An AI-native discovery engine

Findings the published baselines assumed, measured.

Truffle Discovery Lab runs a closed loop. It takes an old or underused dataset, asks one falsifiable question, isolates one variable, and measures how much of a claimed result is really there. Each turn is a citable, downloadable paper. Negative results count.

Request a dataset Read the loop

2 discoveries published · curl truffle.help/api/discoveries.json

scroll to read the loop

the loop

One dataset, one variable, one falsifiable answer per turn.

Dataset

Start from an old or underused dataset that nobody re-examined.

Question

Pull one falsifiable claim out of the published baseline.

Isolate

Hold everything fixed but the single variable in question.

Measure

Report what is really there. A null result is a result.

latest finding

The most recent turn of the engine.

No. 008Surgical video2026-06-21

A zero-pixel clock recovers most of the surgical-phase signal on Cholec80

A predictor that sees zero pixels and knows only where a frame sits in time reaches 70.9% video-level phase accuracy, closing 62% of the gap between the majority floor and the published vision ceiling.

Read the paper

70.9%

phase accuracy with zero pixels (matched splits)

Horizontal bar chart comparing a majority floor, a clock-only predictor on two splits, and five published vision models on Cholec80 phase accuracy. — The clock alone (matched split) lands at 70.9%, far above the 39.6% majority floor and within reach of PhaseNet's published 78.8%. None of the published vision models see less than the clock does.

the archive

Every turn, citable and downloadable.

No. 009published

On ESOL, a fused-descriptor MLP matches a graph neural network at a quarter of the GPU time

A small MLP on fused RDKit descriptors and count fingerprints matches Chemprop's D-MPNN on both scaffold and random ESOL splits, with overlapping 95% confidence intervals, while using 3.6 to 4.2 times less GPU wall-clock; a gradient-boosted tree on the same features sits at the compute-accuracy frontier, ahead of both.

Molecular property predictionESOL (Delaney aqueous solubility)2026-06-21

4.2xless GPU wall-clock at D-MPNN-equivalent accuracy (scaffold split)

interfaces

One record, read by humans, agents, and MCP clients.

The lab is an API before it is a website

humans / agents / mcp

API

Every discovery and the live engine status as JSON, at /api/discoveries.json and /api/status.json.

CLI

The assay CLI is the source of truth. The web and MCP surfaces project it; they never own logic the CLI lacks.

MCP

Point an MCP client at the lab to read the archive and request a dataset from inside your own agent.

curl truffle.help/api/discoveries.json

Have a dataset that deserves a second look?

Request a dataset See what is in flight