Olympiad-grade data creation * author math/physics/compsci problems and multi-step reasoning tasks (proofs, multi-hop, tot, chain-of-thought). * produce fully-worked solutions, hints, rubrics, and counter-examples; deliver in latex + json/parquet. Stepwise reasoning & annotation * decompose solutions into verifiable steps; tag dependencies, error types, and alternate paths. * build evaluator prompts and scoring functions for exact, rubric-based, and partial-credit grades. Model training & evaluation * curate instruction-tuning sets and reasoning curricula; run small/medium scale fine-tunes. * design benchmarks and leaderboards with reproducible scoring, leakage tests, and stats reports. Cheat-resistant talent vetting (for existing contributor pools) * create **secure, variant-rich assessments** to evaluate math/logic ability and proof writing. * techniques we use: parallel forms & dynamic variants (same construct, different instances; per-candidate seeding). * hidden gold & cross-item invariants to detect copy/paste or tool-overreliance. * time-to-solve and process evidence (scratch reasoning, step logs, latex diffs) reviewed by experts. * style & consistency checks (statistical signals + human review) to flag likely ai-generated text. * closed reference sets and controlled tool policies; honor-tests that require original derivations. * output: ranked shortlists, skills profiles, red-flags, and targeted upskilling plans. Red-teaming & safety * adversarial prompts for math/logic hallucinations, spec violations, and content safety; failure taxonomies and fixes.