apps/trainer (parking-trainer): inspect / train / evaluate / publish. Reads the wash
collector's SQLite + crops read-only off its volume; time split (validation = newest
slice); thin classes dropped; damped class weights; `features` mode (frozen ImageNet
backbone, on-disk feature cache, seconds to retrain) and `finetune` mode (light
augmentation). CPU-only torch from PyTorch's wheel index. ONNX export checked against
the torch model; NO model file below the validation floor (exit 3, report still written);
exit 2 = not enough labels. `evaluate` scores a shipped model on labels reviewed after
training + the unlabelled pile; `publish` PUTs a version folder to a Gitea generic package.
Light core deps; the `train` extra is heavy — CI syncs without it, torch tests skip.
apps/vision: BodyTypeClassifier (bodytype.onnx + sidecar = the preprocessing contract:
crop margin, input size, RGB 0-255, normalisation inside the graph) and
RefinedVehicleDetector over YOLOX — refines only `car` or a class the classifier trained
on, min-confidence, `detector_class` on the result; path set but no file = phase B off
without an error; a broken file is a health detail. models/bodytype.version (tracked,
empty) pins the published version the Dockerfile fetches at build (BuildKit secret;
a pin that cannot be fetched fails the build). Verified: a trainer model gives identical
probabilities inside the vision service; both images built and smoke-tested.
Delivery: parking-trainer image in build-images.yml, the `trainer` compose profile on the
collector stack (CPU, read-only data, TRAINER_OUT), commented TRAINER_OUT/PUBLISH_TOKEN in
the wash-collector stack, .dockerignore for both Python contexts, trainer deps synced in CI.
Wiki: bodytype-classifier-training rewritten as built (+ one fleet model not per site,
secrets/access, where the crops live), opencv-anpr-service §Phase B, vision-review-outbox,
vision-service-packaging, fleet-deployment-komodo, index, log.
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
The wash stream is small; the entry camera photographs every car in exactly the view the
classifier is trained on. The booth can now queue entry vehicle reads as pure training
material — crop + the camera's class, no order, no operator, no category.
- Core announces every vehicle read (deviceEvents.emitVehicleRead from snapshot.ts); the
Car Wash module listens, samples entry reads in-process (sampleEntry: exactly one in N)
and queues them (enqueueEntry). CARWASH_REVIEW_ENTRY_SAMPLE=N; 1 = every entry (storage
and bandwidth are not the limit — user); 0/unset = off. Forwarded by compose.
- Packages carry kind: "wash" | "entry". Collector: kind column, entry meta validated
without the operator fields, review screen shows an entry sample as such, export has a
kind column, operator agreement computed from wash items only. Setup line shows
"1 in N entries sampled"; status carries entrySample.
- komodo: park-2's four review lines enabled (collector URL by Netbird DNS name, booth-2,
the shared per-booth secret, every entry sampled) — the collector is up on the overlay.
- Tests on both sides. Wiki: vision-review-outbox (entry stream + the internet-feed
assessment), log.
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
The far end of the Car Wash review outbox (wiki/concepts/vision-review-outbox.md): a small
Fastify + SQLite service in the monorepo (shares the payload contract and the class
vocabulary via @parking/shared), delivered to art-docker-station by its own stack so
nothing booth-side lands there and nothing of it on a booth.
- POST /ingest: bearer token per booth (constant-time), X-Booth-Id must match, multipart
meta + JPEG (magic checked, 2 MB cap), meta validated against the contract, idempotent on
the item id; crop stored at crops/<booth>/<item>.jpg on the volume + one items row.
- /review + /api/*: the reviewer's screen served by the process (Basic auth, one login):
one pending crop at a time, operator's pick and camera's pick beside it, one button/key
per vocabulary class + unusable + skip; stats per booth and per hashed operator
(agree / disagree / unusable — disagree = the reviewer's class is outside the operator's
category).
- GET /export/labels.csv: reviewed usable rows for training; formula-leading cells are
neutralised (booth-supplied names). Crops stay on the volume for the trainer on the host.
- Booth payload now carries operatorCategory.classes so the comparison needs no site setup.
- Delivery: apps/collector/Dockerfile (monorepo context), docker-compose.collector.yml
(bind to the overlay IP; commented `trainer` profile seam for the GPU), a third build
step in build-images.yml, a `wash-collector` stack in komodo/resources.toml with one
secret per booth referenced from both the collector's token list and the booth's own
stack (park-2 lines templated, commented, DNS name for the URL).
- Tests: app.test.ts (ingest ok/dup/refusals, review + stats + export, config). Image
built and smoke-tested locally (health, ingest, duplicate, auth, verdict, export).
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
The operator's category choice is a hypothesis, not truth (user, 2026-09-06): each wash
order with a vehicle read queues a package for a trusted reviewer over the private overlay
(Netbird); the verdict becomes the phase-B training label and the per-operator error rate.
wiki/concepts/vision-review-outbox.md.
- Boxes: the vision service returns the vehicle bbox; snapshot.ts stores the vehicle and
plate boxes on the read as FRACTIONS of the analysed frame (the stored snapshot is a
downscaled copy); vehicleForIdentity() returns them.
- carwash_review_outbox (migration 0031) + review-outbox.ts: crop = detector box + 8 %
margin, ≤ 640 px, plate blurred in place from the plate box; payload carries a
pseudonymous booth id and a keyed operator hash — no site name, no plate, no OSD, no
bystanders; multipart POST with a per-booth bearer; 2xx → sent (image dropped);
400/404/413/415/422 → abandoned; anything else → backoff 1 min·2^n capped 6 h; voided
orders and items older than 14 days abandoned unsent. Nothing queued while unconfigured.
- Enqueue is fire-and-forget off the intake path in createOrder; the loop runs every
CARWASH_REVIEW_INTERVAL_SEC (60) and stops on close.
- GET /api/carwash/review/status (site:read) + a "Remote review" line in Setup → Car wash.
- Env CARWASH_REVIEW_URL / _TOKEN / _BOOTH_ID (all three or off) documented in
.env.example and forwarded by compose.
- Tests: review-outbox.test.ts (crop + blur on a synthetic frame, config/pseudonyms,
queue/drain/backoff/abandon, through the app). Wiki: new concept page, index,
venue-modules As built, log. The collector is not built.
Claude-Session: https://claude.ai/code/session_01FWncR69HgGPuei1dLrW3cU
julian