# DeepBrain interview package: QA report

Initial local verification on 23 October 2026. The guide was subsequently authorized for public publication at **deepbrain.define-labs.com**; see Release_Report.md and Metabase_Parity_Report.md for that expanded verification. No AI Studios account creation, purchase, paid model/API request, application or external message was performed.

## Delivered artifact scope

- 41 static browser routes: 16 authored chapters, 24 exercises, root entry. Shared template/components and one authored content set also produce the complete Markdown guide.
- One-page briefing; full translated PO role, required/preferred qualifications and employee benefits; product/market/research chapters; evidence map from the two supplied CVs; 12 answer-first scripts with short/expanded versions and stopping points; four concentration drills, rubric and study sessions.
- Proposed 23-section localization PRD, strategy stack and editable SVG metric relationship map. No internal company architecture, plan or approval is claimed.
- 32 canonical metric definitions: 28 connected to synthetic analytical views/queries; four additional payment/conversion/version/privacy definitions explicitly uninstrumented. All actual business baselines and targets remain unknown.
- 20 SQLite base tables and three views; 180 fictional accounts, 312 users, 1,300 tasks, 1,241 attempts. No generated demographic attributes. Public/observed/synthetic/proposed provenance stays separate.
- 24 standalone SQL queries, result CSVs, 24 SVG chart/table views, 18 current plot PNGs, editable notebook with 25 executed code cells, schemas, generator/analysis/validation scripts and separate model answers. Personal answer and score fields are blank.
- Separate Metabase database 3, collection 17, 24 saved questions and dashboards 7–10, within the existing local server. It is a separate project/dataset, not a separate server process.

## Automated checks actually run

| Check | Evidence | Result |
|---|---|---|
| SQLite integrity, foreign keys, one task per joined row, full cost parity | `lab_validation.json` | Passed |
| Review NULL/acceptance/critical-error boundaries, milestone chronology | Same | Passed |
| Independent activation/return recomputation, start/end and maturity fixtures, zero-activity users | Same | Passed |
| MRR/entitlement reconciliation and timestamp-ordered journey flow conservation | Same | Passed |
| 24 queries × seven filter/time scenarios, incl mismatched persona/workflow and future empty windows | Same | 168 scenario queries passed; 191 total checks |
| Seed reproduction in a fresh temporary folder; every base-table cell compared | `seed_reproduction.json` | 20 tables passed; canonical database not regenerated |
| Actual Metabase saved-question queries × same scenarios, every returned cell versus canonical SQLite | `metabase_validation.json` | 168 filtered query checks passed |
| Intended parameter mappings | Same | All 24 cards passed; static/offline/planning and financial scopes declared |
| Existing saved-card formula/visualization/name/collection preservation | Same | 248 existing cards passed after canonicalizing Metabase's generated query UUIDs |
| Existing SORI source DB, Metabase snapshot and website build-manifest hashes | Same + `infrastructure_before.json` | Unchanged |
| Existing dashboard name/collection identity | Same | All six unchanged; no existing dashboard mutation was issued |
| Notebook execution against read-only SQLite | `analysis_manifest.json` | 25 cells executed; saved outputs are synthetic |
| HTML routes, local links/anchors/assets/downloads, metric fields and blank exercises | `site_validation.json` | 41 routes; 1,116 local link checks; zero failures in final run |
| Download archive integrity and excluded private inputs | Same | Archive test passed; no raw CVs, session/login files or private Metabase state |
| Portable bundle rebuilt in a fresh extracted folder | `portable_bundle_validation.json` | 41 routes reproduced without Metabase writes |

The source before/after card comparison needed UUID normalization because Metabase emits fresh internal query-reference UUIDs on reads. Semantic structure, formulas and visualization settings were compared with consistent UUID mapping, preserving references rather than ignoring the entire query.

Commands run from this project: `scripts/inventory_reference.py`, `fetch_sources.py`, `build_lab.py` (new database only), `build_queries.py`, `build_metrics.py`, `build_answers.py`, `build_exercises.py`, `build_analysis.py`, `build_glossary.py`, `build_ledger.py`, `build_relationship_map.py`, `build_metabase.py`, `validate_lab.py`, `validate_metabase.py`, `check_reproduction.py`, `build_website.py`, `validate_site.py` and `validate_portable_bundle.py`. Core SQL uses standard-library Python; chart/notebook/site builds used the existing analysis environment. No SORI generator was executed.

## Browser checks actually performed

Mac in-app browser. Detailed record: `browser_validation.json`; dashboard render counts: `metabase-browser.json`.

- Search returned relevant chapters and exercise E12 for entitlement. Page-specific right-pane explanations changed correctly on navigation.
- Notes and completion survived reload; QA-only notes/completion were cleared afterward. Timer advanced and stopped.
- Dictionary lookup, keyboard Enter activation and Escape dismissal worked. Unknown lookup behavior is explicit and implemented; it does not generate answers.
- SQL copy produced the executable query in the clipboard. File/ZIP targets passed HTTP/filesystem/archive checks.
- Single-paragraph selection produced a highlight; note and blue category persisted. Keyboard editing worked. JSON export contained the note/category; invalid backup was rejected; valid import restored quote, note and category after reload. QA highlight was removed afterward.
- Device voice options populated. Read, pause, resume and stop controls changed state correctly. Navigation reset controls/context; code cancels prior speech on pagehide. No paid audio service is used.
- All 24 Metabase questions rendered as charts/tables across four dashboards without visible chart errors. The actual Persona dropdown → training → Update filter interaction loaded the intended dashboard cards. Offline/period cards retained their declared independent scope.
- Guide and exercise layouts were visually inspected in 390×844 CSS-pixel iframes. Controls wrapped and content remained readable. This is a narrow viewport test, not a physical phone/touch-device test.
- Print CSS was inspected through a layout simulation: navigation, audio and worksheet controls hidden, content visible. Native print-preview inspection was attempted but the computer-use layer blocked access to the Codex app. Native pagination and printed output remain unverified. No PDF is delivered.
- Plot contact sheet inspected. A misleading experiment comparison title was replaced with a mixed-assignment audit/table. Candidate plot labels distinguish baseline/candidate. Six evidence views are tables rather than invented graphics.

## Corrections made during QA

1. Initial generator placeholder-count errors were corrected before producing the usable dataset; failed draft files remain only in local QA, excluded from downloads.
2. Four initial billing movement rows had subcent rounding residuals. A narrow repair assigned the residual to the corresponding movement while retaining closing MRR/carry-forward and cash. The corrected generator reproduces every table. See `billing_rounding_repair.json`.
3. A whole-paragraph quote also appeared inside a model-answer block. Empty prefix/suffix had been treated as absent context, making restoration ambiguous. Exact empty boundaries now participate in matching; save/reload/export/import were retested.
4. Q01 now includes only whole UTC weeks fully inside the selected use-date range and before cutoff, preventing a partial final selected week from looking like decline.
5. Q13 now audits mixed variant assignment directly. It does not visualize a treatment-effect comparison from an invalid account experiment.
6. Download rebuilding clears only its generated directory so removed/changed exports cannot remain stale. No browser progress or canonical source data is deleted.

## Source and evidence limits

The full Korean target posting was read in the browser; its translated qualifications and benefits are distinct. Raw research snapshots/access ledger remain local. Twenty primary/publisher source entries were reviewed with explicit limits. Research methods/results/discussion and the architecture image were inspected; complete marketing/demo playback was unavailable. Review platforms were readable through the web tool while direct requests returned 403. Name-confused/incentivized reviews are not treated as population evidence.

The current SORI manifest and catalogs were reconciled against its older handoff. The reference inventory contains 188 fully parsed files with hashes/structure and notebook source cells; large catalogs were inspected structurally and selected relevant text reviewed. Historical material was not reproduced wholesale. SORI music data, targets and distributions were not reused.

The supplied Leadership and Senior Product PDFs are the application baseline, with inspection/file dates and hashes retained in `sources/cv_baseline.json`. They contain a KidsLoop date discrepancy. No CV was edited. The guide does not supply missing historical metric formulas, original experiment records or unrecorded anecdotes.

## Still unverified or blocked

- Authenticated AI Studios import, editing, rendering, pronunciation, export/use, failed-job recovery, billing behavior and permission/deletion enforcement. App access stopped at sign-in; no account was created.
- Current internal strategy, architecture, actual customer/performance numbers, reporting leader/interviewer, interview date/format/language and employee-benefit contract applicability.
- Audible quality/pronunciation of device speech, PJ's actual 20–30/60–90-second delivery, every screen-reader path, physical mobile devices and native print pagination. Device speech controls working does not establish a spoken-answer timing result.
- `deepbrain.define-labs.com` hosting, DNS, SSL and public access. Publication is deliberately pending local review.

These limits do not turn hypothetical expected impact, synthetic output or artifact completion into a historical accomplishment or production release.