Work with us

Three ways in, and none of them start with a form you cannot see the end of.

You can collaborate with us on the research, build the kernel itself, or run it inside your own organization. They are different relationships with different terms, and it is worth saying which one you mean in the first sentence.

[email protected]The research program

We reply within two business days.

Two white machined blocks with interlocking dovetails pulled apart on a pale surface, the mating faces anodised cobalt blue.

All three relationships have the same mating surface: we hand over evidence, in a form that lets you attack it, and we expect you to.

Research collaboration

The problems underneath this product are open ones: identity from evidence, measuring a correctly declared absence, detecting the moment a retained source stops supporting the value read from it. We give evaluation sets with their evidence under a data agreement, co-author the write-up, and publish the protocol before the result and the result whichever way it goes.

What to send

Your group, the question you would put to the data, and what you would need released for the result to be reproducible by somebody else.

The research program and reading list

Build here

For engineers and researchers who want to work on records, provenance and evaluation rather than on a chat surface over a search index. There is no published list of open roles, no headcount plan and nothing to apply to. We read every mail that shows work. A thing you built, the problem it actually solved and what you would do differently beats any description of yourself.

What to send

One artefact, whether code, a paper, a system or a post-mortem, and two paragraphs on the problem it solved and where it failed.

[email protected]

Deploy it in your organization

The kernel runs inside your perimeter, with one database for your organization; sources, model routing, reader clearance and retention stay your policies. Start with one entity your people already describe two different ways, and read the record it produces before any commercial conversation happens.

What to send

The entity, the private and licensed sources you fuse to describe it today, and who in the organization is allowed to read which of them.

How deployment inside a perimeter works

For the third, the request access form is the shorter path.

The practice

What the work is actually like

Six rules the codebase enforces rather than aspires to. They are listed here because they are the parts people find constraining, not the parts that sound appealing.

01
A judgement is recorded as a decision with its evidence, never as a special case in code.
When a record behaves wrongly, the fix is the rule or the decision that was missing, applied to everything it governs. Patching the one entity that revealed the problem is the defect, not the remedy, and it is the most common thing a new engineer has to unlearn here.
02
An absence is written down as an absence.
An empty result is never allowed to become a negative finding. Not found, blocked and not applicable are values with reasons attached, and asserting one requires evidence of the absence itself, cited like any other value.
03
A correction is another entry, never a rewrite.
Nothing in the ledger is edited or deleted. A wrong value is superseded by a new one that says what changed, on what evidence and when, so a reader can reconstruct what the record said on the day somebody acted on it.
04
A measurement is published whichever way it goes.
The protocol is written before the run, and the number that comes out is the number that is published. We would rather be corrected in public than be quietly wrong inside a customer's file, which means work here is regularly shown to be worse than hoped.
05
Reading a sentence is model work; code joins identifiers.
No pattern over prose stands in for understanding it. Deterministic code joins mechanical things, such as a locator, a date or a registry number, verifies excerpts verbatim and assembles the record. The moment a regular expression is working out what a sentence means, it is wrong.
06
The interesting work is small and unglamorous.
Much of it is deciding what a value means, which of two documents is describing the same quantity, and how to prove either. There is no demo-driven roadmap; the surface a customer reads is judged by whether it can survive being disputed six months later.

Before you write

What we are reading

These four are on the site because they are the things we would want an applicant or a collaborator to have read first. None of them is a summary of the others, and each contains something we can be held to.

The position paper

/paperRead this first if you intend to disagree with usThe position the product is built on, including the parts we cannot yet measure and the claims we have deliberately not made. A mail that argues with a section of it is the mail we answer first./researchThe open problems, and the literature we work fromIdentity from evidence, measuring a correctly declared absence, detecting when a retained source stops supporting the value read from it. If your work touches one of these, say which and say where our framing is wrong./evidenceHow the two claims on this site are countedThere are two numbers on this site and nothing else. This is how each is counted, what would make us withdraw it, and which measurements came out worse than we expected./guaranteesThe contract you would be maintainingOne record per real thing, a cited and dated source behind every value, declared unknowns, any past state reproducible. Building here means being bound by these four on days when it would be easier not to be.

what a good first mail containsWhich of the three routes you mean, in the first line. One thing you made, with the problem it actually solved and the place it failed. A sentence on what you think we have got wrong, if you think so. It does not have to be long, and a post-mortem is worth more to us than a description of yourself.

Write to us

Say which of the three, and show the work.

The form below is the same one as on the access page; the last field is where the useful part goes. If you would rather send an attachment or a link, mail is better.

Direct: [email protected]
Research: The reading list and the open problems are on the research page; the position paper is at /paper.
Deployment: Request access and the perimeter answer most of the first round of questions.

Start a conversation

The same form for all three routes. Say which one you mean in the last field. It is read by a person and answered by email.