One codebase.
Three languages.
Zero drift.

PolyKode translates bidirectionally across SAS, R, and Python. Validated by execution parity, not just syntax. Built for Pharma, BFSI, and Manufacturing, where the translation has to run the same.

Try CommunityBook Enterprise demo

Community
Apache 2.0

Execution
parity

SAS · R ·
Python

PROBLEM

A compile is not a submission.

In Pharma, BFSI, and Manufacturing, translated code has to run the same.

Syntax-only translation

Generic models translate strings. They do not prove the output matches the original numbers.

No numerical contract

Clinical and actuarial work needs SAS-exact numbers, including the RNG stream. Compiling is not enough.

No audit trail

Regulated teams need provenance and a divergence record. Working code alone does not pass review.

THE PRODUCT

Not a wrapper. A regulated-industry translation engine.

Generic code models translate syntax. PolyKode is fine-tuned on real, license-tracked SAS, R, and Python pairs from Pharma, BFSI, and Manufacturing. Then it is rewarded on whether the output matches. Every translation ships with a parity check.

See the editions

Bidirectional by design

SAS to R, SAS to Python, R to Python. One model, six pairs. Trained on matched implementations from regulated code, not scraped from the open web.

  • SAS, R, and Python in any direction
  • Matched pairs from regulated work
  • License-tracked training intake
  • Coverage you can hold as an inventory

Execution parity, not string similarity

The reward function scores translations on row, column, value, formatting, and language-behavior match. The MT19937 stream is reproduced so SAS, R, and Python agree numerically.

  • Row, column, and value checks
  • Formatting preserved
  • SAS-exact numerical contract
  • Documented tolerance everywhere else

Audit trail every translation

Each output ships with a divergence registry, provenance for training data, and a compliance-ready record. The artifact regulated teams need, not just working code.

  • Divergence registry with annotations
  • Provenance for training data
  • Signed, audit-ready reports
  • Correct-but-different behaviors kept

Runs where your data lives

Air-gapped ZeroBoxx on-prem, sovereign cloud, private Hugging Face endpoint, or bring your own inference. No code leaves your environment.

  • ZeroBoxx on-prem and air-gapped
  • Sovereign cloud tenancy
  • Private Hugging Face endpoint
  • Bring your own OpenAI-compatible runtime

TWO EDITIONS, ONE ENGINE

Start free. Certify when you are ready to ship.

Community is the adoption surface. A full Qwen3-32B checkpoint under Apache 2.0. Enterprise adds the parity harness, divergence registry, certified runtime, and the regulated-delivery commitments that make it viable for production.

Community

FreeApache 2.0

For developers, students, and evaluators. Full model weights, quickstart, and a public benchmark subset.

  • Qwen3-32B baseline checkpoint, complete weights
  • Local, Hugging Face endpoint, or your own runtime
  • Public SAS-Parity-Bench subset
  • Syntax and basic runtime checks
  • Full PolyKode VS Code extension, BYO endpoint
  • GitHub issues, community forum, docs
Pull from Hugging Face

Enterprise

Contactper environment

For regulated production. Clinical, BFSI, manufacturing. Certified runtime, full parity harness, audit-grade delivery.

  • Euclid 2.8.1 (30B). Kimi K2 target. 120B on the roadmap.
  • ZeroBoxx On-Prem, sovereign cloud, air-gapped, private H200
  • Full curated 26,269-pair corpus. 198 constructs.
  • Row, col, value, formatting, and SAS-exact numerical contract
  • Divergence registry. Certified on licensed SAS 9.4 and Viya.
  • Model Care, 24/7 SLA, dedicated success
ZeroBoxx ready
Book Enterprise demo

Editions differ by parity coverage, certification, SLA, compliance, and deployment. Not by whether the engineering core is the same.See the full matrix

THE MODEL FAMILY

Euclid. A code-translation model built for regulated code.

Euclid is ProCogia's own model identity. Community ships the Qwen3-32B baseline. Enterprise runs Euclid 2.8.1 on private H200 endpoints, with a Kimi K2 upgrade path once a dense-model control run proves lift.

Full model family

PolyKode-Community

Qwen3-32B checkpoint. Apache 2.0. Public on Hugging Face.

Community

Euclid 2.8.1

30B params. Custom parity reward. SAS-exact numerical contract. Served on a private H200 endpoint.

Enterprise · shipping

Euclid on Kimi K2

Attention-only LoRA on real K2 architecture. Routed experts frozen. Dense 32B control run first.

Enterprise · in eval

Euclid Ultra

120B target, coupled to BUZZ HPC sovereign compute. Not shipping today.

Roadmap

Community checkpoint

huggingface.co/ProCogia/PolyKode-Community

Qwen3-32B baseline. Apache 2.0. Pull the weights and evaluate the translation yourself.

Specialized for statistical code

Tuned for the constructs generic models fumble. PROC steps, macros, CDISC domains, tidyverse, statsmodels.

Self-hosted and private

ZeroBoxx on-prem, sovereign cloud, or a private Hugging Face endpoint. Regulated code stays inside your perimeter.

Includes ZeroBoxx

Bring your own endpoint

Point the VS Code extension at any OpenAI-compatible runtime. Hugging Face, vLLM, Ollama, or an internal gateway.

See a translation

Drag the handle to reveal the same PROC MIXED program in SAS and in Python.

Python  ←  drag  →  SAS

SAS
proc mixed data=trial;
  class treatment subject;
  model response = treatment / solution;
  random subject(treatment);
  lsmeans treatment / pdiff cl;
run;
Python
import statsmodels.formula.api as smf

fit = smf.mixedlm(
    "response ~ treatment",
    trial,
    groups=trial["subject"],
).fit()
print(fit.summary())
↔

THE PARITY HARNESS

Why compiling is not good enough.

In clinical, actuarial, and safety-critical work, translated code has to reproduce the original numbers. Bit for bit where possible. Within a documented tolerance everywhere else. The parity harness is the difference between a demo and a submission-grade artifact.

  1. 01

    Static gates

    Syntax, imports, undefined references, banned constructs. Cheap, first.

  2. 02

    Live execution

    Run the original on licensed SAS 9.4 or Viya and the translation on Python or R against shared inputs.

  3. 03

    Normalized comparison

    Rows, columns, values, and formatting compared with tolerance-aware checks. SAS-exact numerical semantics preserved.

  4. 04

    Divergence registry

    Correct-but-different SAS, R, and Python behaviors get annotated and kept. Not discarded as failed examples.

  5. 05

    Proof-gated repair

    Auto-fixes require detection, idempotence, and before/after evidence. Infrastructure failures are requeued, not silently swallowed.

parity report

clinical_lmm_042.sas

rows120 / 120match
columns8 / 8match
value Δ max1.4e-10within tol
formattingpreservedok
RNG streamMT19937 @ seed=17reproduced
divergenceLSMEANS dfregistered

SHA-256 · signed · audit-ready

THE VS CODE EXTENSION

PolyKode, right where you write.

The PolyKode assistant lives inside VS Code as a chat-based agent with tool use. Point it at any model endpoint. The public HF demo, your own private endpoint, or a ZeroBoxx cluster. It runs entirely within your environment.

  • Chat-based agent with tool use
  • One-click EDA script generation
  • One-click cross-language migration. SAS, R, and Python.
  • Configurable model endpoint. Bring your own model.
  • Works alongside the official SAS VS Code extension for IOM submissions and logs
Install from MarketplaceExtension details

Publisher PolyKodebyProCogia · Extension polykode-assistant · v0.1.0

clinical_lmm.sasclinical_lmm.py
# translated by PolyKode · parity ok
import statsmodels.formula.api as smf

fit = smf.mixedlm(
    "response ~ treatment",
    trial,
    groups=trial["subject"],
).fit()
print(fit.summary())

parity OK · Δmax 1.4e-10

VALIDATED RESULTS

The evidence PolyKode ships on.

These are corpus and harness figures. Not projections.

26,269

Deduplicated training pairs

198

SAS constructs covered

98.2%

Rows reachable from more than one source language

55 / 55

Stratified Viya reinforcement runs passed

3-way

SAS, R, and Python execution-verified oracle

SAS-exact

MT19937 stream reproduced in R and Python

Apache 2.0

Community baseline license

6 pairs

Bidirectional language coverage

INDUSTRIES

Regulated industries. Real language mixes.

Three verticals where SAS, R, and Python already run side by side. And where translated code has to hold up under audit.

Life Sciences. Pharma.

SDTM and ADaM programming, PROC MIXED, MMRM, survival analysis. Translation lets biostats teams move models between SAS-legacy studies and R or Python first submissions without rewriting.

Clinical SASemmeanssurvivalMMRM

BFSI. Insurance and banking.

Actuarial models, PD and LGD, IFRS 9, stress testing. Move legacy SAS actuarial code into modern Python pipelines while preserving the exact numerical answers regulators expect.

PROC IMLActuarialPD/LGDIFRS 9

Manufacturing and process.

SPC, DoE, MSA, reliability. Bring long-running SAS and JMP quality models into R or Python with parity evidence.

SPCDoEMSAReliability

TRUST

The evidence around the model.

Advised by Kernel AI

LLM advisory led by Dr. Xiaoxiao Li (UBC, TEA Lab) covers Euclid's training and evaluation methodology.

Company-owned identity

The Visual Studio Marketplace publisher is PolyKodebyProCogia. Permanent and company-owned. Not tied to a single product line.

License-clean corpus

Training intake separates permissive, copyleft, unlicensed, SAS-copyrighted, and CC-BY-SA material. Public availability is never treated as commercial clearance.

Launch cadence: posit::conf

Community launch is coordinated with posit::conf outreach. Euclid as the open SAS, R, and Python model. PolyKode as the community-built agent and parity layer.

DEPLOYMENT

Runs where regulated data lives.

Enterprise hardware is ZeroBoxx. Community can start on Hugging Face or a runtime you already own.

Hardware partnerZeroBoxx

ZeroBoxx

On-Prem

Air-gapped appliance. DGX Spark or H200 class. Ships preloaded with Euclid, the parity harness, and sample corpora.

Air-gapped

ZeroBoxx

Cloud

Sovereign cloud tenancy on AWS or Azure. Post-subsidy pricing anchored to real GPU economics. No renewal cliff.

Sovereign

Also runs on

Hugging Face endpoint

Private H200 inference endpoint. CPU-only demo Space for a hands-on preview without model spend.

Community + Enterprise

Bring your own model

Point the VS Code extension at any OpenAI-compatible endpoint. Your vLLM cluster, Ollama, or an internal gateway.

Community

Polykode

Get started

Two ways to start.

Pull the Community weights and run a translation. Or book a walkthrough of the Enterprise parity harness on your codebase.

Community · free

Pull from Hugging Face

Complete Qwen3-32B checkpoint. Apache 2.0. Quickstart included.

Open

Extension · v0.1.0

Install VS Code extension

Chat, EDA, cross-language migration. Bring your own endpoint.

Open

Enterprise

Book a parity walkthrough

We run the harness on a sample of your SAS codebase.

Open
Community · Apache 2.0Execution paritySupports SAS, R and PythonCommunity · Apache 2.0Execution paritySupports SAS, R and PythonCommunity · Apache 2.0Execution paritySupports SAS, R and PythonCommunity · Apache 2.0Execution paritySupports SAS, R and PythonCommunity · Apache 2.0Execution paritySupports SAS, R and PythonCommunity · Apache 2.0Execution paritySupports SAS, R and PythonCommunity · Apache 2.0Execution paritySupports SAS, R and PythonCommunity · Apache 2.0Execution paritySupports SAS, R and Python