The path the idea took, and the paths it left behind.
Every Ingest Paper report closes with a Trajectory Tree: from your paper's own hidden mechanism, we map the path the field actually walked and the latent branches that were in the idea but never explored — each with a concrete first move. Below is the tree from our real AlphaFold run.
AlphaFold (Jumper et al., Nature 2021)
What it did: predicted 3D protein structure from sequence by learning evolutionary constraints and refining toward a stable state with calibrated confidence.
Hidden mechanism: a reusable engine for inverse reconstruction — infer a hidden high-dimensional structure from a thin signal, borrowing constraints from a whole population and refining to a stable, confidence-scored state.
What the field actually did
The natural path deepened into more biological structure: protein complexes (AlphaFold-Multimer), ligands and nucleic acids (AlphaFold3), vast public structure databases, and downstream drug and enzyme discovery.
Why this was natural: the immediate prize — solving the rest of biology's structures — was enormous and sat directly in front of the field.
Confidence-first science
Latent question: what if the calibrated confidence (pLDDT), not the structure, had been the headline — a general way to build predictors that know what they don't know?
- Why not taken: the structures were the prize; calibration read as a feature, not a paradigm.
- Possible now: every high-stakes inverse model — climate, medicine, materials — needs trustworthy uncertainty.
- First move: lift the confidence head out as a standalone module and test it on a non-protein inverse problem.
Population-as-prior as a general method
Latent question: AlphaFold borrows strength from a whole evolutionary population to fold one example. What if that move were abstracted into a domain-independent method?
- Why not taken: coevolution felt biology-specific; the abstraction was never named on its own.
- Possible now: markets, languages, materials families — any population of related instances can borrow strength the same way.
- First move: formalize "population-conditioned inference" and run it on one non-biological reconstruction task.
Inverse design, not prediction
Latent question: AlphaFold predicts structure from sequence; the latent inverse is design — going from a desired structure back to the encoding that produces it.
- Why not taken: prediction had a clean, benchmarkable scoreboard (CASP); design did not.
- Possible now: diffusion and generative methods make inverse design tractable — the same engine becomes a design tool well beyond biology.
- First move: invert the pipeline on a toy system — specify a target geometry and ask the engine for an encoding that yields it.
Why this matters
The value is not claiming history was wrong. It is seeing that a paper usually contains more futures than the one the field selected — and that the discipline is built right into your report, generated from your own paper. Depth scales with tier: Entry and Standard include the historical path plus two to three latent branches; Lab and Frontier expand the full tree with deeper branch analysis and per-branch verification steps.