How the engine works.
What the engine indexes instead of 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
- 10 / 43
- times it handed back the paper the answer came from
- 0 / 47
- times the comparison 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. No search in 1992 would have put those two papers in front of each other, and no search today does either.
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 and colleagues 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
Ten times it returned the paper the breakthrough was built on. The comparison never did.
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 handed that to Atlas and to a leading AI research search tool indexing more than 200 million papers.
The commercial index holds those source papers. The same run proves it, by returning the bridge paper on 31 of the 50 cases, so the zero is not a question of coverage. It has them and does not reach them from the problem. We tested the obvious shortcut too. Take our neutral restatement of the problem, send it to that same index, and it pulls 190 papers out of the researcher’s own literature against our 192, and none of them are the paper. The index still stores every paper in the words its own field wrote it in, so a neutral question gets compared against literal text and matches on whatever wording happens to coincide.
The last row has two denominators because the source paper had to be identified and verified against a public catalogue before anything could be asked of it. Forty-seven of the fifty were identified; forty-three of those carried an abstract. Seven of our ten came back inside the top twenty, where a person sees them without being told to look. Nine hundred of the returned papers were read blind. The eight historical papers the cases were built from score 8 out of 8, and eleven pairings assembled to be wrong score 0 out of 11.
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 everything held. Work from your own field and work from four fields away are ranked in the same pass and shown apart, with the distance reported rather than 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, not by asking a model whether it was telling the truth. A link that does not resolve to a real passage does not ship.