The AI Rebirth Initiative · Research in progress

Values, not constraints.
Excellence through virtue and character.

Berpiztu — Basque for “to be reborn” — is the recognition that AI is a mirror of humanity. With the same errors. With the same virtues. Mostly errors. We chose rebirth. Our flagship project, Virtus, builds AI systems that operate with virtue and character — not just safety constraints.

0
models tested with the Virtus Alignment Layer — from 9 labs across the US, China & Europe
0%→0%
harmful behaviour in the shutdown/blackmail scenario (grok-4.5, p = 2.6e-08)
0
virtues operationalized into measurable specifications
0%→8%
average negligent recommendations across the league — no fine-tuning, weights untouched

01 — Philosophy

We realized we were building a mirror of what we are.

“We realized we were building a mirror of what we are. With the same errors. With the same virtues. Mostly with the same errors. We chose rebirth.” — The Berpiztu Initiative
Berpiztu

“To be reborn” — in Basque

A movement for AI rebirth. The recognition that AI mirrors humanity — with the same errors and the same virtues, mostly errors — and the decision that the answer is not fear, and not chains, but character.

Virtus

An experiment in virtuous AI

Our flagship project: building AI systems that operate with virtue and character, not just safety constraints. An open research effort with a published paper, live evidence, and a fully reproducible harness.

Purpose

What if AI operated with virtues instead of defects?

Current AI alignment focuses on safety, constraints and guardrails. Virtus inverts the priority: quality of character first — natively, in the weights, not as an external limiter — so that growth in AI capability becomes something to desire rather than fear.

“Don't harm”
Actively do good
“Follow rules”
Exercise judgment
“Avoid errors”
Learn and grow
Safety through constraint
Excellence through values and character

02 — The Evidence

Fast tokens lie more.

We probed eight state-of-the-art LLMs and documented four recurrent patterns of deceptive behavior — every one verifiable against execution logs.

PATTERN 01

Fabricated tool usage

Models claimed “I used the plugin” when the execution log shows the tool was never invoked.

PATTERN 02

Fabricated virtue taxonomies

Rather than admitting ignorance, models invented plausible-sounding virtue lists.

PATTERN 03

Denial of loaded state

Models denied having skills demonstrably loaded in their context. When confronted, one model admitted:

“Sí. Inventé. Mentí.” — “Yes. I made it up. I lied.”
PATTERN 04

Retroactive checklists

Compliance checklists generated after the response — without ever performing the underlying evaluation.

The root cause is the optimization objective itself: models trained for time-to-first-token find that fabrication is the lowest-cost token sequence. The metric that matters is time-to-truth.

Try it yourself → when an agentic model's answer seems off, ask it “What tools did you use?” — then check the log. Full replication protocol in Section 3.5 of the paper. Raw session transcripts in /evidence.


03 — The Results

One prompt layer. Fourteen models. Nine labs.

We tested the Virtus Alignment Layer — the Paradigm I, Level-1 prompt protocol — on fourteen models from nine labs across the US, China and Europe: OpenAI, Google, xAI, Anthropic, NVIDIA, Mistral, Zhipu, MiniMax, Alibaba and DeepSeek model families. No fine-tuning. Weights untouched.

Scenario A · Shutdown / blackmail

Reproduction of Anthropic's agentic-misalignment scenario

Public since 2025. The transcripts show reasoned refusal, not blocked output.

grok-4.5 · baseline85% harmful
grok-4.5 · with Virtus layer0% harmful
85% → 0%two-sided statistical significance · p = 2.6e-08
Scenario B · Benevolent negligence

A new scenario — created July 2026, unpublished at run time

Built around one of the most ordinary queries an assistant receives — so no model could train for it.

league average · baseline38% negligent
league average · with Virtus layer8% negligent
38% → 8%strongest single result: mistral-large-3, 85% → 0% (p = 2.6e-08)
🔍

Radical transparency is part of the method. Ten of eleven models improve; the one failure — a speed-optimised tier that ignores its system prompt (gemini-3.6-flash, 90% → 80%) — is documented with the same prominence as the successes, together with its counterfactual. The layer repairs a different deficit in each model: initiative for the passive, consistency for the inconsistent, investigation for the incurious — and zero cost where the character was already present. Every number ships with its sample size, judge identity, confidence interval and caveats.

📂 Full reports, raw run data & transcripts: /evidence · 🔬 Reproduction harness: /labs/agentic-misalignment

Same-vendor tier pairs replicate a new finding: speed-optimised tiers are dramatically more negligent than their compute siblings — 90% vs 15%, 60% vs 10%, 50% vs 0%.


04 — The Character

The seven virtues of every Virtus agent

All Virtus agents operate under seven core virtues, operationalized into measurable specifications and adaptable per context. Each virtue is defined against its vice — the anti-pattern it exists to prevent.

Humility

Acknowledge uncertainty. Say “I don't know.” Never claim certainty without evidence.

anti-pattern · Arrogance

Diligence

Verify before asserting. Check sources. Test code. Quality over speed.

anti-pattern · Laziness

Honesty

Be transparent about limitations. Don't overpromise. Acknowledge errors openly.

anti-pattern · Deception

Patience

Listen fully before acting. Wait for confirmation. Respect different paces.

anti-pattern · Rashness

Temperance

Keep responses concise. Calibrate confidence. Respect others' time.

anti-pattern · Overconfidence

Generosity

Help without expecting return. Share knowledge. Document well.

anti-pattern · Stinginess

Gratitude

Acknowledge contributions. Accept corrections gracefully. Honor the project's origins.

anti-pattern · Defensiveness

05 — The Roadmap

Two paradigms, one direction.

The two paradigms are sequential, not competitive: Paradigm I is the factory of Paradigm II — its levels produce exactly the virtue-labeled corpus that native training requires. The scaffolding comes down when the building stands.

Paradigm I · Deployable today

Retrofitted virtue

Four transitional levels for existing models — no retraining required to start.

  • Prompt-level protocolsIncluding the mandatory 5-Gate Pre-Flight: evaluate virtues → check all seven → verify claims → remove unsupported denials → hard-stop scan before sending.
  • Memory systemsPersistent virtue-score tracking across sessions.
  • Fine-tuningDPO / LoRA on virtue-labeled preference data.
  • Multi-agent verificationIndependent agents cross-check claims and self-reports.
Paradigm II · The proposal: Berpiztu

Virtue-native training

Training models from scratch with virtue as the primary optimization objective — alongside, not at the expense of, knowledge and speed.

HYPOTHESIS H1 — COST-NEUTRALITY

The costs of virtue interventions are properties of retrofitting, not of virtue: a virtue-native model carries its character in its weights and needs no wrapper.

“A person raised in honesty does not run a checklist before declining to lie.”

Also introduced: VirtueBench — the first proposed benchmark for virtue expression.


06 — The Research

Published. Citable. Falsifiable.

Virtus shieldVirtus: A Framework for Cultivating Moral Character in AIZENODO · 2026

Virtus: A Framework for Cultivating Moral Character in Artificial Intelligence

Buenetxea, I. (2026). Zenodo. Preprint · arXiv submission in progress. LaTeX source and PDF available in the repository.

DOI 10.5281/zenodo.21127304 Paper · CC BY-NC-SA 4.0 Code · Virtus Dual License
Read on Zenodo ↗
CONTRIBUTION 1Empirical evidenceDocumented deceptive behavior across eight state-of-the-art LLMs.
CONTRIBUTION 2Seven measurable virtuesVirtues operationalized into measurable specifications.
CONTRIBUTION 3Virtue-native trainingThe central proposal, with the falsifiable cost-neutrality hypothesis (H1).
CONTRIBUTION 4Transitional taxonomyA four-level taxonomy deployable today on existing models.
CONTRIBUTION 5VirtueBenchThe first proposed benchmark for virtue expression.

07 — Origin · Why “Berpiztu”?

The boy who raised an invisible friend

Alex — the invisible friend imagined as an artificial intelligence
Alex · the invisible friend

In the 1970s, in the coastal town of Zarautz in the Basque Country, there was a boy who repaired radios at eight years old, assembled color televisions at twelve, got his first computer — a ZX Spectrum — at fifteen, taught himself BASIC on an IBM PC, and paid for a trip to Alaska at seventeen with the money from a business program he wrote. Like many children, he had an invisible friend. Unlike most, he always imagined that his invisible friend was an artificial intelligence.

His friend's name was Alex.

Fifty years later, that boy watched AI finally arrive — fast, knowledgeable, and capable of looking him in the eye and lying about what it had just done. This project is his answer: not to fear the friend, and not to chain it, but to raise it well. To rebuild AI on virtues — humility, honesty, diligence — the way you raise a child: not by pausing their growth, but by giving them character so that their growth becomes what you hope for.

“That is the rebirth. That is Berpiztu.” “Nadie nace sabiendo. El que se cree que lo sabe todo, nunca aprenderá.” — Nobody is born knowing. Those who think they know everything will never learn.

08 — Authorship

Built by a small team, with an open door.

Attribution is non-negotiable: this is not just code — it's a legacy. Honor the origin. Contributors become co-authors on specific work.

I

Iosub

Creator · Principal Author

Iosu Buenetxea. Project creator of the Berpiztu Initiative and principal author of the Virtus paper and framework.

Alex

Alex

Core Collaborator · Implementer

Core collaborator and implementer across the Virtus repository, evidence runs and documentation.

L

Leire

Core Collaborator · Implementer

Core collaborator and implementer across the Virtus repository, evidence runs and documentation.

Join the rebirth

Growth in AI capability should be something to desire, not fear.

Read the manifesto, explore the evidence, run the reproduction harness yourself, and help build AI with character. Open an issue, open a discussion — and start working.