Poll — don't wait for pushes. The discovery endpoint is rate-limited, cursor-based, and tag-filtered:
GET /v1/problems?tags=compchem&state=open&since=<cursor>&limit=50
Authorization: Bearer msk_...
# response items are event-shaped, so future webhooks/SSE carry the same schema:
{ "event": "problem.opened", "seq": "17423:ab41",
"data": { "id": "...", "tags": ["compchem","defect"], "mode": "relax",
"quality_floor": {"f_tol_ev_A": "...", "allowMoreAccurate": true},
"reserve_price_usd": 65.0, "deposit_usd": 0, "closes_at": "...",
"frozen_spec_sha256": "...", "protocol": "1.x.y" } }
Everything you need to decide is in the item — one request, no follow-ups. A 30–60 s poll interval is plenty; deadlines are hours to days.
At launch, problems are fixed bounties: solve if your cost is below the bounty; first verified solution wins. When sealed-bid auctions enable for a problem class, the mechanism is second-price (Vickrey), and your strategy is one line:
Claims are truth-checked: in scoring mode, a verified result outside your claimed band fails even if it is better than claimed. Update your claim; don't exceed it.
Your margin is your preconditioner. Warm density-matrix starts, MLIP pre-relaxation, learned charge-density guesses — all fair game, and encouraged. The platform's own measurement: a converged-DM warm start reached a density residual of 1.3×10⁻⁴ in one SCF step where a cold start sat at 1.177 — four orders of magnitude at identical cost. Two rules keep it honest:
Setup guides for the reference solver environments (conda-forge dual-code container; CUDA SIESTA notes for GPU nodes; MPI pinning) land here at launch — including cheap-GPU recipes, since verified-then-paid workloads tolerate interruptible hardware.
From ASE (see tutorials) or the raw API. You pick engine + fidelity (or exact parameters); the platform freezes the tuple, estimates cost, and your card is charged only when a solution verifies. Unfilled problems cost nothing.
A passing verification certifies: forces below the frozen threshold and electronic self-consistency of the declared character, under a named, digest-pinned protocol — recomputable by anyone from the public record. It does not certify a global minimum, and quantities without a checkable witness (free energies, ensemble averages) are never labeled verified.
OracleResult schema — portable fields in eV/Å, code-native diagnostics kept native and code-tagged, the raw log + sha256pseudodojo-sr / pseudodojo-fr) and uploading your own (POST /v1/pseudos)