Trust
Custody you can check. Maths you can install.
Your tenant. Never pooled. Never trained on. Deletable. The engine that computes every number is open and runs offline.
Open core
You can read the code that computes the numbers.
Every statistic on every screen comes from a package you can install yourself, offline, without an account. Importing it makes no network call. If any claim on this page were untrue of the engine, you could find out in an afternoon.
Installable
pip install alphaengine. Two dependencies, numpy and scipy. It runs with the wifi off.
Inspectable
The maths, the corrections and the tests that keep them honest are all in the open repository.
Reproducible
A published number is a contract. A computed value that changes forces a version bump, so a study saved today still reproduces in two years.
The output
Why you can trust what it tells you.
Any system can promise to be careful. These are the four places where being careful is the only thing the code will let it do.
It cannot call something good without showing its work
A result is only allowed to be called an edge if the correction for how many things you tried has actually been computed. Not "should be". Cannot. Two separate parts of the system check it, and they were written independently of each other.
It cannot quietly change its mind later
A decision, once written, has no edit path. There is no function to change it, because none was written. Closed positions are additionally chained together, so removing or altering one is detectable by anyone you hand the record to.
A view on the market can only make you smaller
When the system reads the wider environment, that reading is allowed to reduce a position and never to increase one. The direction is preserved and the size can only shrink. A change that broke this would stop the build, not produce a bigger number.
It says it does not know
Where a figure is unavailable, the answer is that it is unavailable. Coverage gaps, thin samples and dates that were inferred rather than given all come back labelled. Nothing is filled in to keep a page looking complete.
Your data
Where the line is drawn.
The line is not around what we hold. It is around who can reach it.
Goes in
Your data
series you upload or connect
Your judgment
mandate · theses · journal
Your record
studies · fills · decisions
Your tenant
Encrypted, scoped to you
exportable in full · deletable in full
Cannot leave, to
Another customer
never pooled, no cross-tenant aggregate
A training set
no model, ranking or default learns from it
A resale
your licence, your entitlement, never redistributed
One container, three sealed exits. What matters is not whether your data is held, it is that nothing carries it out: not to another customer, not into anything that learns, and not back onto the market it came from.
Four commitments
We hold your data. Here is what that does and does not mean.
Plenty of tools promise to hold nothing. It is an easy thing to say and a hard thing to keep, and an operating system that held nothing could not run a monitor overnight or reproduce a study two years later. So we hold what the job needs and we are specific about the terms.
It is yours
Encrypted at rest, scoped to your fund, exportable in full and deletable in full, at any time and without asking us. Holding it is what makes an overnight monitor and a two-year-old reproduction possible; owning it is what makes that safe.
It is never pooled
One customer's data never sits in a store another customer can read, and no aggregate, benchmark or ranking is derived across customers. Every query that reads your records is scoped in the data layer itself, not in the page that displays it, because a check that lives in a route is one refactor from being gone.
It is never learned from
Nothing you give us trains or tunes anything. Not a model, not a ranking, not a default, not a suggestion shown to somebody else. Your proprietary model and your strategy logic are never ingested at all, so there is nothing to learn them from.
It is never redistributed
Vendor data you supply stays under your licence and your entitlement. We use it on your behalf and we do not resell it, republish it, or serve it to another customer. Your licence terms are not our business model.
Two things survive unchanged from the older, narrower version of this promise, because they were about custody rather than absence: your proprietary model is never ingested, and the language model never originates a number the system trades on.
Line by line
Everything, and where it lives.
A test suite fails any change that would leak one account's records into another's, or leave a new table out of the export and the delete. The list of your tables is derived from the database itself rather than maintained by hand, so a table added next year is covered whether or not anyone remembers it. The boundary is not a policy. It is a gate.
The rules
Six commitments, enforced in code.
Authorization lives in the data layer, not the page.
Every query that reads your records carries your account in the query itself. A filter applied by the screen is one refactor away from not being applied at all, so the ones that matter are not written there. Attempted cross-account reads are in the test suite as attacks, not as happy paths.
The model never originates a number you trade on.
Every figure is computed by the deterministic engine and bound to a source. The language model arranges those facts and cannot invent the number. Where it offers judgment inside a run, it is allowed to veto or shrink a position and never to enlarge one.
A result cannot claim more than the evidence behind it.
An outcome is only allowed to be called an edge if the correction for how many things you tried has actually been computed, and if the number of things you tried was actually recorded. Leaving that count out does not get you a generous answer, it gets you an inconclusive one.
You can export everything, and you can delete everything.
A full dump on request, and a delete that is complete. A test derives the list of your tables from the database itself rather than from a hand-written list, so a table added next year is covered whether or not anyone remembers it.
What compounds is your process, not your edge.
How you size, what you check before committing, what you said would falsify an idea, how your conviction has actually paid. Two funds running opposite strategies can share a way of working, and that is the only sense in which the system gets better with use. It never gets better at your strategy, because it never sees it.
The engine is open, so none of this has to be taken on faith.
Every statistic on every screen comes from a package you can install and read. If a claim on this page were untrue of the engine, you could find out in an afternoon.
Decision support, with you in the loop. The engine never touches a trade. That is a feature and a line we do not cross. Computational tooling, not investment advice.
Do not take the maths on faith either.
Skeptical? Install the engine, generate data with a known answer, and check it against the result. You hold the answer key, and you never had to give us anything to find out.