Platform · Performance
Measured, not promised
Every performance claim on this site is an assertion inside our own test suite, with a millisecond ceiling that fails the build when it is crossed. Here is each number, what it measures, where it was measured, and the caveat that travels with it.
The methodology
Budgets, not adjectives
Three rules make a number quotable here. Every row in the tables below obeys all three.
Gates in every build
A performance claim is a test with a ceiling. 54 of 439 backend test files assert a timing bound; a crossed ceiling fails the build. The gates run on one dedicated machine so timings are consistent, and a budget is never moved to accommodate a slow machine.
One machine, two setups
Every number was measured on one machine in our lab, in one of two setups. The single-node setup runs the embedded columnar fact store in-process. The scale-out setup runs the scale-out column store with distributed job runners, the cache and a control database with split roles. Neither is a cloud region or a cluster.
Our own test data
Synthetic worlds and demo tenants with fictional brands, never a customer workload, because there is none to quote. Nothing on this page is a forecast-accuracy claim; the accuracy machinery is a capability we surface, not a result we have delivered.
The numbers
Every number we quote, and what it actually measures
Read the caveat column as carefully as the first. A number without its framing is a number we would not put on this page.
| Number | What it measures | Where it was measured | Caveat |
|---|---|---|---|
| 33 / 34ms | a grid pivot page (p50 / p95) over a one-million-row world, brand × store, mixed SUM and AVG, a 100-week window, served through the real HTTP grid door; a 100-row result page | single-node setup · an automated build gate | warm (3 warm-ups, 15 samples); in-process test client, no network hop; a synthetic test dataset. The tight budget was ≤ 100 ms. This sits 3× inside it. |
| 12.5ms | rolling one million fact rows up to one thousand groups | single-node setup · an automated build gate | the test asserts under 1 s: 12.5 ms is the recorded value, not the assertion; single user, no concurrency. The scale-out setup later measured 18.5 ms. |
| 1.02s | one million rows written in bulk through the full storage-contract path | single-node setup · an automated build gate | the same synthetic test dataset. |
| 1,003,750,000rows | a retail day-level fact table on the scale-out adapter: 2,500 items × 2,500 stores × 730 days, 22 % listed-pair sparsity, 438.6 MiB on disk | scale-out setup · our own benchmark | our own test data; not a production workload. |
| 2.39Mrows/s | write throughput at billion-row scale, through the storage contract | scale-out setup · the same benchmark | the same benchmark, the same caveat. |
| 127.8ms | a workbench-shaped pivot page, department × region, a one-month window, over the billion-row table | scale-out setup · the same benchmark | cold; 13.8× faster than the path it replaced. It missed our own 100 ms tight target by 28 %, and we recorded that. |
| 65.6ms | a 500-store × month drill on the billion-row table | scale-out setup · the same benchmark | inside the tight budget. |
| 124.8ms | an interactive pivot read immediately after a fresh 100,000-row write at billion-row scale | scale-out setup · the same benchmark | the churn case; the earlier ~6 s churn tail is gone. |
| Number | What it measures | Where it was measured | Caveat |
|---|---|---|---|
| 70ms | a single cell edit reaching its downstream measures: edit → buffer → flush → event bus → dependency graph → recompute, the full stack (re-measured at 74 ms) | single-node setup · an automated build gate | one edit, one tenant; the gate asserts under 500 ms. Not a concurrency number, and not the same thing as the live rule cascade in the next row. |
| ~505ms | a live rule cascade, steady: the full downstream rule chain firing after an edit on a live demo tenant (from 2.3 s; the era began at 24.1 s) | a live demo tenant | a different measurement from the 70 ms gate: a whole rule chain, not one edit’s propagation. We name which is which and never blend them. |
| 2,875,321 → 6rows | one cell edit’s downstream cascade, before and after bounding: the rows it re-evaluated, and the time it took: 24.1 s → ~2.2 s | a live demo tenant | the ~2 s remainder is a named open seam, not a finished number. |
| < 500ms | the platform’s edit-visibility budget: an edit visible to its editor | a budget every gate holds | a budget we hold, not an average we report. What was proven under load is the next row. |
| 10 / 193 to 212ms | editor latency (p50 / p95) under sustained load on the scale-out setup: 25 edits/s + 10 reads/s for 60 s, run twice, zero failures | scale-out setup · an automated build run | inside our 500 ms limit. This is not “50 concurrent editors”. That target has not been run. The single-node setup under the same load recorded p95 526 ms, above our own limit; we record it as a miss, not a pass. |
| Number | What it measures | Where it was measured | Caveat |
|---|---|---|---|
| 48,034 → 105ms | an all-history 24-month rollup over the full billion rows: before and after a declared rollup layer; 456× faster served, identical totals | scale-out setup · the same benchmark | quoted only as the pair: that shape needs a declared rollup layer, built once asynchronously in 48.7 s. On its own the 48 s is the shape that did not fit the interactive path, and we say so. |
| identicaltotals | the storage swap-proof: both storage engines return the same totals at both scales; one configuration value is the whole switch, zero caller changes | both setups · a permanent automated build gate | not “engine A is faster than engine B”: the embedded engine wins small aggregates and the scale-out engine’s curve is flatter. The claim is the right engine per tier, the same answers. |
| 208.5Mcells | two live tenants re-materialised from one storage engine to the other, at about 2M cells per second, parity exact | demo tenants | demo tenants, our own data. |
| 74ms | the scenario engine’s gate: the branch chain, the root rule, the cascade and the save walls, together | single-node setup · an automated build gate | single-node setup. |
| Number | What it measures | Where it was measured | Caveat |
|---|---|---|---|
| 143 / 167 / 170ms | resolving and compiling one principal’s effective access (p50 / p95 / max, cold) across 1,000 users · 1,000 roles · 10,000 grants; p95 0.02 ms warm | an automated build gate | a bulk-provisioned test dataset with a real hierarchy closure. “Flat, no cliff” is the claim. |
| 482ms | provisioning a schema-isolated tenant over HTTP: 94 tables at the same migration head, 89 policies in its own schema | single-node setup | single-node setup. |
| Number | What it measures | Where it was measured | Caveat |
|---|---|---|---|
| 36,220 · 31,151series | a full-portfolio model contest as a governed job: series classified · series judged; 1.16M verdict cells and 800k published forecast cells in 108.9 minutes; 1,090 series skipped honestly, 0 refused | a demo tenant | we quote the throughput only, never that run’s accuracy figure. |
| 14 vs 169ms | the same lifecycle run with a policy setting off, then on: the “off provably skips” proof; the footprint ladder that reports what a change reaches, 32 ms | single-node setup · an automated build gate | post-fix figures, after the adversarial review pass. |
| 9.2 % → 3.7 % | the median gap between the explainer model and the winning model’s forecast: how faithfully the explanation reproduces the number it explains; the worst-5 % tail 26 % → 17 % | a demo tenant | use as “the explanation got twice as faithful”, a fidelity figure, never an accuracy claim. |
| 0 of 2,310series | the measured verdict that a weather driver did not earn its seat: paired champion score 0.5093 → 0.5104 with weather in the room | a demo tenant | a demo world whose demand was generated without weather. This is the refusal story: a platform handed real weather that measured it and said no, not a finding about weather. |
| 26ms | a 10,000-edge bill of material exploding: multi-level, where-used, with quantity-per, scrap, effectivity, diamonds, cycle-safe | single-node setup · an automated build gate | the engine carries the bill of material by design; the manufacturing module is built with our first manufacturing customer. |
| 21Mcells | a demo tenant’s history landed through the governed connector pipeline, 0 rejects | a demo tenant | demo tenant. |
| Number | What it measures | Where it was measured | Caveat |
|---|---|---|---|
| 41 · 20 · 1 | widget types · studio surfaces · the one route the whole planner application renders from: every screen is metadata | the source tree | counts as of September 2026. |
| 8contracts | the platform’s swappable interfaces: control store, fact store, compute plug-in, connector, event bus, cache, job runner, feature store | the source tree | none |
| 59packages | the demand module, shipped as metadata packages rather than engine code | the source tree | count as of September 2026. |
| 54 of 439test files | backend test files that assert a timing bound: the performance budgets enforced in every build | the source tree | count as of September 2026; “budgets are tests”, not a coverage claim. |
| 0lines | of module code to build a whole application from an empty install, through the product’s own doors: the platform core’s final acceptance gate | the acceptance record | one application, walked end to end; the platform core passed every one of its acceptance gates before the first module was built. |
This page’s own vitals
What this page scored, dated
A performance page should be measured too. These are this page’s own lab results from the same gate every page of this site passes before it ships.
- Lighthouse · mobile
- 100 · 100 · 100 · 100 performance · accessibility · best practices · SEO
- Lighthouse · desktop
- 100 · 100 · 100 · 100 performance · accessibility · best practices · SEO
- Largest contentful paint · lab, mobile
- 1.7 s
- Cumulative layout shift · lab, mobile
- 0.00
- Total blocking time · lab, mobile
- 0 ms
- Measured on
- 2026-09-15 simulated mobile and desktop, the site’s own release gate
Third-party scripts on this page: none. Fonts: self-hosted. Stylesheet: inlined.
Measure it on your own data
Bring a slice of your history and your hierarchies. We show the engine working on them: the same budgets, your numbers.
Request a demo
a working demo on your data · no commitment