Capabilities reference

A brain is built around one or more capabilities — plain-English descriptions of what the brain can do. The wizard’s auto-resolver picks the smallest set of priors that covers every capability you selected, preferring paper-backed priors and specialists.

Live capabilities

Capability	Icon	What it delivers	Paper-backed priors
Predict a number	🔢	Given a few examples, predict a numeric value for a new input	pfns-reference (Müller 2022)
Forecast forward	📈	Given the past of a series, predict where it is heading	tabpfn-ts (Hoo 2024), ifbo (Rakotoarison 2024)
Pick the next experiment	🧪	Given results so far, suggest which configuration to try next	pfns4bo (Müller 2023), ifbo (Rakotoarison 2024)
Extrapolate a partial curve	📉	Given the start of a curve, predict where it will end up	lc-pfn (Adriaensen 2023), ifbo (Rakotoarison 2024)
Show confidence	🎯	Output not just a prediction but how sure the brain is	pfns-reference, lc-pfn, ifbo, pfns4bo

A capability is live if at least one paper-backed, non-niche prior in the catalog delivers it. The wizard derives state from the catalog — adding a new paper-backed prior with the right tag automatically flips a capability live.

Coming soon

Capability	Icon	Why it’s not live yet
Sort into categories	🏷️	Needs TabPFN (Hollmann 2022/2025) integration. A two-moons demo prior exists but isn’t paper-backed.

Coming-soon cards are disabled in the picker with a ⏳ badge naming the missing study. The card still shows the icon + description so users can see what’s on the roadmap.

How the resolver picks priors

For each contract group (priors sharing a tensor shape), the resolver:

Enumerates every minimum-size cover — every subset that covers all selected capabilities
Across all candidates, sorts by:
- Fewer priors (smaller cover wins)
- More paper-backed priors (paper-backed beats demo)
- Fewer bonus capabilities (specialist beats generalist)
- Lexicographic tie-break

So:

You pick	Resolver returns	Why
Forecast forward	tabpfn-ts (covers forecast + confidence)	Paper-backed, 1 bonus capability (vs ifbo’s 3)
Forecast forward + Pick next experiment	ifbo	Single prior covers both; ifbo is the only paper-backed match
Predict a number + Show confidence	pfns-reference	Canonical PFN reference, covers both
Forecast + Extrapolate a curve	ifbo	Single prior covers both; lc-pfn only covers extrapolate
Predict a number + Forecast forward	refused at picker	Different specialities — cross-contract mix isn’t possible

Mix rules — the contract constraint

Mixture training requires every chosen prior to share a tensor contract (input + output shape). The wizard tracks four contracts:

regression_1d — X (N, 1) + y (N,)
regression_vard — X (N, D) + y (N,) (D varies per task)
classification — X (N, F) + labels
temporal_tf — X (T, F) + temporal y

The picker greys capabilities that would force a different contract. Picking one anyway replaces the current selection (clarified via the tooltip).

Specialised brains (niche priors)

Some priors are tied to a specific domain and don’t fit a general capability checkbox. KinPFN (RNA folding kinetics) is the only one in the v1 catalog. Niche priors:

Have niche: true on their wrapper
Don’t contribute to capability state (a niche prior alone doesn’t make a capability “live”)
Don’t appear in the main picker grid

A Specialised brains section beneath the capability grid surfaces them when applicable.

Paper-pinned reproductions

If you pick exactly one paper-backed prior and no others, the wizard switches to a paper-pinned route — model architecture, hyperparameters, eval suite, and prior overrides all come from the paper’s own seedRun, not from the wizard’s recipe sliders. See Paper reproductions for the details and trade-offs.

How a coming-soon capability flips live

A capability becomes live the moment a paper-backed, non-niche prior in the catalog provides it. The wizard, the brains-list chips, and the brain-page hero all update automatically — no separate switch to flip.