The reviewer is not the author
Whatever proposes a hypothesis never gets to judge it. The reviewer's instruction is to attack it on evidence and then rule honestly — including ruling that its own attack failed.
Cross-domain discovery engine/beta
Don't wait for
serendipity.
Invite it.
Every point is a piece of published work; every cluster, a field that grew on its own, in its own vocabulary. The library has never had corridors — this is Intercognix looking for the crossings.
In 1986, two medical literatures — dietary fish oil, Raynaud's syndrome — had never cited each other. Read together, they implied a treatment; a trial confirmed it three years later. Seismology's aftershock model beat trained analysts at forecasting burglary. Spin-glass physics explained how a protein finds its fold. Every time, both halves were already published. Only the bridge was missing.
Intercognix goes looking on purpose. Give it a problem you are stuck on: it searches deliberately distant fields for a mechanism that fits, carries it back into your terms, and hands it to an independent reviewer that tries to break it. What survives carries a measurable variable, a decisive test, and the observation that would kill it.
Not a summary. Not an answer. A proposal you can test — or discard in an afternoon, which is equally the point.
Intercognix n. — a candidate insight discoverable through previously unrecognized relationships among independently known things. inter, between · cogni, knowing · x, the unknown Access is granted to researchers. The engine is in beta and in active development; the people running expeditions now are shaping what it becomes.
What a run produces
- Hypotheses ranked into three tiers, with the reasons stated
- Every citation rated against its source — verified, partly, not found
- A decisive test and an explicit falsifier for each one
- An auditable record of every step, so a reviewer can check the reasoning
- A research-edition PDF you can circulate
A redacted sample report is shared with researchers once access is granted.
How a run works
Four movements, each one recorded, so any conclusion in the report can be traced back to the step that produced it.
- 1Your problem, stated as phenomena It is restated as what is unresolved rather than as a hypothesis, so the search is not quietly aimed at an answer someone already holds. What you want ruled out as a standalone cause is recorded and honoured.
- 2Distant fields, chosen on the evidence Candidate disciplines are scored for how far they sit from yours and how well their mechanisms fit the behaviour you described. A field that does not earn its place is not searched — the search goes where the transfer is worth something, not where it is easy.
- 3The mechanism carried back into your terms What is found is translated into a concrete, measurable claim about your problem, with a decisive test and the conditions under which it must fail written down before anyone likes it.
- 4Independent adversarial review A reviewer that did not write the hypothesis searches the literature for prior art, checks every premise, rates every citation against its source, and tries to reject it. What survives is ranked by whether you could test it now, and with what.
What makes it different
Distance is a budget, not an accident
Every run spends a set share of its search on near, medium, far and maximally distant fields. The far bands are where the transfer is worth anything; left to itself, a model stays home.
Citations are rated, not trusted
Each supporting claim is checked against the source it cites. Verified, partly, not found or junk — printed in the report next to the claim it was supposed to support.
Falsifiers are mandatory
A hypothesis with no stated way to be wrong does not survive the pipeline. Every result you read comes with the observation that would kill it.
Nothing is hidden behind a summary
Every step of every run is recorded and hashed, so you — or a reviewer, or a journal — can confirm that what the report claims is what actually happened, and that nothing was edited after the fact.
Prior art is an output
When a hypothesis turns out to be already published, that is recorded with the citation — a finding in itself, and one the next run starts from.
What a report looks like
A full expedition report runs to a hundred pages or more: every hypothesis with its mechanism, its decisive test, its falsifier and its reviewer's verdict, every citation rated against its source, and the complete record of how each conclusion was reached.
A redacted sample is shared with researchers when access is granted, rather than published here. The engine is in beta and the first hypotheses are still being tested; until that feedback is in, the honest thing is to show the work to the people doing it rather than to advertise with it.
What it costs, and what is yours
During beta, nothing is charged by Intercognix
You add your own provider API keys and pay your own model usage. A full expedition scores ~280 candidate fields, searches the ~40 that earn it, and formulates and independently reviews ~25 hypotheses — for about $20–30 on the default model mix. Nothing is billed through us, you choose the models, and you see the estimate before a run starts.
The hypotheses are yours
Test them, discuss them, put them in a grant, publish them. Intercognix claims no ownership, no licence and no right of review. Two courtesies are asked: tell us before you publish, so anything unprotected on the method side can be filed first, and name the tool.
Your runs are private
A run is visible to the account that started it and to nobody else. Problems, hypotheses and reports are never shown to another user, and are not used to train anything.
The method is what is protected
The confidentiality agreement covers how Intercognix works — its stages, scoring and prompts — and the unpublished filings behind it. That is the whole of what it asks you to keep to yourself.
What it would cost to do this properly
One expedition is a probe. The interesting question is what a systematic search costs — a whole programme's worth of problems, or one problem searched exhaustively. Because independent review dominates the bill, the cost scales close to linearly with the number of hypotheses assessed, and the arithmetic is unusually easy to check.
| Hypotheses formulated and reviewed | Distant fields scored | Model usage | What that buys |
|---|---|---|---|
| 25 one expedition | ~280 | ~$28 | a first look at one stuck problem |
| 100 | ~1,100 | ~$110 | one problem searched from four independent angles |
| 1,000 | ~11,000 | ~$1,100 | a lab's whole portfolio of open questions, or one question exhausted |
| 10,000 | ~110,000 | ~$11,000 | a programme-scale sweep: every unresolved problem in a field, against every field |
Measured, not modelled: about $1.10 per hypothesis formulated and independently reviewed, taken from completed runs ($0.75–1.35 across them). The figure follows the models you choose — a heavier reasoning model on the review side costs more per hypothesis and is worth it on a hard problem, a lighter one costs less and is enough for a first sweep. You set that per expedition, and you see the estimate before it starts. At ten thousand candidates the whole search costs less than one month of one postdoc, and returns ten thousand hypotheses each carrying a decisive test and a falsifier — of which the ones that survive review are the only ones you would ever have read anyway.
No elapsed time is quoted here on purpose. The stages are independent and the work parallelises; how quickly a sweep finishes is a function of how much of it you run at once and of the rate limits on your own provider accounts, and that ceiling is being raised rather than promised.
What it is not
Every specimen remains a candidate until validated. Each output is a proposal generated by a language model and examined by another; nothing has been run in a laboratory. Tier 1 means "testable now, with data or a platform you can reach" — not "true". Most candidates collapse under evidence, prior-art checks and expert scrutiny, and they should.
The review can be wrong in both directions. It can miss prior art that exists, and it can reject something sound because it could not find the supporting work. The rating on each citation is there so you can see which happened.
It does not replace domain judgment. The value is in putting a mechanism from a field you do not read in front of someone who knows what would be surprising. That someone is you.
The engine is in beta. Stages are still being sharpened and the scoring still moves. Expect the occasional failed run, and expect the reports you read this month to be better next month — partly because of what you tell us.
Who is behind this
The word Intercognix was coined by Jay Gontu in 2026, to name what Don Swanson had found and not named: not a forgotten fact awaiting retrieval, but an implication of what humanity already knows that humanity has never known. The idea, and a growing catalogue of candidates, is set out at intercognix.com. This service is the engine built to hunt for them.
It is built and run by one researcher, not a company. An expedition you request is read by a person, and the feedback you send is read by the same person.
Running a lab, a group or a programme? Cross-domain search works best against a portfolio of stuck problems rather than one. Say so in the form and mark it as a group — pilots are arranged directly.
Invite serendipity to your problem
Access is granted to researchers. Submit the form below to run an Intercognix Discovery Expedition on the problem you are trying to solve — a search for hypotheses that could measurably shorten the road to an answer. Tell us what you are working on and what you are stuck on; that is what the decision is made on. Everything an expedition produces is confidential, and you will be asked to accept a confidentiality agreement before your first run.