Docs · Data
The data under the engine, and its limits
The engine reads a generated inventory of panels, each counted in its own unit. The development scores come from a separate benchmark. Both carry their commit and their date.
Data
Measured panels, each in its own unit
The rows below come from a generated inventory of the files the engine and the benchmark reference. A record is counted in the unit of its own panel: a series, a test variant, a weekly store × product observation, a dated filing, a company-year. Units differ between panels and stay separate.
| Records | Real units | Status |
|---|---|---|
| Monthly time series | 48,000 series | training |
| Cold-start library | 29,479 series | training |
| A/B tests | 212,600 test variants | development |
| Retail price panels | 22,717,635 weekly store × product observations | development |
| Retail transactions | 524,950 product-store-week transactions | development |
| Aggregated trial arms | 241 aggregated arm rows | development |
| Business failures by section and month | 4,187 section-month observations | development |
| Annual filings | 58,456 dated annual filings | development |
| Business failure records, panel A | 1,499 company-years | development |
| Business failure records, panel B | 1,432,589 company-years | development |
| Historical confirmation lot | Sealed | sealed confirmation; historical and reused |
The inventory spans 28 files and 1,796,943,023 bytes, the closed historical lot included. Opened files carry their byte size and a SHA-256 hash. The historical confirmation directory is counted by filename and size, with its files left closed.
Dated records keep their dates, and a series point dated after the decision is refused.
Method
How a claim is tested against a null
Each benchmark family has its own null baseline. G measures the improvement over the best null: net advantage for choices, CRPS gain for trajectories, Brier gain for distress. The global G is the unweighted mean of bounded G values across the real global families.
A development candidate needs a positive paired gain, no degradation beyond paired noise, a failing placebo, a recovered positive control and a green suite. The release decision requires one declared reading of a new sealed lot against seven release criteria.
Scores
Development scores, family by family
The global G on the development lot is 0.1517, against a release target of 0.20 on a new sealed lot. The register contains 22 development runs. Engine commit ad9184c23038, register date 2026-09-29T00:56:49.
| Family | Development G |
|---|---|
| A/B tests | 0.0568 |
| Price response, retail panels | 0.1281 |
| Series level | 0.2656 |
| Series cold start | 0.1508 |
| Business failures by section and month | 0.1400 |
| Distress, annual filings | 0.0941 |
| Synthetic false positives | −0.0278 |
| Margin, retail panels | 0.4555 |
| Distress, company panel A | 0.0250 |
| Distress, company panel B | 0.0443 |
| Zero-data case | 0.0304 |
| Price response, retail transactions | 0.3528 |
| Aggregated trial arms | 0.0769 |
These scores measure progress from one commit to the next. A general accuracy claim waits for the fresh sealed lot.
Limits
What the data volume leaves open
Data volume says nothing of predictive power on its own. Some price changes are observational, some test data carry aggregate outcomes, and some series share origins. Development data have influenced the engine. A trace shows that a calculation repeats; predictive accuracy needs its own test.