# Yotta Content

> Yotta Content builds original tasks, reinforcement-learning environments, verifiers, trajectories, repairs, and private holdouts to help frontier models hillclimb specific benchmark capabilities. It is part of the Ulam AI data ecosystem.

## Core pages

- [Homepage](https://yotta-content.vercel.app/): Product overview, production loop, principles, and benchmark catalog.
- [Benchmark catalog](https://yotta-content.vercel.app/benchmarks/): Ten benchmark-shaped data and RL environment families.
- [Machine-readable evidence](https://yotta-content.vercel.app/assets/evidence.json): Evidence status and supplied run metrics.
- [Expanded LLM guide](https://yotta-content.vercel.app/llms-full.txt): Detailed capability and evidence summary.

## Evidence-backed families

- [GDPval-AA-shaped professional work](https://yotta-content.vercel.app/benchmarks/gdpval-aa.html): Three judged 44-task GPT-5.6 Sol Pro blind-run bundles with artifact, rubric, trajectory, and integrity evidence.
- [Terminal-Bench 2.1-shaped environments](https://yotta-content.vercel.app/benchmarks/terminal-bench.html): Two judged five-task GPT-5.6 Sol Pro blind-run bundles with container tasks, hidden verifiers, and action/observation trajectories.
- [Humanity's Last Exam-shaped reasoning](https://yotta-content.vercel.app/benchmarks/humanitys-last-exam.html): Judged 50-task GPT-5.6 Sol Pro blind run across text and multimodal synthetic tasks.
- [GPQA Diamond-shaped science](https://yotta-content.vercel.app/benchmarks/gpqa-diamond.html): Judged 33-task GPT-5.6 Sol Pro blind run across biology, chemistry, and physics.
- [CritPt-shaped research physics](https://yotta-content.vercel.app/benchmarks/critpt.html): Judged 36-checkpoint GPT-5.6 Sol Pro multi-attempt blind run across 12 synthetic physics families.

## Additional benchmark-shaped programs

- [tau3-Banking-shaped environments](https://yotta-content.vercel.app/benchmarks/tau3-banking.html)
- [SciCode-shaped scientific programming](https://yotta-content.vercel.app/benchmarks/scicode.html)
- [AA-Omniscience-shaped factuality data](https://yotta-content.vercel.app/benchmarks/aa-omniscience.html)
- [AA-LCR-shaped long-context reasoning](https://yotta-content.vercel.app/benchmarks/aa-lcr.html)

## Ulam Math Data

- [Ulam Math overview](https://yotta-content.vercel.app/benchmarks/ulam-math.html): AIME-style exact-answer RLVR, graduate and research Math RL environments, and open research problems for rollout-based post-training.
- [AIME++ public sample](https://huggingface.co/datasets/ulamai/AIME-Plus-Plus)
- [Math RL Tasks public sample](https://huggingface.co/datasets/ulamai/Math-RL-Tasks)
- [SOTA-Math public sample](https://huggingface.co/datasets/ulamai/SOTA-Math)

## Important boundary

Yotta Content task families are independently created, benchmark-inspired systems. They are not official benchmark releases, replicas, affiliations, or leaderboard results. Quantitative claims are limited to supplied synthetic-run artifacts or clearly linked public samples.