Reports
Hand-written analyses that combine run data with findings that sit outside the automated pipeline (push-notification delivery, leak events, prompt-quality judgments). Each report points at the runs it references so you can open them and verify.
Stage 1 + 1b — Flash vs Pro × normal vs strict on photo OCR
infoFirst production run of the F1–F6 routing infrastructure, extended with the strict-mode dimension after a bridge bug was found and fixed. 3 photo fixtures × 2 models × 2 prompts × 3 iters + collection pair = 42 runs. Verdict: keep Flash + normal as default. Strict didn't kill the luggage hallucination on either model — but Pro+strict produced byte-identical output every iteration, a determinism win worth offering opt-in.
Workstream F1 + F2 — Lambdas now self-report and accept runtime overrides
infoHigh-level ship update with animated flow charts. Foundation step of Workstream F: every OCR Lambda now stamps provenance (provider/modelId/promptId/promptSha256/lambdaVersion=*-rev2) on every response and accepts modelId+promptId on the request body. 75/75 tests passing, 4 Lambdas updated, zero behavior change yet — the plumbing F3-F7 ride on is in place.
Workstream F — hot-swappable prompts + Gemini model variants (plan)
infoExecutive view of the next cycle. Why Lambda-baked prompts/models cost us; the 7 sub-steps (F1 self-ID, F2 prompt YAML, F3 S3 config, F4 iOS picker, F5 sweep harness, F6 dashboard provenance, F7 Flash-vs-Pro sweep) with complexity, risk, sequencing, and the verification gate per step.
Photo + collection strict mode — paired sweep (G5)
infoMirror of the video strict-mode rollout, landed on photo + collection OCR. Parallel strict-capable Lambda, iOS Settings toggle, dashboard notes plumbing. 3 seeds × 2 modes — AAAA0011 code screenshot shows strict preserving typos and formatting normal silently corrects.
v2 OCR rollout — everything that shipped
infoEnd-to-end write-up of the v2 migration. Covers all three flows, the dashboard observability layer, and a deep dive on strict polish mode (what it does, which products, default status, and how to enable it).
Unit economics — interactive cost calculator
infoSlide a user through the Free / Premium / Ultra tiers and a photo / collection / video usage mix to see the per-user margin. Breaks out Gemini + Lambda + Rekognition unit costs against Apple-adjusted revenue.
Upcoming — landing page + model-switcher plan
infoWork queued after the v2 rollout: (1) migrate flashcopy.app → zapcopy.app with a Claude-design brief, (2) add a Lambda/model switcher + per-scan provenance records for easy A/B and provider-change audits.
Gemini polish `additional_notes` audit
review5 unique v1 polish-notes blobs analyzed. Gemini silently rewrites OCR content it thinks is wrong (nickBase→pickBase, getWebView→getWebview), fabricates try/catch structure, and non-deterministically picks regex variants. 4 prompt-iteration recommendations.
v2 Video Pipeline — Decision Hub
reviewFour decisions: ship v2 as default, BBBB0003 −32.4% word delta verdict (visible char diff), push validation A+B+C tracks, Workstream C go/no-go. Each with evidence and a recommendation.
v1 vs v2 OCR Lambda — paired video sweep
infoBatches 2026-04-21T17-19Z (v1) and 2026-04-22T03-23Z (v2). 10 runs × 2 seeds × v1/v2. Zero v2 fence leaks, 15% faster, rawResponse captured server-side.
v1 vs v2 OCR Lambda — paired photo sweep
infoBatches 2026-04-21T14-32Z (v1) and 2026-04-21T14-35Z (v2). 24 runs total across 4 photos × 3 iters × 2 variants, including 2 fence-prone adversarial fixtures. v2 structural defense holds at 100%.
Full-coverage sweep analysis
action neededBatch 2026-04-20T16-21Z (30 runs). Real JSON leak observed on 1 of 10 video runs; push notifications likely not delivering; photo + collection outputs extremely deterministic.