After v2 — landing page migration & model-switcher

The Flash Copy landing page at /Users/oscar/Desktop/dev/flash-copy-landing predates the ZapCopy rename and is stylistically conservative. We want to replace it with a Zapcopy-branded site that leverages cutting-edge interactive design while keeping the core narrative (what the app does, social proof, pricing, CTA).

E1 — Asset + content inventory (first step)

• Catalog every piece of existing copy and imagery under flash-copy-landing/ (index.html, privacy, terms, company, claw) — export into a single brief/content.md.
• Catalog usable visuals from /Users/oscar/Desktop/dev/FlashCopyAssets/ and ZapCopyAssets (demo videos, app screenshots, output examples, seed inputs).
• Pull the current pricing story from UsageTrackingService.swift (Free / Premium / Ultra) and the feature matrix from the in-app Subscription variants so landing-page claims line up with what the app actually does.

E2 — Design-brief prompt pipeline

• Assemble a prompt bundle that includes: the inventory above, our in-app visual language (neon-accent cyan, Outfit typography, dark UI, subtle bolt motifs), competitor landing pages we admire (for inspiration / anti-inspiration callouts), and explicit "must-have" sections.
• Feed the bundle into Claude's design capability and request 3 distinct directions (playful-editorial, command-center dense, hero-marquee cinematic) — each as interactive HTML mockups we can host locally and A/B.
• IP-safety call: evaluate whether Claude's design mode needs the same Stitch-style hard-gate we apply to Google Stitch prompts. Since Claude is our primary model we're more tolerant, but the final prompts should still avoid "make it look like [brand]" patterns. Save the resolved stance as a feedback memory.

E3 — Pick a direction, implement, deploy

• Side-by-side review the three mockups (screenshot → voting HTML pattern we already use for subscription variants).
• Scaffold the chosen direction as the new zapcopy-landing repo under /dev; reuse the Cloudflare Pages deploy config from flash-copy-landing/CLOUDFLARE_PAGES_SETUP.md.
• Point zapcopy.app DNS at the new Pages project; keep flashcopy.app alive with a 301 to the new domain for at least 6 months.

E4 — Decision points to raise during brainstorming

• Does the site get its own blog / release-notes surface, or do we keep change logs inside the app?
• How much interactive "try it" content? (A drop-in image → live OCR demo is technically feasible with a public demo Lambda, but adds ongoing cost.)
• Do we replicate /claw (the SteelHub subsidiary page) or quietly sunset it?

Today, if Gemini changes behavior overnight (silent model nudge, safety-tuning shift), we find out because users report bad scans — we have no fast path to swap providers or compare the new output to the old one. The v2 rollout proved that a per-run lambdaVariant field is the seed we need; F builds out the full infrastructure to make switching routine.

F1 — Provenance fields on every scan

Add the following to the existing OCR result envelope (photo + collection + video):

Persist on Photo / Collection / Video and forward to the dashboard ingest. Back-fill nulls for pre-F1 scans (label as provenance: "pre-F1" so charts can exclude them).

F2 — Lambda/model swap without a deploy

• Move the current static OCRLambdaEndpoints into a remote config fetched from a cheap Lambda at cold start (with on-disk cache + 1 hour TTL).
• Config shape: { flow, route: { primary, fallback, canary }, rolloutPct } so we can dial a canary at 5% → 25% → 100% without pushing a new iOS build.
• iOS emits telemetry per scan (success, latency, parse-rate) back to a lightweight stats Lambda so the dashboard can render canary deltas in real time.

F3 — A/B harness in zapcopy-qa

• Extend scripts/run_full_sweep.sh with a --provider arg that injects the relevant launch-arg override per run.
• New dashboard page /model-compare/[batchA]/[batchB]: paired seed-by-seed word-delta, latency, cost, and polish-notes concerns side-by-side.
• Add a "provider" facet to the existing flow pages so a user can filter /flow/video by Gemini vs Claude with one click.

F4 — Provider adapters

• Write a single OCRProvider Python interface and implement it for: Gemini 2.5 Flash (current), Claude Haiku 4.5, GPT-4o-mini.
• Each adapter takes {image, prompt, mode} → {text, rawEnvelope, tokens, elapsed}. The v2 Lambdas become thin callers of an adapter chosen by config.
• Parallel canary: at 1% of calls, run both primary and canary providers, record the diff for offline analysis, return the primary's result to the user.

F5 — Alerting on provider drift

• Nightly cron: rerun the regression sweep against the current primary provider, store stats in data/drift/<date>.json.
• If parse-rate, fence-leak rate, or mean-latency shifts by > 2σ from a 30-day rolling baseline, open an issue and post to the #zapcopy-ops Slack channel.
• On alert, the runbook is: flip canary config → 0% primary, 100% backup; regression sweep the new primary; file follow-up with the drifted vendor.