Every morning, the photos I already took on my shift become a report I didn't have to write.

Which photos even belong in the report is itself an engineering decision, and I didn't hand it to the model. A code-signed PhotoKit binary, photos-query.swift, does the selection: it queries an overnight time window from config, then applies four independent post-filters — within the geofence (centered on the work site), from the rear camera (detected by reading the EXIF LensModel and excluding anything marked "front"), not a screenshot, and carrying real GPS and EXIF. It logs its own filter stats so a bad night is debuggable.

The systems detail that makes this real: Photos and iCloud access only works from a code-signed binary spawned by launchd — the TCC permission is identity-bound and survives recompile only when the same identity re-signs. The binary exits with a distinct code on a permission denial, and the Python orchestrator retries with backoff. Getting an unattended 06:30 job to reliably touch the photo library is a systems problem, not a scripting one.

Each photo is resized and converted to JPEG (≤1024 px, quality 70) — HEIC isn't a valid vision input, and the resize caps token cost. The vision call returns structured JSON: a category (gate, vehicle, damage, document, other), a description, and for vehicles the make, model, color, plate, plate state, a plate-reading confidence, and a fractional bounding box for the plate. The category rules are hard-won — a "gate" means only the perimeter gate, never an interior door; that distinction only got written after the model got it wrong in the field.

The signature move is the two-pass plate enhancement, and it's the most transferable idea in the system. A re-read fires on either of two triggers. The first is obvious: low confidence (< 0.80). The second is the interesting one: a plate that contains a known-confusable character (O 0 Q D I 1 B 8 S 5 Z 2 G 6 9) gets re-read even at high confidence (< 0.95). The insight, in the code's own words: a model can be confidently wrong on a character whose glyph is ambiguous. The field failure that motivated it — a plate read as OJG659 that should have been QJG659, the Q's tail lost to streetlamp glare while the model's prior leaned toward O.

On a re-read the pipeline crops the plate region (model bounding box, falling back to a heuristic bottom-center crop), grayscales it, boosts contrast 1.8×, sharpens 2×, and upscales small crops, then re-reads with a prompt that threads the first-pass state hint while explicitly telling the model not to anchor on the first read. Acceptance is guarded: a lower-confidence re-read is accepted only if it changed confusable-character positions and nothing else — a wild swing or a length change is rejected. The general principle, which outlives plates: when a model does OCR in a domain with known-confusable tokens, the presence of a confusable token is itself a signal the confidence may be mis-calibrated. Force a second look with domain context baked in.

At the start of every run, jetson_client.probe() calls the Jetson's /ready endpoint with a 5-second timeout. If it answers, vehicle photos go to /v1/plates for the local detector's bounding box and OCR; if it times out, the run keeps the Claude read, logs the fallback, and continues. The breaker is that timeout plus the route decision — and it fires every morning in production: the logs show plate reads routing to local inference daily, with Claude as the floor underneath.

The merge logic is deliberately cautious. The local detector's bounding box is always preferred — it's more accurate than the model's estimated box — but its OCR text is trusted only if it passes three gates: at least five characters, at least 0.70 confidence, and not meaningfully shorter than the existing read. Fail any gate and it's "bounding-box only": keep the model's text, but feed the better crop into the enhancement pass from Step B. That guard came from a real bug — a low-confidence two-character local read overwriting a valid seven-character one — and every rejection is logged so a regression is visible. Local when available, cloud as the floor, with the seams instrumented.

The model reads photos one at a time; the report is about the night as a whole, and that gap is where structural errors live. Four deterministic passes close it. Vehicle dedup collapses the same car shot twice from different angles into one, using multi-signal scoring — fuzzy plate distance, make/model/color agreement, time proximity — against a grouping threshold. A critic_review pass checks the assembled manifest for four defect classes — malformed plates, vehicle overcounts, category errors, implausible plate states — and its return is constrained to corrections the orchestrator applies programmatically, so a review note can't quietly become report prose. Cross-day repeat-offender matching runs the same fuzzy multi-signal comparison against a 90-day history, so a returning car is caught even when OCR got its plate slightly wrong on a different night. And a silent-loss detector reconciles the exported photo count against the count that made it into the report's sections, and warns on a mismatch — born from a real incident where surplus photos in a single slot dropped silently.