GERRA
Gerra / Research indexOpen record
Research, methods, and source material

Evidence has
a paper trail.

Studies, technical notes, field research, and the underlying datasets in one public record. We publish what the data supports, what it does not, and enough of the trail to inspect the difference.

Research works
09
Source datasets
19
Field notes
17
02 / Papers

Research register

Papers, specifications, and technical notes. Ordered by release, with the measured substrate shown beside each claim.

  1. R-08The GPU Rental Market: Price Dispersion and the Cost Curve of ComputeA four-year, cross-provider panel measuring the unusually wide price dispersion of identical accelerators and testing whether GPU prices contain an equity signal.Study panel: 6,216 observations / 38 providersResearch paper/PreprintMay 2026
  2. R-07Retail Financial Sentiment as a Systematic SignalEighteen years of human-labeled bullish and bearish messages tested as a point-in-time systematic signal, including a direct comparison with NLP sentiment from X.426M+ messages / 18 yearsResearch paper/PublishedMay 2026
  3. R-06Mapping Prediction Markets to SecuritiesA cross-platform framework that maps live event probabilities to exposed securities with signed direction, sensitivity, and point-in-time probability history.300+ events / 500+ securitiesResearch paper/PublishedMay 2026
  4. R-05The Falling Price of InferenceA preliminary cross-provider study of hosted-model token prices, separating the clear directional decline in cost from the limits of a mixed-model sample.488 provider-model observationsResearch paper/PreprintMay 2026
  5. R-04From Trained Model to SiliconA reproducible pipeline from a trained logistic decoder to synthesizable Verilog RTL and an auto-generated testbench, measured before and after fixed-point quantization.0.9781 accuracy / simulation passedResearch paper/PublishedMay 2026
  6. R-03Operational Telemetry: A Cross-Tool Activity GraphA data-availability report on consented company activity normalized across communication, documents, CRM, tickets, repositories, people, teams, and tasks.5 companies / 38 toolsTechnical note/Working draftMay 2026
  7. R-02Empirical Characterization of the Sim2Real GapA joint-level comparison of simulated and real trajectories on a full-size bipedal humanoid, isolating where model fidelity breaks down across the body.23 joints / measured transfer gapResearch paper/PublishedDec 2024
  8. R-01Open Robot Training Format v1.0A common representation for robot demonstrations across multimodal observations, action spaces, sensor calibration, coordinate frames, and embodiments.Interoperable training-data schemaSpecification/Working draftDec 2024
03 / Field notes

The archive stays live.

Longer investigations sit beside the formal papers: component histories, market structure, deployment economics, and observations that are useful before they become a specification.

AllResearchCodeAI TrainingMethodData QualityGuideMarketsEvaluationRSS
  1. 01Why Identical GPUs Rent at Wildly Different PricesThe same H100 rents for $1.20 to $22.02 per hour. What a 38-provider panel shows about GPU price dispersion, and why the equity signal is a null.Research explainerMay 12, 2026
04 / Source data

The dataset is part of the argument.

Research is linked to the data products it can actually support. Each catalog record carries its collection scope, history, delivery form, and technical detail forward.

A

Markets & signals

Point-in-time data for compute, capital, attention, and event-driven research.

  1. 01Retail Investor SentimentExclusive retail investor sentiment from the largest social finance platform.426M MESSAGES · SINCE 2009
  2. 02Sports Query & Answer LogsThe only sports data source that powers real-time AI tool calls for the largest language models.1BN+ QUERIES · 12 YRS
  3. 03Startup TechnographicsPrivate company technographics mapped to public market tickers. See what startups adopt before the market prices it in.10M+ COMPANIES
  4. 04Federal Software ContractsTicker-mapped US federal government spending on enterprise software. See which vendors win before the street does.120+ TICKERS MAPPED
  5. 05Prediction Market ProbabilitiesCross-platform prediction market events mapped to affected securities with real-time probability streams.500+ SECURITIES
  6. 06GPU Rental PricesHourly GPU rental prices across every major cloud, normalized to canonical SKUs. A leading indicator of AI capex before it reaches earnings.262,146 OBSERVATIONS · 40 PROVIDERS
  7. 07Inference Token PricesReal-time inference economics across hosted-model providers - token prices, throughput, and latency as a demand-side read on AI compute.2,397 QUOTES · 381 MODELS
B

Software & organizations

Repository and operating histories for agents that need to reason across real work.

  1. 01Cross-Tool Workplace RecordsCross-tool operational data from collaboration systems. Entity-resolved into a unified schema for AI training and workflow evals.40+ EVENT TYPES
  2. 02Engineering Delivery RecordsEngineering coordination data across source control, issue tracking, and data platforms. Unified SDLC schema for coding agents and SWE evals.50+ SDLC EVENTS
  3. 03Full-History CodebasesClone and inspect 178 real-world repositories as complete engineering systems, with 15,238 unique commits preserved for training, reasoning, and evaluation.15,238 COMMITS
  4. 04Whole-Company Operating ArchivesComplete operating histories of real companies - code, business data, communications, documents, and databases - as a training corpus for frontier models.WHOLE-COMPANY DEPTH
C

Agent tasks & RL environments

Runnable tasks with machine graders, where the reward signal is verified rather than asserted.

  1. 01Repo Build & Repair TasksNatural-language specifications and real issues turned into runnable repositories with held-out test suites, each verified to fail before the fix and pass after it.2,400 SPEC TASKS · 24K REPO TASKS
  2. 02Exploit & Patch TasksContainerized vulnerable services with working exploits, patches, and a machine-gradeable subtask ladder - multi-stage chains, not single-bug toys.4,000 CTF TASKS
  3. 03Full-App Build TasksNon-technical prose specifications for whole web applications, graded by a browser driving the running app rather than by reading the diff.2,400 APP SPECS · 3,200 WEB TASKS
  4. 04Back-Office Work TasksReal de-identified business operations turned into RL instances with exact-value oracles - the accounting close as an agent task, graded to the cent.16,000 RL INSTANCES · 24K TRAJECTORIES
  5. 05ML Research TasksGPU-backed environments where the task is to run machine-learning research: win the competition, replicate the paper, or beat the base model.400 COMPETITION RUNS · 200 REPLICATIONS
D

Embodied systems

Demonstration and sensor data grounded in bodies, environments, and time.

  1. 01Robot Teleoperation DemonstrationsSuccess-labeled robot manipulation episodes collected via human teleoperation across diverse tasks and embodiments.400K+ EPISODES
  2. 02First-Person Human MotionFirst-person human video paired with full-body 3D motion capture - the human-demonstration layer for embodied pretraining.POV + 3D MOCAP
  3. 03Robot Sensor RecordingsHigh-frequency proprioception, inertial, and audio streams with sub-millisecond synchronization across embodiments.<1MS SYNC
Open the complete data catalog
05 / Evidence standard

Claims should survive contact with the trail.

Our job is not to make every dataset look predictive. It is to establish what was observed, test the strongest interpretation, and leave the boundary visible when the result is weaker than the thesis.

  1. 01

    Start at the source

    Origin, collection context, licensing, and temporal coverage stay attached to the data.

  2. 02

    Preserve the trail

    Manifests, timestamps, mappings, transformations, and checks make the result inspectable.

  3. 03

    Try to break the claim

    Point-in-time tests, holdouts, permutation checks, and leakage review come before the headline.

  4. 04

    Publish the boundary

    Null results, sample limits, exclusions, and unresolved uncertainty belong in the conclusion.

Evidence currently in the record
426M+human-labeled messagesPoint-in-time retail sentiment15,238unique commitsCloneable full-history repositories262,146price observationsCross-provider GPU panel23humanoid jointsMeasured simulation transfer

Open collaboration

Have a hard data problem, a result to reproduce, or a corpus that should exist?

research@gerra.com