EVIDENCE / NOVEMBER 5, 2026
See the work
behind the numbers.
Measured operations, defined datasets and independent numerical checks. Each result has a specific scope.
Latest Windows validation
GalaxyGrid 0.17.0-d2-test15. One million synthetic data rows, 32 logical processors, 32 requested workers, an effective parallel cap of 31 and adjustable RAM with a 192 MiB worker scheduling allowance.
| Workload | Result | Scope |
|---|---|---|
| 12 stock indicators | 10.21 seconds | One million source rows; end-to-end case time |
| CSV export | 3.25 seconds | One-million-row regression export; end to end |
| Installed features | 16 / 16 passed | Imports, charts, references, workflows and local queries |
| Workload correctness | 36 / 36 passed | The runner's documented workload cases |
| Native keyboard entry | 16 / 16 preserved | No background errors recorded |
Export was about eight times faster than the previous preferred-profile run (26.00 seconds). This comparison covers the same regression export workload. Hardware load and data characteristics affect results.
Download the summarized Windows results (JSON).
Full 6.7-million-row indicator audit
6,719,040 rows. Twelve indicator columns. 80,628,480 positions checked, including warm-up blanks. Zero mismatches outside tolerance.
Test15's actual indicator script calculated a synthetic series with variable prices and volume on the Linux build host. The separate saved-value checker recomputed each expected result from source High/Low/Close/Volume inputs using independent Python/NumPy calculations.
The acceptance rule is abs(actual-reference) ≤ 1e-9 + 1e-9 × abs(reference). The scan covers minute return, typical price, price × volume, SMA 20/50, EMA 20, true range, Wilder ATR 20, average volume 20, rolling VWAP 20 and Bollinger upper/lower bands.
Download the full-scale accuracy summary (JSON).
Native snapshot and read-only checking
The completed synthetic results were exported through GalaxyGrid’s native .ggrid exporter. The checker verified the archive’s embedded database checksum and scanned an isolated temporary copy without changing the source. This is a defined stock-indicator test; it does not certify every spreadsheet function.
What remains under evaluation
The Windows combined result is over_budget: nine of ten small-UI latency cases exceeded their original thresholds. These are performance misses, with no correctness errors reported in that run. An earlier intermittent keyboard issue has not been reproduced in the latest tests.
The stock audit does not certify every possible formula, exact-decimal mode, compact/lazy overlays, other indicator definitions or a customer's unsaved workbook. The updated checker accepts native archive versions 1–3 when the saved worksheet is materialized. Evaluate your own workbook and workload before relying on an output.
Independent checks you can run
The installation includes a combined regression runner and a separate saved-workbook checker under GG-Accuracy. Use the current checker update for newer native archives. Its report includes total positions, mismatch counts, maximum errors and bounded examples. The checker never uploads your workbook.