Company

We build the record that an organization’s decisions can rest on.

Reliable Context makes one product: a system of record for organizations that fuse private and public information and are asked, afterwards, to defend every value they acted on. It runs inside the customer’s perimeter, under the customer’s policies.

Request access Work with us

Engagement: How an engagement works
Contact: [email protected]
Deployment: Inside your perimeter

A cobalt blue keystone seated at the crown of a pale limestone arch, photographed straight on against a light grey wall.

A keystone carries load only once it is seated. So does a value: until it carries the sentence, the locator and the date it was read from, it is holding nothing up.

What we are

One product, one kind of buyer, one place it runs.

The company is deliberately narrow. Everything below is a constraint we accept rather than a market we are keeping open.

How the product works

the thing we build

A system of record, not a research assistant

Organizations in pharmaceuticals, financial services, government, legal, insurance, intellectual property and supply-chain compliance hold two halves of the same picture: internal material nobody outside can see, and public material anybody can contest. Fusing them produces a file that somebody will act on. We build the record that file is drawn from, with four properties a reader can check on any value: one canonical record per real thing, a cited and dated source behind every value, unknowns declared in words, and any past state reproducible.

The four guarantees

who we sell to

The person who has to defend the value

Not the person who wants a summary faster. The buyer is the analyst, reviewer, counsel, compliance officer or diligence lead who is asked, sometimes years later, where a number came from and why it was believed. That person does not need more retrieved text; they need the one sentence that asserted the value, the document it sat in as it stood that day, and a way to reproduce what the file said at the moment they signed it.

The problem, stated

where it runs

Inside your perimeter, one database per organization

The kernel is installed on the customer’s infrastructure and holds one database for that organization. Which sources may be read, which engines may see which material, which readers are cleared for which source, and how long documents and their derived values are retained are stated as the customer’s policies, not settled case by case. Values leave carrying their chain of custody, or they are labelled as withheld and attributed.

Inside your perimeter

The four neighbouring products each answer a question and return text. A system of record answers a question and returns the evidence for the answer, which is a different object and is built differently.

What this is not, argued

not a search product: Search returns documents ranked by their resemblance to your words and leaves the reading to you. The record has already been read: a value exists here because a sentence in a retained document asserted it, on a date.
not a data vendor: We do not sell a corpus, a feed or a licensed dataset, and nothing arrives already assembled from somewhere else. The record is built out of your own sources, inside your perimeter, and belongs to you the way your database does.
not a model: We train nothing and we sell no weights. Models read documents here, under your routing policy, and what they produce is admitted only as a value with a citation that a person can check against the source it came from.
not a dashboard: There is no chart, no score and no health colour. A record is read one value at a time, and the interesting question is never the aggregate: it is what stands behind this particular cell, and what would change it.

Why it exists

Why the company exists.

The company exists because of one failure we kept meeting in organizations that were doing the work properly. They had read the filings, the contracts and the internal memoranda. They had argued about them. And six months later, asked why a particular number was on a page, they could reconstruct the argument but not the evidence. The reading had been real; there was no record of it.

Almost every tool offered for that gap treats it as a retrieval problem: find the document again. That is the wrong shape. By the time a value is disputed, finding a document that supports it is easy and proves nothing, because you can find a document supporting nearly any value. What has to survive is the specific sentence that was actually read, in the document as it stood on the day it was read, together with the competing sentence and the stated grounds on which it was not carried.

So we build one thing and refuse the four adjacent things it could become. The discipline is that nothing reaches a reader without its evidence, that an absence is spoken aloud with a reason instead of appearing as an empty cell, and that any state the record was ever in can be put back. Those are the properties that make a file defensible, and they are unpleasant to retrofit onto a system that did not start with them.

We would rather a record say not found, and here is what was searched than be confidently wrong inside a file somebody signs. That preference is the whole company. Everything in the product is downstream of it, and so is everything we refuse to claim.

What we believe

Five positions the product is built to obey.

These are not values in the corporate sense. Each one rules something out, and each one costs us something we would otherwise be able to ship.

01
A name is never an identity.
A name is evidence about which thing you are looking at, and often weak evidence. Two companies share one trading name; one company carries three over a decade. A system that treats the string as the key will merge two counterparties or split one, and both errors are invisible at the point of use.
02
Absence is not evidence.
An empty cell reads to almost every reader as a negative. It is not one. Not found, blocked by a licence, and not applicable to this entity are three different statements about the world, and each of them owes the reader a reason and the record of what was searched and when.
03
A value without its sentence is a rumour.
Naming a source is not citing it. The unit of evidence is the sentence that asserts the value, quoted as it was written, with the locator, the date it was read, and a retained copy of the document. Anything less asks the reader to trust the intermediary instead of the source.
04
A record that cannot be replayed is a claim.
A decision is defended months or years after it was made, against a record that has moved since. Unless the state the decision rested on can be reproduced exactly, with the sources that supported it on that day, there is nothing to defend with except memory and a current page that no longer says the same thing.
05
Disagreement between sources is reported, not averaged.
When two credible sources conflict, the conflict is the finding. Picking the more recent, the more official or the more frequent hides a real dispute behind a single confident number. The competing value stays on the record with its own sentence, its own date, and the stated grounds for not carrying it.

How we work with organizations

Six commitments, half of them refusals.

Stated as properties of an engagement rather than as a sales process, because each one is something you can hold us to on the first entity and on every entity after it. There is no phase in which a value is allowed to appear without its evidence, and no phase in which material leaves your boundary to make the work easier for us.

What we ask for in return is narrow: the entity, the values that are contested, and the names of the people who disagree about them. We do not need your archive and we do not want a transformation programme attached to a record.

Inspect the demonstration Request access

One entity you already argue about
Not a corpus and not a pilot programme. A single company, facility, counterparty or programme whose file is contested inside your organization today, where you already know which values are hard and who disputes them.
The record is built inside your perimeter, under your policies
One database for your organization, on your infrastructure. Which sources may be read, which engines may see which material, who is cleared to read which source, and how long anything is retained are your settings, stated once and applied to every value.
Every value is defended or declared
A value carries the sentence it was read from, the locator and the dates. A value that has no evidence is a named negative with a reason, not a blank. There is no third state in which something appears on a surface and cannot be accounted for.
Nothing leaves without its chain of custody
An export carries the source, the excerpt, the locator and the dates alongside each value, in a form a reviewer can read without access to the system, and can be reproduced as of the date it was produced.
We do not ask you to move your data
There is no upload step, no shadow copy and no index of your material on our side. The kernel is installed where your documents already are. A source that may not leave your network is an ordinary case here, not an exception negotiated for you.
We do not name you
You will not appear in a logo wall, a case study, a conference slide or a press release, and nothing we publish depends on your permission to use your name. The two claims on our evidence page are written so that they can be made without identifying anybody.

How we work with researchers

What a research group gets, and what it is asked for.

The problems underneath this product are open problems in provenance, knowledge fusion, truth discovery and attribution evaluation. We would rather be corrected in public than be quietly wrong inside a customer’s file.

Research program

Data access under agreement: Evaluation sets released with their evidence rather than as answer keys: the retained documents, the verbatim excerpts, the locators and the dates that produced each value, so that a disagreement with our reading is checkable rather than a matter of opinion. The agreement fixes what may be redistributed, because part of the material is licensed and part of it is private.
Co-authorship on evaluations: If your group designs the measurement, you are an author on it. We are not looking for a validation exercise with our name on the result: we would rather the counting rule was written by somebody with no stake in which way it comes out.
The protocol before the result: What is being measured, on what data, with what counting rule and what would count as a failure, dated and published before the run. A protocol we publish after the numbers are in is a description of the numbers, and we treat it as one.
How to cite: The position paper carries the argument the product rests on, with its citations; the research page carries a BibTeX entry for it. If you cite a measurement rather than the argument, cite the dated protocol and the dated result separately, because they are two documents.
No embargo and no approval right: We do not read your write-up before submission, we do not hold a veto over a conclusion drawn from material we provided, and there is no clause that lets us withdraw data after seeing a result. The only constraint in the agreement concerns redistributing material we are not free to release, and it is written down rather than left to a conversation.

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

A joint only holds if both faces were cut to the same tolerance. An evaluation only holds if the counting rule was agreed before either side saw the numbers.

open problems we would like to be corrected on

Whether a decision made by a stated rule is actually better than one made by a similarity threshold, measured on cases where the two disagree, rather than asserted from first principles as we do on this site.
How to score a declared unknown, when the thing you would be scoring it against is also an absence and nobody has established it independently.
Whether a reader shown a competing value and the grounds for not carrying it decides better than a reader shown one confident value, or merely more slowly.
What a benchmark for a fused record could even be, when half of the evidence is private and cannot be published alongside the task.
How to detect the moment a retained source stops supporting the value that was read from it, without re-reading everything on a schedule.

The position paper Evaluations and reading

A group working on any of those is welcome to write directly. This is not the same thing as a commercial engagement, and it is not routed through one: the request we make of a research group is that the protocol is agreed in writing before the run, and that the result is published whichever way it goes.

The two things we claim

Everything else on this site is a property, not a number.

We make two factual claims about the company and nothing else. No logos, no named customers, no accuracy figures, no funding. The method behind both, and what each one would take to falsify, is on the evidence page.

Evidence and method

claim 01

In production use in two sectors.

Installed inside customer infrastructure, reading live private and public sources under those customers’ own policies, with values being acted on by people who do not work here. We do not name the customers and we do not publish volumes.

methodOne instance counts when its values are acted on outside this company; a pilot, an evaluation or anything running on our own machines does not count, and two customers in one industry would be one sector.

Counting rule and falsification

claim 02

More than six thousand identity decisions made by rule, none since reversed.

Each decision answers whether two records describe one real thing. Each was made by a stated rule against cited evidence rather than by a similarity score, each is readable back with the evidence it rested on, and none has since been reversed. The count includes decisions taken during evaluation runs; no independent audit has been performed.

methodA reversal is a later decision contradicting an earlier one about the same pair; because nothing is overwritten, reversals are counted rather than inferred, and the count is cumulative across deployments.

Counting rule and falsification

Neither claim is a performance measurement. The second is a count of decisions and a statement that no correction has been filed against any of them; it is not an accuracy rate, and a reversal, if one happens, changes the sentence rather than being netted off against the total.

Contact

Bring one entity you already argue about.

Write to us with the entity and the values that are contested. We will tell you what the record would carry for each of them, what it would declare as not found, and which of the contested values we think would stay contested after the evidence is in.

Research groups should say which of the open problems above they work on, and what data an evaluation would need. There is no sales step in front of that conversation.

Email: [email protected]
Enquiry: Contact and the structured enquiry
Engagement: Work with us
Access: Request access
Claims: Evidence and method