validation.md

Local validation — 2026-10-14

Live browser results

Trial	Result	Evidence summary
ChatGPT, manually initiated neutral category query, captured by CLI	Completed	8,764 characters, 18 visible source URLs; no literal BeatAPI mention
ChatGPT, fully automated fresh chat through the new collector	Completed	404 characters, 5 visible source URLs; no literal BeatAPI mention
Gemini, fully automated fresh chat, 120-second completion window	Timeout	No completed answer scored; brand mention remains null

The first and second questions differ; these are smoke tests, not a benchmark or visibility estimate. The manual capture is explicitly labeled operator-supplied prompt, not independently matched. The third result is a collector timeout, not proof Gemini failed to generate an answer: pending background rendering was observed in DOM inspection. No retry was submitted automatically.

Raw answers, account context and tab identifiers stay local and are excluded from Git. The Browser Lab UI was inspected in the real Chrome browser with a screenshot. No login data was exported.

Code verification

• Browser collector behavior tests: completed evidence, failed-run denominator, source URL filtering, no duplicate submission after ambiguous transport failure, local storage, no-brand scoring and verification handoff.
• BeatGEO provider, configuration and text-extraction tests: 101 passed before final documentation-only updates.
• Environment registry tests: 6 passed.
• Shared library TypeScript check passed.
• Whole workspace Biome check passed (818 source files at the first check).

Final collector suite: 9 passed. Final whole-workspace lint passed after the last implementation fixes. Local HTTP checks returned 200 for health/history and 403 for disallowed Host, foreign Origin and unauthenticated writes. CI only runs the dependency-free browser collector test suite; the BeatGEO checks above were local.

Not verified

Full BeatGEO dashboard + database + worker + browser collector end-to-end; unattended scheduling; all engines and regions; model pinning; BeatAPI analysis/billing integration. No database migrations or production deployment were performed.

TaskSpace default and desktop workbench follow-up

• Browser/evidence suite: 15 passed. Default collection chooses the supplied profile, scopes every page command, never falls back to ordinary tabs, and retains pages after completion or ambiguous submission.
• Workspace lint: 819 files passed. The installed pnpm wrapper had an execution-format error; checks used the installed pnpm 12.4.1 JavaScript entrypoint, with the workspace security settings unchanged.
• Isolated headless Chrome: visibility/citations/prompts/raw-answer switching, denominator handling, filters, selected profile submission and local draft saving passed without page errors. Fixture data was intercepted only inside the disposable test context, never written into the user's evidence store.
• Daily Chrome: the local workbench opened inside a named BeatBrowser TaskSpace, alongside the signed-in Profound reference page. The profile-discovery endpoint confirmed a connected 1.3.1 extension and TaskSpace mode. This was a UI/read check, not a fresh model submission.
• Public UI assets use Lucide (ISC) and Geist (OFL), with included licenses. Private source-page captures and the reference font remain untracked.
• Ego Lite is not a runtime/package dependency. Its local application was moved to Trash at the owner's request; profile data was retained.

Full product/worker/database integration, all Profound modules, native macOS focus guarantees, and a new live ChatGPT/Gemini collection with the default TaskSpace path remain unverified. No migration, npm release or production deployment was performed.

BeatGEO MVP identity and monitoring verification

Verified locally with Node.js 24.19.0 and pnpm 12.4.1:

• Workspace unit suites: 1,391 passing tests across config, CLI, lib, deployment and web.
• Browser collector/evidence: 16 passing Node tests; simulated transport, not fresh live-platform collection.
• Workspace TypeScript checks and Biome lint passed; the web production build passed with dependency bundling warnings.
• Headless workbench UI with synthetic observations: independent brand selection, 50% from 1/2 scored answers, timeout excluded, error plus partial text retained, elapsed time, JSON export and no JavaScript errors.
• Interactive product blueprint: 12 clickable steps, no JavaScript errors; README preview generated from that file.
• No production database migrations, account changes, paid provider calls or user browser windows were used for this verification.

PostgreSQL → worker → BeatBrowser end-to-end scheduled monitoring and live Gemini stability still need deployment acceptance. The README status table is the supported MVP boundary.

Monitoring loop and BeatAPI integration — 2026-10-15

This entry supersedes the earlier unverified database/worker boundary for the specific path below; it does not certify every engine or unattended operation.

• Created a disposable PostgreSQL 17 instance on loopback port 55439, with a dedicated beatgeo_mvp_test database. Applied all migrations, including collection attempts and optimization actions. No BeatAPI main-site or gateway database was changed.
• Real pg-boss → worker → local collector → signed-in ChatGPT: one fresh category question completed, yielding 1,457 answer characters and 10 visible source links. The worker stored the answer/citations/observation atomically and scheduled the next cycle 24 hours later. The answer did not literally mention BeatAPI. This is one smoke test, not a visibility estimate; the future cycle was queued, not waited for or observed executing.
• An earlier attempt stopped on PAGE_BUSY before submission. Added bounded, read-only composer readiness polling, then completed the trial above. Regression tests cover transient/persistent busy states and human pause without refresh, focus changes or duplicate submission.
• PostgreSQL integration: 40 tests passed, including five new loop tests for baseline → implementation → actual queue → matched retest, duplicate job handling, timeout, empty answer, cross-brand isolation, changed questions and transaction rollback when citations cannot be saved. Provider replies in these integration tests are synthetic.
• Workspace unit suites: 1,408 passed (config 119, CLI 35, lib 660, deployment 52, web 542). Browser collector/evidence: 19 passed. All workspace TypeScript checks and Biome lint passed. Web production build passed with dependency directive/chunk-size warnings.
• Local authenticated dashboard UI: saved a synthetic recommendation and baseline, recorded implementation, queued a retest, and displayed missing comparable coverage. Desktop and mobile inspected; no horizontal overflow or browser page errors. UI fixtures are explicitly labeled synthetic and kept out of Git.
• BeatAPI provider contract tests cover search → page reading → structured analysis, fixed credential destination, API-simulation separation, truncated/empty answers, and no automatic retries for 401/402/429/5xx. These tests mock HTTP responses. Anonymous public capability discovery found Web Search; it does not validate billed requests.

Still unverified: a real BeatAPI key/model/charge, new Gemini live collection, overnight cadence execution, multi-worker/device leasing, account/region/model pinning, complete citation side panels, production deployment and long-term reliability. SEO dedicated endpoints are not connected. Website technical audit and a complete content publication workflow remain roadmap features. See self-hosting and product/API boundaries.

Guided local setup acceptance — 2026-10-15

• A new pnpm local:setup created a dedicated PostgreSQL 17 container/volume on loopback port 55440, generated independent secrets, applied migrations, and prepared the pinned BeatBrowser 1.3.1 extension. --resume completed against the unchanged generated configuration. No main-site/gateway configuration or database was touched.
• From a fresh database, a headless local browser registered the first owner, opened Connections without a research key, selected the existing Chrome profile, created a brand and manually entered one question. The local organization admin can manage Connections; ordinary members and non-local modes cannot.
• That question ran through pg-boss → worker → bundled BeatBrowser → a signed-in ChatGPT TaskSpace. It completed in approximately 163 seconds with 2,663 characters and six visible citations, persisted to PostgreSQL, and queued its next cycle 24 hours later. This is one smoke test; the future cycle was not observed running.
• New connection contract/security tests passed, including unauthorized access, fixed credential destination/no redirects, rejected key/model updates, and sanitized status. PostgreSQL integration passed 41 tests, including encrypted key/model persistence and retaining the previous configuration after rejection. External BeatAPI responses in these tests were mocked; no billed model acceptance is claimed.
• Workspace unit suites passed 1,430 tests (config 119, CLI 35, lib 660, deployment 52, web 564). Browser tests passed 20; isolated local-runtime tests passed two. Workspace type checks and the web production build passed. The setup page was inspected at desktop and 390px width, with no mobile horizontal overflow.
• Source-based setup still requires Chrome's explicit extension installation when none is present. It does not silently install extensions, migrate Chrome profiles, or prove website login from bridge connectivity. There is no native one-click desktop installer, remote browser leasing or SQLite backend in this release.

Website-driven question bank — 2026-10-15

• Added 12 reusable question intents, eight starter categories and English/Chinese generation. Website research supplies the actual category, audience and use case; the template fallback requires a category choice and is explicitly not website-verified or measured search volume. Confirmed questions use the existing storage and schedule.
• Unit suites passed 1,445 tests, including website-profile generation, arbitrary non-AI businesses, Chinese prompts, unknown businesses and limits. Research/model replies in these tests are mocked; billed website research remains unverified.
• In a separate beatgeo_bank_test database on loopback 55439, a fresh owner created a brand without a research key. Headless UI verification generated 12 Chinese image-business questions (nine neutral, three branded), switched to English travel questions, verified an enabled tracking action and inspected 390px mobile layout without overflow or page errors. No live model submissions were triggered by this test.
• The creation flow exposed a retained parent-route cache that could produce a 404 for a just-created brand. Creation now opens a fresh document; a newly created test brand reached its question setup without refreshing manually.
• Workspace type checks, Biome and web production build passed. The user's local service on 3140 was restarted only after confirming no active collection or queue work. Test data remained in the isolated test database; the main-site/gateway database was not accessed.

First-use journey and prerequisites — 2026-10-15

• Added a read-only local:doctor check for Node 24, pnpm 12.4.1, Git, Docker engine and Compose. The source installer uses the same preflight before installing dependencies or writing database configuration. Checked it against the local environment; all five requirements passed. Missing tools, wrong versions and a stopped engine are covered by tests.
• Setup now orders research key/model → Chrome → website → ongoing monitoring. It distinguishes saved credentials, a connected Chrome profile and website login, includes extension-folder copying, and explains local data, API usage, restart and uptime requirements.
• The isolated authenticated setup page passed desktop and 390px mobile checks, including the clipboard value, ordered steps, masked key field, database/restart help, no page errors and no horizontal overflow. No real credentials were saved and no paid request or AI website submission was made.
• Web unit tests: 564 passed. Script tests: 16 passed. Web type checking, Biome and production build passed. This did not repeat a fresh installation or migrate a database; the existing local installer acceptance above still describes that path.

Branded MVP acceptance — 2026-10-15

• Unified the app, sign-in, setup, report and chart-export identity around the BeatAPI mark and BeatGEO name. Setup has four navigation anchors and consistent cards. Raw response data is collapsed by default; mention labels explicitly refer to tracked brands, and browser model versions are shown as unpinned account defaults.
• In the isolated beatgeo_bank_test database, the actual UI created BeatAPI, generated an English API-industry starter set, retained one question and started tracking. pg-boss → worker → BeatBrowser → signed-in ChatGPT completed in 219 seconds, storing 5,511 answer characters and 14 visible citation links. BeatAPI was not mentioned. The answer and citations were read back in the product; the next job was queued approximately 24 hours later. This is one live smoke test, not a visibility benchmark or an observed overnight cycle.
• Ten authenticated routes were inspected at 1440px and 390px: setup, overview, responses, citations, visibility, share of voice, opportunities, prompts, competitors and models. The brand asset loaded; no horizontal overflow or Playwright page errors were observed. Raw-data disclosure and the actual answer presentation were checked separately after collection completed. Vite reported ResizeObserver notifications during some viewport changes; this is not a claim of a completely silent development console.
• A separate empty beatgeo_release_ui_test database verified first-admin registration and access to Setup. An already initialized instance exposed an SSR error at /auth/register; the route now returns a 307 to sign-in while preserving returnTo. A fresh browser followed the redirect, signed in and arrived at Setup successfully.
• Workspace unit suites passed 1,445 tests; PostgreSQL integration passed 41, browser collector tests 20, script tests 16. Workspace type checks, Biome and the web production build passed. The web checks were repeated after the final UI/auth changes. Integration and research-provider replies remain synthetic; the one live browser answer above is separate evidence.
• The two test launchers were stopped after confirming no active collection; their databases were retained. The user's local app on 3140 and its database, shared BeatBrowser bridge, Chrome and main BeatAPI project were left running/unmodified. UI source changes are available in the existing local dev server after refresh.

Still unverified: a real BeatAPI key and billed website research / opportunity generation, a future unattended cycle, new Gemini live collection, a clean-machine installation on another computer, and full Profound feature parity. Current BeatGEO configuration contained no research key during this acceptance. The earlier installation acceptance remains the installation evidence; this pass exercised fresh registration on a new database, not a complete reinstall.

Team source release — 2026-10-15

• Pinned BeatBrowser to 8a295920d06cb83cd69e79e43eafa8e7b62b9efe, which includes task cleanup, clearer linked-page labels and the unframed logo. Frozen installation downloaded the GitHub dependency successfully without changing supply-chain policies.
• Prepared the downloaded extension in a disposable directory and verified the manifest, task backend, popup and stylesheet match the dependency. The resulting popup includes cleanup controls. The user's loaded extension, browser sessions and database were not reconfigured.
• Repeated Biome (841 files), workspace types, 1,445 unit tests, 20 browser collector tests and 16 script tests successfully. Prior real signed-in ChatGPT evidence remains documented above; this dependency update does not add claims about paid research, overnight execution or other model websites. PostgreSQL integration also runs in the PR's isolated CI database.
• README now explains that a BeatGEO fork retains a pinned dependency, not a second repository fork. Each team member installs their own local instance and explicitly loads the Chrome extension.

README and interface languages — 2026-10-15

• Added English-primary README plus Chinese, Japanese, German and Korean versions, with an independent open-source Profound-alternative tagline and current MVP boundaries. All seven READMEs across BeatGEO and BeatBrowser passed local-link and heading-anchor checks. README Refiner's only remaining warning misreads the documented pnpm version as a command.
• Web lint, type checking, 568 unit tests and the production build passed. Two committed Playwright language regressions passed against both the development server and the built server: switching preserves entered form values, reload remembers the choice, and another tab receives the change.
• Headless Chromium with a disposable profile exercised sign-in, setup and nine brand routes against the existing isolated beatgeo_bank_test database. Checked English/Chinese switching, saved language after refresh, an unsaved setup input surviving a switch, desktop and 390px setup layout, and original answer text remaining visible. No uncaught page errors were observed. The database's previously captured answer was reused; this was not a new AI collection.
• The first-click regression exposed an interactive control before React hydration finished. The language control now remains disabled until the client is ready. No database migration or main BeatAPI change was needed.
• No model or paid search request was made. The test instance had no research key; opening Opportunities displayed its research-configuration failure. Generated research, CLI messages and unknown provider errors are outside interface translation; see language scope.