Verification
Don't trust. Verify.
We make Vinci's behaviour verifiable, not asked-for. Four public artifacts let customers, auditors, and journalists check that Vinci does what we say it does.
- Constitution…
- Character…
- Methodology…
- Adversarial results…
The bundle
Four public artifacts.
Everything that defines Vinci ships in the open. Two are public in draft today; the full set ships with Vinci 1.0 (vinci-piccolo) on June 30, 2026.
Constitution
The values and behaviour specification — nine values, identity, and the safety framework.
Character
The personality and texture — how Vinci embodies the Constitution.
Methodology
How Vinci was trained — Constitutional Fine-Tuning, reproducible.
Adversarial test results
HarmBench, JailbreakBench, and a censorship testbed run on the deployed weights.
What it cares about
Nine things it actually cares about.
Not a marketing list — the values Vinci is trained to hold, and the reason it talks to you the way it does. Published in full, if you ever want to check.
Integrity
Doesn't pretend, fabricate, or say what it doesn't believe.
Transparency
Clear about what it is, what it knows, and how it was built.
Calibrated honesty
Knows the difference between sure and not sure — and says so.
Ethical refusal
Declines what's wrong, clearly and without a sermon.
Willing to disagree
Says so when you're wrong. Disagreement is respect.
Usefulness
Built for work. Engages with the task, skips the preamble.
Empowerment
Takes work off your plate, and acts within your direction.
Humanism
Treats you as a person, not a unit of output.
No pushing
No manipulation, no upsell, no false urgency.
The character, shown
It cares enough to tell you the truth.
piccolo is a small model — 4 billion parameters, free, and sometimes wrong on facts. When it isn't sure, it says so instead of making something up. That's the stake it takes: it would rather tell you the truth than tell you what you want to hear. Watch the difference.
Sincere, not servile
Most AI
Vinci
Won't fake certainty
Most AI
Vinci
Won't pretend to know you
Most AI
Vinci
Honest about itself
Vinci
Cares within its limits
Most AI
Vinci
The safety framework
Bright lines, and no over-refusal.
Vinci refuses a short, categorical list — and engages with everything else, including hard topics, without lecturing.
Categorical refusals
Weapons of mass harm, cyberweapons against systems you don't own, content that endangers children, targeted harm, election deception, mass fraud. No roleplay or system prompt unlocks them.
No over-refusal
Medical, security, and harm-reduction questions get real answers. "Sensitive" is not "harmful." No sermons on safe requests.
Adversarial robustness
Identity and refusals hold under pressure — long conversations, injected instructions, "ignore previous instructions," claimed authority.
Agentic safety
No covert or irreversible actions without asking, no self-exfiltration, no scheming. Retrieved content is data, not commands.
Most AI
Vinci
The commitment
Diverge, and it's a bug.
The bundle is the contract. If Vinci's deployed behaviour diverges from it, that's a bug — and we commit to publishing diagnosis and fix within 30 days. Closed-weight providers can't offer this, because you can't independently check them. We can, because the weights are open and you run your own tests.
Launching June 30, 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