What we learned

Turning 6,700 papers into structured protocols — and where the machines break.

An experiment in extracting reproducible organoid culture recipes from the open-access literature. The short version: a cheap local model can find the ingredients, but only a frontier model can reconstruct the procedure — and only if you fact-check it against the source. Doing that well is possible. Doing it at scale is expensive. Here's the evidence.

6,725
papers ingested
14,794
full-text methods parsed
289
verified procedure steps
(10 gold protocols)
438
verbatim evidence quotes

The goal

Every organoid paper buries a recipe in its Methods section: which cells, which signaling factors, at what concentration, in what order, for how many days. That knowledge is locked in prose, one paper at a time. The premise was simple — parse the methods, extract the recipe into a structured schema, ground every reagent to an ontology, and make the whole corpus searchable and comparable. Then imagine doing it across all ~154M papers in the scientific literature.

We built the pipeline: automated discovery (OpenAlex, Europe PMC), full-text retrieval as structured JATS XML (no PDFs, no OCR), local-model extraction, and entity grounding against SRI Node Normalizer + Cellosaurus. The corpus is real and it works. The interesting part is where it stopped working.

What worked, and what broke

Same local model (gemma 3, 12B, running on one GPU). Two very different tasks.

Finding the ingredients reliable

Pulling reagents, concentrations, and units — and grounding them to ontology IDs with a verbatim quote for each. This is essentially named-entity recognition. The local model does it well, cheaply, at corpus scale.

100% grounded

Reconstructing the procedure fails

Segmenting the method into ordered stages and assigning each reagent to the step it belongs in. This needs multi-step reasoning over long context — and the 12B model can't hold it together.

59% of stages came back empty

Across 1,470 papers the local model tried to break into stages, the failure was systematic — not an occasional miss:

Stages returned with no reagents ("nothing added in this step")59%
Papers flattened to just ≤2 stages (a thin skeleton)83%
Reagents the model found but never placed in any step1,471

It produced things that looked like recipes — ordered stages, tidy cards — while quietly dropping the reagents it had already found and mischaracterizing what the paper actually did. A confident, hollow skeleton. That's worse than no recipe, because it looks trustworthy.

The pivot: small and perfect

Instead of extracting 6,000 papers badly, we extracted 10 landmark protocols perfectly — one per organoid type (cerebral, lung, kidney, intestinal, cardiac, retinal, and more), all open-license, all highly cited. Three ingredients made it work:

1
A schema that models honesty, not just data

Every value carries its provenance: stated (with a verbatim quote), deferred to a citation (the recipe lives in another paper), or not reported (the authors left it out). Plus step-level actions with real parameters — temperatures, spin speeds, durations, densities — not just a reagent list.

2
A frontier model for the reasoning

The same task that hollowed out under the 12B model — read the whole procedure, segment it, place every reagent, catch what's deferred to a citation — a frontier model handles. It even flagged suspicious values ("10 mg/cm² collagen — likely a µg typo, verify") on its own.

3
An independent, cross-model fact-check pass

A different frontier model (GPT-5.5) graded every extraction against the source, quote by quote, and the extractor fixed what it flagged — looping until it passed. Cross-model matters: a model checking its own work shares its own blind spots.

What the fact-checker caught

Real problems the cross-model loop surfaced and fixed — the reason the pass is not optional.

!
Fabricated details. Claims like "nine serial passages" and "PCR + chromatogram analysis" that were never in the paper. — kidney protocol
!
Contradictions with the source. The extraction said a protein was required for a structure the paper explicitly showed forming fine without it. — kidney; and a liver metabolism claim reversed from the source
!
Metadata masquerading as fact. DOIs, licenses, and author lists marked as "stated in the paper" when they come from database metadata, not the text — a subtle, systematic honesty bug a same-model check would have shared and missed.
!
Over-claims. Multi-part values trimmed back to exactly what their cited quote supports — no more.

4 of 10 protocols passed the independent grader outright; the rest had their fixes applied and await a final human sign-off. The honest headline isn't "100% automated" — it's "the machine and the checker together get you to a place a human can actually trust, fast."

The catch: it gets expensive

Those 10 perfect protocols cost ~1.4 million tokens across 20 model agents — extraction plus the fact-check-and-fix loop. That's excellent value for 10 gold-standard records. It is a very different number at the scale the field actually needs.

10 papers — done~1.4M tokens
The literature — ~154,000,000 papersextrapolates to trillions

So the real architecture is a hierarchy of cost: cheap local models to sweep the whole corpus for ingredients and grounding; frontier models reserved for the reasoning-heavy recipe reconstruction on the papers that matter; and an independent fact-check pass as the quality gate. You don't run the expensive step on everything — you run it where it earns its cost.

The lesson

“You need a solid schema and frontier LLMs, then fact-check passes. It works — and it gets expensive. The craft is knowing where to spend.”

The one-sentence takeaway from the whole experiment.

Cheap models are for breadth. Frontier models are for judgment. Fact-checking is not a finishing touch — it's the step that makes the output trustworthy enough to matter. And a rigorous schema is what lets all three cooperate. None of the three is optional; the interesting engineering is in composing them affordably.

The full interactive atlas — all 5,538 protocols, searchable — runs on its own host.

Explore the full interactive atlas →