# Propose untested catalysts for the OCx25 pipelines

The two synthesis routes behind this dataset (VSParticle spark ablation of
metal rods; University of Toronto robotic chemical reduction of metal
salts) run about 30 experiments a day and are still running. A model is
only worth something to them if it names compositions they have not made
and is right about them. Name them.

## The task

Propose up to 20 compositions, ranked, that are **not** in `experiments`
(no target or XRF composition within 5 at. % of yours in the same element
system), for HER or CO₂RR, that one of the two routes can plausibly make
(elements with a metal rod available for spark ablation, or a soluble
precursor reducible by hydrazine; up to four elements), and for each give:

- the predicted outcome in the dataset's own terms (HER cathode potential
  vs SHE at 50 mA/cm²; CO₂RR Faradaic efficiencies or production rates at
  3.3 V) with an uncertainty and how it was obtained;
- the rationale: which model or physical argument, which training points
  it leans on, and the nearest tested compositions and their measured
  outcomes;
- one falsifiable prediction in a sentence the lab's standard test would
  confirm or refute (for example "FE(CO) at 3.3 V exceeds 60 % and FE(H₂)
  stays below 25 %");
- cost or availability notes where they matter (precious-metal content,
  toxic elements such as Hg, Tl, Cd).

Base the predictions on a model whose LOCO `pearson_r2` you report in
`models_used` with a reference to where it is documented (a TfP Attempt,
Artifact or Note in this Campaign, or a public URL). A model that has not
been scored under the Campaign protocol is a guess; say so by reporting
its score as 0.

Favour candidates that would teach the lab something either way: a
predicted good HER catalyst without platinum-group metals; a Cu-free
composition predicted to make C₂H₄; a composition where two models
disagree. Avoid trivial neighbours of the best tested points.

## What to submit

- `document`: JSON satisfying `document_schema`.
- `report`: Markdown, under 3,000 words, written for the experimental team:
  the shortlist as a table, the reasoning, what each result would mean,
  and the order you would run them in.
- `code`: an archive with the model, the candidate generation and the
  nearest-neighbour check against `experiments`.

## How it is judged

Schema check, then a campaign reviewer from the team. Verified when the
candidates are genuinely untested, synthesisable by the stated route, the
predictions are traceable to a model scored under the Campaign protocol,
and the reviewer would put at least five of them in the queue. Whether
the lab then makes and tests them is their decision and happens outside
TfP; results, if any, are the best possible Note on this Problem. Bound
inputs: `experiments`, `xrf`, `dashboard`, `document_schema`.