DarkSigmaVisit Atlas
DarkSigma, Inc.MethodFresno, California

How the engine works.

What the engine indexes underneath the words, shown on two papers that share none of them, and measured against fifty transfers that already happened.

50
cross-field transfers that already happened
21 / 50
times it reached the field that solved the problem
3 / 50
times a leading AI search did

The reduction

A physics paper stopped a heart arrhythmia. Nothing in it mentions hearts.

Ott, Grebogi and Yorke published a way to hold a chaotic system on a regular cycle in 1990. Two years later a cardiology group used it on a rabbit heart made arrhythmic by a drug. Given the cardiology problem statement alone, the engine returns the 1990 paper at rank 7.

Cardiology, 1992

In a rabbit heart made arrhythmic by a drug, the intervals between beats become irregular instead of settling into a steady rhythm. The two available responses both act on the whole tissue at once: more drug, or one large shock.

Nonlinear dynamics, 1990

An unstable periodic orbit is held near a saddle fixed point by small perturbations applied at the Poincaré section, timed from the return map rather than from the full trajectory.

What both of them say, with the vocabulary removed

A system whose cycle has gone irregular is returned to a regular one by small, precisely timed corrections applied at the moment the cycle repeats, instead of by a single large intervention.

Garfinkel, Spano, Ditto and Weiss made that connection by hand and published it in Science in 1992. It is one of the fifty cases in the benchmark below.

The benchmark

Fifty problems whose answer is already on record, run three ways.

Fifty times in history, a researcher cracked a problem by taking a mechanism from a field they had no reason to read. Every one is on record, with the paper that made the connection. We restated each problem in the asking field’s own words, as of before the transfer existed, and on 2 and 3 September 2026 gave that restatement, and nothing else, to the three arms.

Read the benchmark
AtlasA leading AI searchThe same search, our wording
Came back with work from outside the researcher’s own field49 / 5013 / 5039 / 50
Came back with the field that historically solved the problem21 / 503 / 502 / 50
Papers it found in that field, out of 300 read per arm3873
Returned the specific paper the breakthrough was built on10 / 430 / 430 / 43

The third column is the same commercial index, run a second time on our neutral restatement of the problem. It leaves the researcher’s own literature almost as well as we do, 190 adjudicated papers against our 192. It reaches the field that solved the problem twice, against three for the literal wording and twenty-one for us. Leaving a literature and arriving in the right one are separate capabilities, and the rewrite supplies only the first.

The last row scores all three arms over the same forty-three cases: the source paper had to be identified against a public catalogue, predate the bridge paper, and carry an abstract. Forty-seven of the fifty were identified and forty-three of those could be ingested. Seven of our ten came back inside the top twenty, where a person sees them without being told to look.

Depth

Scrolling further down does not get you out of your own literature.

Share of each list that was work from outside the field the researcher already reads, by position in the results.

4 to 9%

of everything a leading AI search returned, at every depth from rank 1 to rank 20, was work from outside the field the researcher already reads. Atlas ranged from 57 to 72 per cent.

RanksAtlasA leading AI search
1 to 557%5%
6 to 1059%4%
11 to 1570%9%
16 to 2072%7%

In 37 of the 50 cases the comparison returned nothing from outside that field at any depth. The paper states all of this the other way round, as the share of each list that was the asking field’s own literature: 95, 96, 91 and 93 per cent for the comparison, against 43, 41, 30 and 28 for us. The figures above are those subtracted from a hundred.

Controls

Every check, and what it returned.

CheckWhat it would have caughtResult
Answer keyRelevance judgements written by the people running the test50 cases, each recorded by its own researchers, DOIs in the paper
Leak auditSource-field vocabulary sitting in the problem statement251 terms checked, 0 occurrences
Blind adjudicationA scorer who can see which system produced an item900 items, no arm label, fields identical
Second instrumentAn adjudicator that is consistently and confidently wrongAgreement on 837 of 900, both label sets published
Title probeA zero that is really an absent paper43 of 43 returned from the comparison index, 37 at rank 1
Corpus probeA ranking failure counted as missing literature46 of 50 covered, the 4 misses kept in every denominator
Adjacent pairsClose fields carrying the difference between the armsRemoved: 18 of 43 against 1 and 1
Positive controlA judge that passes anything put in front of it8 historical source papers, 8 passed
Negative controlA judge that fails nothing11 pairings assembled to be wrong, 0 passed

The rules

Two things the engine is not allowed to do.

  • It cannot filter by field before it ranks

    An incoming problem is matched against all 113,394 papers in the index. Work from your own field and work from four fields away are ranked in the same pass and shown apart. The distance between them is reported and never used to exclude anything.

  • It cannot return a claim it cannot resolve

    A connection is traced back to the paper it came from by code that resolves the reference to a stored passage, with no model asked to confirm its own output. A link that does not resolve does not ship.

The engine in that table is the one Atlas runs on. Researchers join the waitlist at atlas.darksigmalabs.com.