GreenDynamics · Pilot samples · November 2026

Nature as the Verifier

Physically verified environments and expert data that improve the scientific discovery capability of frontier models: planning experiments, reading real measurements and converging on new results in chemistry and materials.

Examples

RL environment · Echo and Live tiers · GD Byte Infra

Rainbow

full-score regionunsafe zoneself-improving search pathdrag to rotate
  • Complex spacemany peaks and ridges; most recipes score low
  • Self-improving searchthe agent climbs, jumps and converges on the best peak
  • Safety constraintsunsafe mixes are scattered across the space; entering one costs the run
  • Many ways to full scorethe top region is an area, not a single recipe
Illustrative model of the search space, not measured data.

Formulate a target colour from dye stocks and deionised water. Every task is grounded in real lab measurements, checked by a separate verifier, and logged step by step.

Tasks
20–50 target colours
Scoring
reach the colour, then safety and efficiency decide the score
Runs
hosted on GD Byte Infra; you connect over a secure API
Expert evaluation data

Characterisation evaluation set

Example item image: a powder X-ray diffraction pattern with five peaks between 38 and 83 degrees two-theta.

Scientist-written questions on diffraction patterns, spectra and microscopy images. Each item has a verifiable answer, so it can be scored automatically for evaluation or used as a reward for training.

Items
double-reviewed, with a held-out set we keep
Evidence
instrument data plus rendered figures
Scoring
exact match, numeric tolerance, or weighted rubric
Harness

How Rainbow plugs into your harness

Choose a tier. The task files, tool interface and verifier are the same in each; only the world behind the API changes.

your sidevisible to the agentworldhidden from the agent

Echo · hosted simulation. Your runner, agent and trainer stay on your side. The environment, world model and verifier run on GD Byte Infra, reached through a secure session API. The agent sees the instruction and tool results only; the verifier runs in a separate sandbox after the episode. Seconds per step, for training at scale.

YOUR HARNESSGD BYTE INFRA · ECHO · hosted by GreenDynamics, reached over a secure APIRunneropens a session over theAPI; task.json sets limitsAgentyour model + scaffold;sees the instruction andtool results onlyTrainer / eval metricsreceives the reward foreach episodetask.jsontimeouts · budget · tierinstruction.mdtarget colour · budget · rulesenvironment/on GD Byte InfraWorld · Echo tiergrounded in real labmeasurements(swapped per tier)Toolspropose_recipe()query_history()Step logevery action, observation and timestamp,recorded for the verifierverifier/separate sandboxrubric.json · score.pyscores the whole path:colour gate, safety, stepswrites reward.jsonhiddenwitnesses/paths proven to scorefull; used to test theverifier, never shownto the agenthiddenlimitsopen sessionpromptaction → tool callobservationtrajectoryreward.json
Evaluation data

Example item

A diffraction question in the format our evaluation sets use. Reading a plot looks easy; ranking peaks correctly is where multimodal models still fail.

Example · XRD relative intensity · multiple choice
Powder X-ray diffraction pattern: peaks at about 38.6, 44.9, 65.3, 78.6 and 82.7 degrees two-theta with decreasing heights 100, 48, 31, 36 and 11.

What is the correct order of peak positions (2θ) from lowest to highest intensity in the XRD pattern shown?

  • A44.9, 38.6, 78.6
  • B44.9, 78.6, 38.6
  • C78.6, 38.6, 44.9
  • D ✓78.6, 44.9, 38.6

Why it is hard: models find the tallest peak easily but often misorder the smaller ones, scoring close to random on this task type.

Evidence

What comes with each pilot

The four things your team will check first, and the file that answers each.

It runs

One command connects to the hosted endpoint. Latency, uptime and concurrent sessions are measured and reported.

OPS_REPORT.md

It has headroom

Frontier models score 20–60 out of 100: plenty of room to improve, and enough signal to learn from.

02060100baselines.json

It resists gaming

Tested against known reward-gaming strategies before delivery. Results shared under NDA.

ANTI_HACKING.md · under NDA

Rights are clean

All measurements produced in our own lab. Contributors consented to commercial and AI-training use.

RIGHTS.md
Next steps

From sample to delivery

IntroThis overview

Examples, harness and evidence.

Scope callAgree the pilot

Environments, data, volume, tier and acceptance criteria.

ContractPilot agreement

Access to environments and data starts once the agreement is signed.

Pilot and deliveryRun, then scale

Run your models under the agreement, then expand to full packs.

Request a pilotTell us which environments and data you are interested in. We reply with a scope call slot.
future@greendynamics.ai