# Browser and integration test record

Executed 2026-09-15 in Google Chrome on macOS against the local application. This record describes observed actions; it is not a claim of a cross-browser or accessibility audit.

## Simulator workflow

| Step | Observed result |
|---|---|
| Open app and Connections | App rendered; optional API-key and storage controls were present |
| Local sign-in | The loopback sign-in returned to the app with storage marked Connected |
| Open saved Brightbank study | All eight journeys loaded; product was Food & beverage, price $28.80, with ad/PDP/cart/checkout evidence attached |
| Inspect ad inputs | Instagram, headline, complete body copy, $9.50 CPM and price-not-visible setting matched the fixture |
| Upload creative through file picker | Source count changed from two to three and the new image preview appeared |
| Detach the duplicate upload | Source count returned to two |
| Request GPT review without a key | Connections opened for key entry; no paid request was made; continued without AI analysis |
| Save study | Study saved through the UI |
| Run 10,000,000 | Live unscaled progress appeared; eight journeys completed in 35.9 seconds |
| Inspect results | Baseline 207,282 clicks, 33,552 carts, 22,634 checkout starts and 4,843 purchases; final combined alternative 18,837 purchases |
| Save frozen result | UI confirmed the run was saved; local API readback preserved it |
| Compare browser and engine | Every scenario result and paired comparison exactly matched the published run; see `browser-replay.json` |
| Import example inputs | **Load Brightbank inputs** loaded eight editable journeys and the correct product/price |
| Import example results | **Open completed example** loaded the full frozen run without recomputing |
| Inspect an individual | Displayed consumer ID 56 replayed the recorded unanswered-product-questions outcome and its probability/draw trace |
| Example index | Rendered actual JSON results; switching to the final combined journey changed the funnel to 38,739 carts, 26,439 checkouts and 18,837 purchases |

The exported JSON and CSV files are generated by the reproducible script and were schema/accounting checked. The browser's separate download buttons were not independently validated for filesystem destination or browser download preferences.

## Storefront mockup workflow

- Complete PDP visually displayed the creative, $28.80 price, $6.95 shipping, sugar disclosure, ingredients, nutrition, storage, returns and support information.
- Add-to-cart opened the matching cart with one 12-pack.
- Entering BRIGHT10 showed an accepted-code message and carried the code into the checkout URL.
- Empty required contact fields prevented progression.
- Contact and shortened address steps accepted fictional test entries and advanced in order.
- The shipping step displayed discounted merchandise $25.92, shipping $6.95, tax $2.30 and total $35.17. The undiscounted path displayed $28.80 + $6.95 + $2.50 = $38.25.
- Express wallet selection disabled the card fields and displayed a simulated-payment notice.
- Review submission displayed **Simulated order complete** and confirmed that no payment was collected or data saved.
- The product-facts PDF was rendered and visually checked as a legible single page.

The mockup's five-step guest/wallet path was exercised. PayPal was verified as an available option but not completed as a separate path. Baseline card requirements and remaining scenario forms were covered by source generation and API HTML extraction, not eight separate manual checkout sessions. Mock form timing/error probabilities are input assumptions, not measured behavior of these forms.

## Automated checks

- TypeScript, ESLint, all 14 engine regression tests and the production build passed.
- The local HTTP integration suite passed authentication, origin, ownership, source-byte preservation, malformed-upload rejection, inert HTML, private-URL rejection, canonical evidence, analysis provenance, concurrent duplicate saves and missing-key rejection.
- The example script uploaded 27 actual assets, saved and reopened the study, evaluated eight journeys across 10M IDs, saved/read the frozen result, evaluated six prices across 10M IDs, and ran three explicitly labeled 200k-ID sensitivity cases.
- It checked funnel/segment/loss reconciliation, paired counts, zero purchases from structurally excluded profiles, checkout-stage isolation, deterministic partition replay and 57 individual outcome/economics records.

Source-presentation corrections after the first browser run aligned field counts, address fields, coupon carry-through and charge timing with the declared assumptions. The final assets were uploaded and the complete automated study rerun. Numerical outcomes remained identical to the browser run.

## Not established

Paid GPT creative/PDF interpretation, positive live URL import through Jina, real ad-platform delivery, real payment providers, mobile-device interaction, a complete accessibility audit and predictive validity against customer purchases were not tested. No provider key or real campaign was used. HTML uploads preserve text/form evidence; they do not run the submitted site inside ten million browsers.