A system of proof
What does your data actually do?
Cosmogenesis reads the code your business runs on and shows what it does, graded by how well each part is understood. Not what someone said it does.
In private beta with a small number of institutions · Patent pending
balance < 0 → <= 0
meaning changed here · 41 days ago
A one-line change, in a weekend release. Every test passed.
Nothing broke and no alarm fired. From that day, the figure was calculated differently.
So the estate was read, end to end.
The same read, run before the release, would have found it.
Somewhere in your company, right now:
- a number the board sees every month that no one can fully explain.
- a control you trust, resting on logic no one has read in years.
- a migration that's been priced twice and never started, because no one can see how it all connects.
- one person who knows how it actually works, and a retirement date.
Your catalogue can name all four. It can’t answer for any of them. Neither can the lineage tools it came with.
Every enterprise has a system of record: a place that holds what people say about their data. None has a system of proof. Cosmogenesis reads the code your business runs on and holds what it does, something you can understand well enough to change safely.
A system of record is maintained by hand, and out of date the day it is signed. A system of proof reads the source, shows what each part does, grades how well it is understood, and returns the same answer every time. Read-only, on your own infrastructure. What it cannot read is marked, not left out, and every finding is traced to the code it came from.
From a column up to the control that depends on it.
The read follows the path all the way down, past where data moves and into what the logic means. More than one source feeds the same logic; every strand is followed, not just the one that broke.
Every link carries a confidence grade.
Measured, inferred, or simply unknown; we won't dress a guess up as a fact. Where confidence drops, risk sits unexamined, rated and bound to the control that should own it. The real screens are just below.
The platform
The engine, reading an estate.
A recording and three screens of the engine reading a real estate: seven stages across SQL, Alteryx and R. Every name invented, nothing mocked up.
Recorded live from the engine.
One read of the whole estate.
Findings ranked by exposure.
The record shows its own gaps.
The system map: the domain opened into its seven stages, one stage locked for its dossier, then drilled into its operator graph.
seeded with the failure modes the engine looks for · see a sample record →
Where to point it first.
A change ships
A one-line change alters what a trusted figure means. The read finds it, dates it, grades it.
▶ play it in the demoA migration lands
Both estates read the same way, the move scoped from the map, completion measured against the read of the rebuilt estate.
▶ play it in the demoAn LLM comes in
Before a model uses the estate, every field is graded by how far it can be trusted, the same way on every run, so what is safe to feed is separated from what is fenced off.
▶ play it in the demoThe expert leaves
The map and the documentation stay, and regenerate on demand.
▶ play it in the demoControl coverage, traced
The read follows the critical flows down to the code and shows whether each control is wired to what it claims to cover. Where a control lives in code, we show the wiring as written; where it lives with people, we say so. The gaps surface in your own review, not in someone else's.
Logic observability
Once an estate has been read, it can be re-read, and changes and drift surface as they appear, including the ones that alter what a number means.
There are more: due diligence, data-quality forensics. Each is the same read asked a different question. Point it at the part of the estate you trust least.
What it gives you.
One reading of your estate. Everything below follows from it.
legibility
Your estate, readable end to end, including the parts no one has looked at in years.
impact analysis
What a change would touch, mapped before you make it.
documentation
Generated from the code itself, regenerated each time the estate is re-read.
evidence
The path behind a number and the controls around it, in a record you could put in front of a regulator.
a foundation
A graded picture of the estate as it is today: the starting point for anything you build on it.
What changes, seat by seat.
Attestation rests on walkthroughs assembled by hand, and on trusting that they were assembled well.
It rests on generated evidence: each step graded, and the same record on every re-run.
A proposed change means days of impact analysis, and the answer still arrives as an opinion.
What a change would touch is known before it reaches production, with the confidence of each link stated.
Sign-off is rebuilt from scratch every period, from queries, spreadsheets and memory.
It is supported by artifacts the read generates, which carry their own evidence and disclose their own gaps.
Using a figure safely means finding the specialist who remembers what it means.
What a figure means, and how far to trust it, can be confirmed before it is used.
Where this goes
What will your AI stand on?
Everyone is being asked for their AI story. Ours is quieter: an agent cannot be trusted on an estate nobody has read. So we read it first, as it stands.
None of that reading is AI. It’s deterministic: the same estate in, the same answer out, every time.
What reads your estate once can keep reading it: understanding first, then a standing watch.
When we don’t know something, we say so.
Why we built it.
Built by people who spent years inside some of Australia’s largest regulated data estates, and kept finding logic nobody had read doing work nobody could account for. So we built the system of proof we needed then.
A small team in Sydney, early-stage by design, proving the platform ahead of a broader release.
See it against your own estate.
Start with a conversation. Then watch it read a slice of your estate, inside your perimeter.