We run real studies on the datasets we originate, from the GPU rental market and retail financial sentiment to prediction markets, code corpora, and physical AI. Methods in the open, results honest, null findings included.
Price formation in the markets behind AI. Cross-provider dispersion in the GPU rental market and the falling cost curve of hosted inference.
Turning alternative data into systematic signals. Retail financial sentiment at scale and prediction-market probabilities mapped to the securities they move.
Grounded corpora for coding agents and enterprise evals. Provenance-clean repositories, cross-tool activity graphs, and trained-model-to-silicon RTL generation.
Multi-modal data from real robot operation. Teleoperation, human demonstration, and a measured characterization of the sim-to-real gap.
A four-year cross-provider panel of GPU rental prices (6,216 observations, 38 providers, 32 SKUs) reveals extreme price dispersion - identical accelerators rent for an 18x range on H100, 96x on A100, and 100x on RTX-4090 - alongside a steady secular decline in consumer-GPU rates. Includes an honest null result on GPU prices as an equity signal under purged k-fold and shuffle-test validation.
Human-labeled bullish/bearish sentiment from the largest social-finance platform: 426M+ messages since 2009 across 50,000+ tickers, distilled into a point-in-time per-ticker score refreshed every five minutes. The ground-truth signal drives long/short strategies at 18.1% CAGR and 1.20 Sharpe on the NASDAQ-100, and beats an X/Twitter NLP-sentiment baseline 122% to 95% on cumulative return.
A cross-platform framework that maps 300+ active prediction-market events to 500+ affected securities across Polymarket, Kalshi, Limitless, and Metaculus. Each event links to 2-8 tickers with a signed exposure type and a 0-to-1 sensitivity score, drawn from point-in-time probability series. On 500+ resolved probability moves greater than 10%, prediction-market changes tend to lead the mapped security price moves by hours to days.
A preliminary cross-provider snapshot of hosted-inference token pricing across 488 observations from Together, Fireworks, DeepInfra, and Replicate. Among the cheapest available models, output price per million tokens fell from roughly $0.13 in mid-2024 into the $0.01-$0.03 range in 2025-2026. The sample mixes model tiers, so the direction of the decline is robust while absolute levels remain preliminary.
A reproducible pipeline that turns a trained logistic decoder into synthesizable Verilog RTL plus an auto-generated testbench, with weights quantized to 16-bit fixed-point. On an 8-ROI decoder it reaches 0.9781 eval accuracy and 0.9986 AUROC, loses less than 2% accuracy after quantization, and passes iverilog simulation against the floating-point reference.
A founder-owned, consented, and sanitized corpus of 1,400+ private repositories with 150,000+ commits, full diff history, pull requests, reviews, linked issues, CI runs, and release tags. Tickets are joined to the commits and PRs that resolved them, giving end-to-end issue-to-code traces for grounded coding-agent training and point-in-time SWE evaluation.
A data-availability report on a normalized, entity-resolved activity graph from 5 consented partner companies across 38 tools and 10 categories: 38.4M chat messages, 11.2M emails, 3.6M files, plus CRM, ticketing, and repository metadata, resolved across user, team, org, document, and task. Built as grounded substrate for agent training and workflow evals under a consent and PII-stripped-at-source model.
A comprehensive quantitative study of the simulation-to-reality gap using the Booster T1 humanoid robot. Analysis of 240 seconds of real-world data reveals leg joints exhibit 2× higher error than arms, with knee pitch joints showing 18.7° MAE. Provides actionable insights for domain randomization and sim2real transfer.
A standardized format for robot teleoperation data. Defines schemas for multi-modal observations, action spaces, sensor calibration, and coordinate frames. Designed for interoperability with LeRobot, RLDS, and Open X-Embodiment.
Every result ships with its method, and where we can, runnable code. If we claim it, we show our work.
We publish null results. A signal that does not survive validation is reported as one. We would rather kill a claim than sell false alpha.
Backtests run on point-in-time data, purged and shuffle-tested. No look-ahead, no survivorship, no leakage.
If you are training frontier models, running systematic strategies, or building embodied AI, and you need data that does not exist yet, we would like to hear from you.
RESEARCH@GERRA.COM