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.

Data governance & security
  * GDPR-first operations (ICO registered), least-privilege ac

Poindexter Labs

Org chart

Teams

Offices