Polars and DuckDB handle import and export in the open. The C calculation core stays closed for IP — but every metric is documented and backed by peer-reviewed research.
A black box that you can’t inspect is not a tool, it’s a dependency. So we made the data layer transparent — you can see every byte coming in, every byte going out. The calculation core is closed for commercial reasons, but every formula it computes is documented and grounded in academic literature you can verify.
The import and export layers are built on Polars and DuckDB and shipped under CPAL 1.0 — you can read the source, fork it, and audit exactly what touches your data. Schema detection, format conversion, BI star schema generation: all transparent.
Between them sits the calculation core, written in C, kept closed for IP protection. But closed does not mean opaque: every metric the core computes is documented with its formula, its assumptions, and its peer-reviewed academic source. You can’t compile the binary, but you can verify the math.
The data layer is built on Polars and DuckDB and shipped under CPAL 1.0. Read the source, fork it, audit it. Eight enterprise platforms supported out of the box, plus standard formats — CSV, Parquet, JSON, SQL, Excel — and a star schema export for any BI tool. Your data flow has no black boxes.
The C kernel handles the hot path — the loops, the joins, the conformance checks, the Monte Carlo runs — with SIMD instructions and zero-copy memory. Eighty to one hundred times faster than pure Python. The binary is closed source for IP reasons. Every formula it computes is fully documented in the public reference.
Every metric in the core cites a peer-reviewed source — van der Aalst on alpha-algorithm conformance, Goldratt on theory of constraints, Shewhart and Deming on statistical process control. We’re not inventing measurements. We’re implementing the field.
Every metric the core computes ships with its formula, its assumptions, and its peer-reviewed citation. You don’t need to read the binary — you can verify the calculation by reading the paper. Any statistician with the source data can reproduce our numbers using a different implementation, and they should arrive at the same result.
# Reference excerpt — value_yield metric METRIC value_yield RETURNS ratio in [0.0, 1.0] FORMULA vy = sum(events where value_add = TRUE) / count(all events in case) CONFORMANCE Lean value-stream mapping (Womack & Jones, 1996); token-replay alignment (van der Aalst, 2016). ASSUMPTIONS · Each event mapped to value/non-value via activity_semantics column · Rework events classified as non-value · Queue/wait events excluded from numerator CITATIONS Womack, J. & Jones, D. (1996). Lean Thinking. van der Aalst, W. (2016). Process Mining: Data Science in Action. 2nd ed., Springer. # Full reference: ghostcitadel.com/observatory/documentation
Closed-source calculation does not mean closed behavior. The library does not phone home, does not log usage, does not require authentication. You can verify the network silence with the same tools you’d use on any binary — closed or open.
The library does not initiate any outbound network connection. Verify with strace, tcpdump, or an air-gapped install. Closed-source code, observable behavior.
All engineering is based in the UK and EU. The codebase, the kernel, and the maintainers are subject to European data protection law — not extraterritorial demands from jurisdictions you never agreed to be governed by.
Every metric is documented to the formula. Anyone with the source event log and the public reference can reproduce our numbers using a different implementation. Closed binary, open verification path.
Continue Exploring
Value Yield, Flow Geometry, Monte Carlo futures. The economic outputs.
See the diagnostics 05 · DocumentationEvery metric formula, every academic citation, every deployment recipe.
Open the docs ← HubAll six paths into the engine, from one map.
Back to the hub