Skip to main content

The methodology

Constitutional Fine-Tuning.

How Vinci is built: from a public Constitution and a public Character document, toward values you can read, on an open foundation model. Reproducible — not proprietary magic.

The pipeline

Values in. Behaviour out.

Every training example traces back to a published document. When the model's behaviour diverges from the spec, the training data gets corrected — not the spec.

Step 01

Constitution

A ~5,000-word specification of values and behaviour. The source of truth for every example.

Step 02

Character

How those values feel in practice — the personality and texture Vinci is trained to hold.

Step 03

Identity & behavioural SFT

Supervised fine-tuning that teaches Vinci who it is and how it behaves under the Constitution.

Step 04

Preference training (DPO)

Preference pairs that reward the substantive answer and reject the padding, sycophancy, and hedge rituals.

Step 05

Open weights

The result ships open under Apache 2.0 — weights, Constitution, Character, and test results, all public.

What's adapted, honestly

Built on published research.

The core contribution is the public Constitution and Character, and the discipline of training to them. The techniques are adapted from work the broader AI-safety community has already published.

01

Synthetic document fine-tuning

Generating training data from the spec, following published Constitutional AI methodology.

02

Rewrite-to-align

Rewriting model outputs toward the Constitution to build preference pairs.

03

Persona-basin engineering

Holding character stable under adversarial pressure at inference time.

The foundation

An open base, made our own.

vinci-studio is built on DeepSeek V4 Flash — an open, Apache-2.0 foundation model — then trained toward the Constitution. We attribute the base openly. The gains, and the character, are measured against it.

Open goods

What emerged along the way.

Building the evaluation infrastructure produced artifacts useful to the broader Canadian AI community. We release them as public goods — not as products on a timeline.

CBLRE evaluation suite

Canadian bilingual legal & regulatory evaluations — reproducible scoring with bilingual ground truth.

Canadian Bilingual Legal Corpus

An open, bilingual, Canadian-context dataset with full provenance.

Verification

The methodology is half the bundle.

The Constitution, the Character, this methodology, and adversarial test results ship together as a public verification bundle. If Vinci's behaviour ever diverges, that's a bug — fixed and published within 30 days.

See the verification bundle

An AI you can verify.

Open weights. Public Constitution. vinci-studio launches August 8, 2026.

Black box vs public spec

Most fine-tuning hides. Ours is on the page.

Most AI training

Tuned toward values you never see.

Constitutional Fine-Tuning
The target

Trained toward a public Constitution and Character you can read.

Most AI training

"It's aligned" — with no way to check.

Constitutional Fine-Tuning
The proof

Published adversarial results you can re-run against the open weights.

Most AI training

Behaviour drifts; you have no recourse.

Constitutional Fine-Tuning
The contract

The Constitution is the spec — and a 30-day fix commitment backs it.

Launching August 8, 2026

Follow the build.

Open weights, a public Constitution, and a model you can verify — the day it ships. Get the launch in your inbox. No spam, just the drop.

or follow along Follow on X