One notebook. 105–118 GB/s.
Apple M3 Max class, 16-core CPU, 64 GB unified memory, CPU paths only. Compressed columnar storage, dictionary encoding, declared primary-sort layout. One cold run + five warm runs per shape; best pass and median retained. Throughput = logical decoded-equivalent GB/s.
| Shape | Rows | Layout | Best | Median | Throughput |
|---|---|---|---|---|---|
| byte_key × SUM(Int32) | 1 B | sorted dictionary key | 32.0 ms | 32.1 ms | 116 GB/s |
| byte_key × SUM(Int64) | 1 B | sorted dictionary key | 69.1 ms | 69.4 ms | 118 GB/s |
| byte_key × SUM+MIN+MAX+COUNT(Float) | 1 B | sorted dictionary key | 67.2 ms | 69.7 ms | 112 GB/s |
| town × SUM(price) | 100 M | sorted string dictionary | 3.6 ms | 3.6 ms | 108 GB/s |
| property_type × MIN/MAX/AVG/COUNT | 100 M | sorted string dictionary | 3.4 ms | 3.4 ms | 111 GB/s |
| Q1-style linestatus reduction | 100 M | dictionary key + predicate | 16.0 ms | 16.3 ms | 99 GB/s |
Scope, stated plainly: these are internal measurements of selected execution paths under a declared physical layout. No formal TPC-H/TPC-DS claim. No public head-to-head claim until comparable engines run on the same machine, same data, same layout assumptions and same cold/warm policy.
Not a kernel. A whole pipeline.
103,000,000 input rows through a 21-stage workflow — PII protection, deduplication, enrichment, SQL aggregation, quality checks and multiple outputs — in 24.4 seconds: ≈ 4.2 million rows per second across the entire workflow, on one notebook.
This demonstrates end-to-end workflow performance including orchestration, transport, checkpointing and multi-output writes — not an isolated analytical kernel. Full stage list and per-stage timings in the workflow report.
4.8 ms. Same corpus, same machine.
| Metric | Indago | macOS Spotlight | Factor |
|---|---|---|---|
| Warm query, median | 4.8 ms | 81.7 ms | 17× |
| Index build (48,899 files, 1.19 GB) | 37.2 s | 320.0 s | 8.6× |
| Morphology / word-form queries (hit@1) | 30% | 0% | — |
| Compound-split / hyphen queries (hit@5) | 100% | 0% | — |
| Typo / fuzzy queries (hit@1) | 50% | 0% | — |
macOS 26.5 · Apple M3 Max · 64 GB · mixed real-document corpus (48,899 files, corpus hash pinned) · 07/2026. Fairness of capability-class probes verified empirically, not by lexical disjointness. Reproducible harness included in the pack.
Declared layout. Declared data.
Numeric reductions: 1,000,000,000 generated rows (low/mid-cardinality dictionary keys, compressed value columns). String-key reporting: 100,000,000 records derived from public UK property price-paid records — source, release date, transformation to benchmark scale and checksums are part of the methodology pack, together with reproduction instructions and raw timings.
Counted, not estimated.
| Catalog | Count | As of |
|---|---|---|
| Protor connectors & processing modules (combined) | 200+ | 07/2026 |
| Indago knowledge sources | 223 | 07/2026 |
| Indago analysis tasks (enabled) | 242 | 07/2026 |
Catalogs are generated from the live registries and exported with every release.
One core. Every platform.
| Platform | x64 | ARM64 | Notes |
|---|---|---|---|
| Windows | ✓ | ✓ | Protor engine + apps |
| Linux | ✓ | ✓ | incl. Raspberry Pi (edge / embedded) |
| macOS | ✓ | ✓ | Protor engine · Indago desktop, daemon, CLI, MCP |
Cross-platform core with platform-specific high-performance kernels. Distributed cluster execution is shipping; the formal cluster-scaling benchmark is scheduled and will be published here.
Externally verifiable.
The benchmark brief is written for hostile reviewers: declared layout, decoded-equivalent throughput, stated cold/warm policy, no comparison claims without a same-machine run. It is circulating for methodological review. We use the phrase “externally verifiable” — and will upgrade it to “independently replicated” only when a named party has reproduced the results.