# DeepBrain Metabase parity report

23 October 2026. Separate database 3, root collection 17, dedicated SQLite snapshot. The existing SORI connection, canonical source files, cards and dashboards are preserved.

## Implemented

- 20 base tables and three views; 27 detected foreign-key relationships.
- Original 24 operational SQL questions and four dashboards.
- 13 reusable grain-specific models, 27 reusable metric objects and 27 saved views referencing those shared metrics.
- DB01–DB32 implementation registry; DB10 uses exact nearest-rank percentile SQL Q07. DB29–DB32 require new instrumentation and have no fabricated values.
- Eight decision-stage subcollections, source/view collections and a nine-tab metric library dashboard (11).
- Three relationship dashboards (12–14) with 13 views: definition comparison, drivers, dependency boundaries, chronological task paths, missingness/recovery audit, mature user flow, cohort rates/counts, activity bands, weekly trends, held-out calibration/error and user evidence rows.
- 627 metric dimension instances, 182 raw fields and 23 tables described. Metadata changes apply only to database 3 and registered DeepBrain objects.
- Full frozen-snapshot library scope distinguished from date-filtered operational questions. Complete edge weeks and financial months; task costs preaggregated; signup users retained with zero activity.
- Standalone parameterized SQL, default result CSV/SVG tables, an executed editable relationship notebook, blank practice worksheet and separate model answers.

## Validation evidence

`metric_library_validation.json`: 27 metric formulas and 27 shared metric views checked cell by cell against SQLite. `relationship_validation.json`: all 48 definition/persona/workflow combinations for all 13 views, 624 actual Metabase queries, canonical SQLite parity and internal task/user flow conservation. `metabase_validation.json`: 168 operational query scenarios; all 24 mappings; original 248 card definitions/viz and SORI file hashes preserved. Browser rendering results are recorded separately in `metabase_parity_browser.json` after inspection.

The held-out default example deliberately does not support a success story: 45 held-out users are covered, the early-signal Brier error is 0.117468, versus 0.116467 for the earlier training base rate. Improvement is -0.001001. The early-value held-out cell has only one user. Do not claim predictive lift, activation causality or company performance.

## Equivalent scope and limits

The capability types match the relevant SORI patterns; object counts do not. Music/MIDI, learning-gain, GPU-capacity and token-cache tables were not invented for DeepBrain. Canonical DeepBrain definitions remain smaller because each supports a real interview decision. No production company data, official North Star, company target or internal architecture is asserted.

Metabase remains local on this Mac. Publishing the browser guide does not publish the Metabase database/server or make localhost links work on another device. The downloadable lab supplies the underlying files. Existing personal notes and source data were not overwritten.