# Neo Norma
# https://www.neonorma.si/

Neo Norma builds RL evaluation tasks and verifiable Harbor/Docker environments for science.
Each environment ships with a deterministic Python Grader that runs in under 50ms and with Preflight verification.
There is no LLM judge in the Grader.

## What an environment contains
- Harbor, Docker, and Gym harness targets
- A containerized task with de-identified clinical, wet-lab, or scientific tables
- A deterministic Python Grader
- A Preflight verification report (reference solver 1.0000, shortcut traps capped at 0.42, schema checks, zero web contamination)
- A de-identification and provenance manifest

## Public benchmark rewards
These are the estimated rewards shown on https://www.neonorma.si/ for the four held-out tasks.
Reward 1.0000 is the reference solver. Means below are the average of the four public task rewards.

Task | Fable 5.1 | Claude Opus 5.5 | GPT-6.1 Sol | GPT-6 Astra | Reference solver
Plasma cfDNA liquid biopsy and aging-blood filtering | 0.7800 | 0.7000 | 0.5000 | 0.4500 | 1.0000
Molecular Residual Disease (MRD) Duplex Sequencing | 0.7300 | 0.6500 | 0.4500 | 0.2000 | 1.0000
CRISPR Perturb-Seq and Drug Combination Synergy | 0.7800 | 0.7000 | 0.4500 | 0.7000 | 1.0000
Oncology Trial Biostatistics (Competing Risks and RMST) | 0.7300 | 0.6500 | 0.5000 | 0.7500 | 1.0000

Mean | Fable 5.1 0.7550 (75.5%) | Claude Opus 5.5 0.6750 (67.5%) | GPT-6.1 Sol 0.4750 (47.5%) | GPT-6 Astra 0.5250 (52.5%) | Reference solver 1.0000

## Container datacards
NN-ENV-01  Plasma cfDNA liquid biopsy and aging-blood filtering
NN-ENV-02  Molecular Residual Disease (MRD) Duplex Sequencing
NN-ENV-03  CRISPR Perturb-Seq and Drug Combination Synergy
NN-ENV-04  Oncology Trial Biostatistics (Competing Risks and RMST)

Each datacard lists harnesses (Harbor, Docker, Gym), Grader runtime, Preflight verification status, and this container tree:
- instruction.md: task prompt and clinical or assay objective
- task.toml: Harbor runner config, resource limits, and timeout
- environment/: Docker image and task tables (counts and row volume only)
- verification/: deterministic Python Grader and Preflight verification suite
- provenance/: at-source de-identification report and chain-of-title manifest

Exact schemas and container artifacts ship with the evaluation sample under mutual NDA.

## How to make contact
AI lab post-training researchers: request a 5-task evaluation sample.
Clinical and biotech data holders: offer an inventory count. Do not send raw files in a first note.
Investors: ask for the environment refinery and licensing model.

Email: buyan@neonorma.si
Book: https://calendly.com/buyan-neonorma/30min
Site: https://www.neonorma.si/
Agent instructions: https://www.neonorma.si/for-agents.md