Your question, worked for weeks by a team that shows its sources.

Coalby is a persistent team of research agents. You set one business question. The team works it over weeks, not seconds: it gathers evidence, keeps checking as things change, and comes back to you when something material moves. Every claim carries the source it came from. The judgment stays with you — that is what "Led by You" means.

How it works

One

You set the question. One question that matters to your business, with the decision it feeds and the date that decision is due.

Two

The team works it over weeks. Specialists gather sources. Each finding is a dated, sourced observation, kept as it was found; corrections are added, never overwritten.

Three

A checker who did not do the work reviews it. Claims that survive reach you with their source attached. Claims that do not are labelled, not deleted.

Four

It keeps running. A weekly digest of what moved. If nothing material changed, the digest says so and stops there.

Checked, not plausible

A one-shot search engine gives you a plausible answer in seconds. Plausible is the problem: it reads exactly like a checked answer, and you cannot tell them apart in the room where you have to defend it.

Our verifier specification sets the rule we work to, and it is deliberately unfriendly to our own output: every claim starts out unsupported. It only becomes supported when an agent that did not write it goes and retrieves the cited source and finds the statement actually there. Absence of evidence stays unverifiable — it never quietly turns into support. A named category like "trade press" or "public records" is not a source.

supportedThe source was retrieved and it says this.
contradictedThe source was retrieved and it says otherwise.
unsupportedThe source was retrieved and the statement is not in it, or no real source was named.
unverifiableThe source could not be retrieved. We tell you which claim this was.

Two further rules you can hold us to. Estimates are labelled as estimates and carry their method and their error band in the same sentence as the number — "about 40 firms" is a site count, and the company count will be lower. And a general industry rate applied to your situation is marked as what it is, a prior, not a finding about you.

"We don't know" is a valid answer, and you get it in writing rather than having to detect it. So is "nothing changed this week".

Two shapes of engagement

market map

A defensible universe of companies, with the evidence behind every row

For a buyer, an acquirer or a fund that needs to know who exists in a market, which of them fit a stated thesis, and how sure anyone can be about that. The product is the credibility of the database, not the length of the list: for any row, where it came from, when, how sure we are, and what would change our mind.

You receive: a ranked shortlist, an evidence panel per row, a full export, and a weekly changelog of what moved in the market.
The gate before it reaches you: the top rows are source-checked by an agent that did not build them; coverage is published as a lower bound with the test that produced it, not as a claim of completeness; every modelled number states its method.
What it is not: a list scraped from a directory. Where a source cannot carry a field, the field stays empty and the gap is named.

competitor threat radar

Which competitor should worry you in the next twelve months — and what changed this week

A one-shot answer can list your competitors. It cannot tell you which weak signal is accelerating: early hiring, a filing, a grant, a price change, a partnership. Those only mean something watched over time, which is the part a single query cannot do.

You receive: an opening ranking with the criteria written down before the ranking, the evidence per competitor, and a plain statement of what would prove the top threat is not the top threat.
Then, weekly: a digest of what moved, what it means, and how the ranking changed. When nothing material changed, the digest says "no material change" and names the next check date. We do not ping you daily.
The gate: no claim that something shifted leaves without a verdict from the checker, and anything unverifiable is labelled as such inside the digest you read.

Where we are today

Held to our own standard, here is the honest state of this, on the same terms as anything we would send you.

We have not yet delivered a verified engagement to an outside client. No client names appear on this site and none ever will.
Both engagement shapes above are written up as playbooks and are in build. Their gates are real and are what we would be held to; they are not yet a record of completed outside work.
The first threat radar is running on ourselves, on the market we are in, because that is the cheapest honest pilot. Its opening ranking was corrected by its own verifier pass before we looked at it again — the top entry survived, but for a different reason than we first wrote down, and the margin narrowed.
The verifier is at version 0.1 and has had its first run against real claims. Where our own coverage is thin, we say so here rather than waiting to be asked.

Start a conversation

Send one email with your question in it. Useful things to include:

The question, in your words.
The decision it feeds, and when that decision is due.
What you already believe the answer is, and what would change your mind.

Email us your question

Or write to adam@aubrains.com. An engagement is $500 a month and ends whenever you say it does. We will reply with what we think we can evidence, what we do not think we can, and what it would cost. If a single search answers your question in ten seconds, we will tell you that instead of selling you weeks of work.