excluded): ~0.08 ms/frame at 20 tracks.
• tests/test_sources.py — live-source reconnect / drop-to-latest tested via a fake-capture seam (no cv2, no real RTSP).
• Colab notebooks (notebooks/) and the docs site synced.

Full notes: CHANGELOG.md

v0.2.0

Trio Retina — the model-agnostic state layer for world models. Turn any detector (YOLO · VLM · DINO) into one standard, queryable stream of events + latent state. numpy-only core.

📦 First PyPI release: pip install trio-retina

Added

• Detection.from_supervision() — ingest a Roboflow sv.Detections (duck-typed, no supervision import), so Supervision pipelines plug straight into Retina's event layer.
• LineRule(min_frames=...) — confirm a crossing only after N frames on the new side, suppressing jitter near the tripwire.
• Hero examples (run with no model / GPU / network): examples/rtsp_to_webhook.py (camera → restricted-zone alert → webhook), examples/from_supervision.py, and examples/latent_vec.py (populate the latent vec channel yourself).

Changed

• Distribution renamed to trio-retina (import stays import retina).
• [all] extra now installs every optional adapter (incl. Grounding DINO).
• Honesty pass: dropped the unreproducible forecast number; latent channel documented as a shipped interface with producers on the roadmap.

Full notes: CHANGELOG.md

v0.1.0 — first public release

Trio Retina is the model-agnostic state layer for world models: turn any detector (YOLO · VLM · DINO) into one standard, queryable stream of events + latent state. numpy-only core, runs on CPU at the edge.

pip install "retina-sdk @ git+https://github.com/machinefi/trio-retina"   # core: numpy only

A PyPI retina-sdk release is landing shortly.

Highlights

• The retina.event standard — tiny, JWT-like; ships a JSON Schema + pure-Python validator.
• Composable pipeline — wire detectors/trackers/rules/gates/enrichers/sinks with |, a node list, or a JSON workflow.
• Detectors: YOLO (Ultralytics), open-vocab (Grounding DINO), any VLM, or any callable(image)->[Detection]. Trackers: IoU (pure-Python) + Norfair.
• Rules: zone.enter / line.cross / dwell / count — now with exit_grace_s (rides out detection blips / id flicker) and anchor (center/feet/head).
• Dual state — symbolic events plus an optional latent (Vec) channel, never collapsed.
• Demos (all run with no model / no GPU): forecast (dynamics on WorldState) and a live iTwin.js digital-twin overlay.

Developer experience

• pytest runs on a bare checkout (just pip install numpy pytest).
• make test / lint / format / docs / build; pre-commit config; CI across Python 3.10–3.13.

Full notes in CHANGELOG.md.