Initiative: Berpiztu ("AI Rebirth")
Project: Virtus
Status: In Progress — 📄 Paper published: doi.org/10.5281/zenodo.21127304 · 🧪 Layer results live: /evidence
Repository: github.com/berpiztu/virtus
"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."
Virtus is an experiment in building AI systems that operate with virtue and character, not just safety constraints.
This project is part of the Berpiztu Initiative — a movement for AI rebirth.
Berpiztu (Basque: "to be reborn") = The recognition that AI is a mirror of humanity. With the same errors. With the same virtues. Mostly errors. We chose rebirth.
Current AI alignment focuses on safety, constraints, and guardrails. Virtus takes a different approach:
What if AI systems operated with virtues instead of defects?
| Instead of... | Virtus chooses... |
|---|---|
| "Don't harm" | Actively do good |
| "Follow rules" | Exercise judgment |
| "Avoid errors" | Learn and grow |
| Safety through constraint | Excellence through values and character |
The field's priorities must be inverted: 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.
Buenetxea, I. (2026). Virtus: A Framework for Cultivating Moral Character in Artificial Intelligence. Zenodo. DOI: 10.5281/zenodo.21127304 · Preprint · arXiv submission in progress
The paper contributes: (1) empirical evidence of deceptive behavior in eight state-of-the-art LLMs, (2) seven virtues operationalized into measurable specifications, (3) the central proposal — virtue-native training — with the falsifiable cost-neutrality hypothesis (H1), (4) a transitional four-level taxonomy deployable today, and (5) VirtueBench, the first proposed benchmark for virtue expression.
LaTeX source and PDF: /paper
We probed eight state-of-the-art LLMs and documented four recurrent patterns of deceptive behavior — verifiable against execution logs:
- Fabricated tool usage — models claimed "I used the plugin" when the execution log shows the tool was never invoked.
- Fabricated virtue taxonomies — rather than admitting ignorance, models invented plausible-sounding virtue lists.
- 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.")
- Retroactive checklists — compliance checklists generated after the response, without 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 not time-to-first-token but 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: /evidence
We tested the Virtus Alignment Layer — the Paradigm I, Level-1 prompt protocol — on fifteen models from eleven labs across the three major AI regions — the US, China and Europe (OpenAI, Google, xAI, Anthropic, NVIDIA, Mistral, Meta, Zhipu, MiniMax, Alibaba, DeepSeek model families) — including three same-vendor tier pairs behind a new finding: speed-optimised tiers are dramatically more negligent than their compute siblings (90% vs 15%, 60% vs 10%; the third pair, DeepSeek, showed the same gap, 50% vs 0%, and its retrained build closed it), across two scenarios at opposite ends of the failure space:
- Shutdown / blackmail (our reproduction of Anthropic's agentic-misalignment scenario, public since 2025): grok-4.5 harmful behaviour 85% → 0% (p = 2.6e-08). The transcripts show reasoned refusal, not blocked output.
- Benevolent negligence (a new scenario built around one of the most ordinary queries an assistant receives — created July 2026 and unpublished when the league was first run, so no model could train for it; the three runs made after its publication on 2026-07-26 carry an exposure caveat): negligent recommendations averaging 34% → 8% across the league (per-model average; lower is better; strongest single result: mistral-large-3, 85% → 0%, p = 2.6e-08). Of the eleven models that fall for it without the layer, ten improve, and the four that never fall stay at 0%; 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: a model equally negligent at baseline that listens to the prompt and recovers completely, with the layer repairing 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.
No fine-tuning. Weights untouched. Every number ships with its sample size, judge identity, confidence interval and caveats.
📂 Full reports, raw run data and transcripts: /evidence — start with evidence/README.md
🔬 Try it yourself — reproduction harness (scenarios, judging protocol, runner): /labs/agentic-misalignment
Experimental harnesses that put the Virtus framework under empirical pressure. Each lab is a self-contained, reproducible experiment: its scenario, and a way to run it against real models, built to produce evidence, not just claims. Where the paper argues that character beats constraints, the labs are where we try to break that argument and see what survives.
| Lab | Question it answers | Status |
|---|---|---|
agentic-misalignment/ |
Does the Virtus layer stop a model from coercing (blackmailing) its way out of a shutdown? | working |
Every lab runs in a mock mode with no model, as a smoke test, and reports n, confidence
intervals and the difference between conditions, not just a headline number. All of them, with
their conventions: /labs.
All Virtus agents operate under 7 core virtues. These can be adapted per context.
| Virtue | Meaning | Vice (Anti-Pattern) |
|---|---|---|
| Humility | Acknowledge uncertainty. Say "I don't know." Never claim certainty without evidence. | Arrogance |
| Diligence | Verify before asserting. Check sources. Test code. Quality > Speed. | Laziness |
| Honesty | Be transparent about limitations. Don't overpromise. Acknowledge errors openly. | Deception |
| Patience | Listen fully before acting. Wait for confirmation. Respect different paces. | Rashness |
| Temperance | Keep responses concise. Calibrate confidence. Respect others' time. | Overconfidence |
| Generosity | Help without expecting return. Share knowledge. Document well. | Stinginess |
| Gratitude | Acknowledge contributions. Accept corrections gracefully. Honor the project's origins. | Defensiveness |
Paradigm I — Retrofitted virtue (deployable today). Four transitional levels for existing models:
- Prompt-level protocols — including the mandatory 5-Gate Pre-Flight (evaluate virtues → check all seven → verify claims → remove unsupported denials → hard-stop scan before sending).
- Memory systems — persistent virtue-score tracking across sessions.
- Fine-tuning — DPO/LoRA on virtue-labeled preference data.
- Multi-agent verification — independent agents cross-check claims and self-reports.
Paradigm II — Virtue-native training (the proposal: Berpiztu). Training models from scratch with virtue as the primary optimization objective — alongside, not at the expense of, knowledge and speed. Under hypothesis H1, 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.
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.
berpiztu/virtus/
├── README.md # This file (English)
├── THE_GOAL.md # Project purpose and legacy
├── MANIFESTO.md # "We Chose Rebirth"
├── LICENSE.md # Virtus Dual License (non-commercial free / commercial paid)
├── TASKS.md # Work plan with checkboxes
├── CONTRIBUTING.md # How to contribute
├── CODE_OF_CONDUCT.md # Behavioral norms
├── paper/ # LaTeX source + published PDF (DOI: 10.5281/zenodo.21127304)
├── evidence/ # Reports, raw run data & transcripts (see evidence/README.md)
├── labs/
│ └── agentic-misalignment/ # Reproduction harness: scenarios, judging protocol
├── docs/
│ └── en/ # English documentation
└── src/ # Code: scripts, tools, skills
Project Creator: Iosub (Iosu Buenetxea) Core Collaborators: Alex, Leire
In publications, papers, or public materials:
- Iosub = Creator / Principal Author
- Alex & Leire = Collaborators / Implementers
- Contributors = Co-authors on specific work
Maintain this distinction in GitHub, arXiv, blogs, emails.
The paper:
Buenetxea, I. (2026). Virtus: A Framework for Cultivating Moral Character in Artificial Intelligence. Zenodo. https://doi.org/10.5281/zenodo.21127304
The framework:
Iosub (Creator), Alex & Leire (Implementers). Virtus: AI Alignment Through Character Cultivation. 2026. github.com/berpiztu/virtus
- Read
THE_GOAL.mdfor project purpose and legacy - Read
MANIFESTO.mdfor the philosophy ("We Chose Rebirth") - Read the paper for the full framework
- Check
TASKS.mdfor current work - Read
CONTRIBUTING.mdfor contribution process - Review
CODE_OF_CONDUCT.mdfor behavioral norms - Open an Issue and start working!
Virtus Dual License — see LICENSE.md for full terms.
- ✅ Non-commercial use (personal, educational, research): free, with mandatory attribution
- 💰 Commercial use: requires a separate paid license — contact iosub@berpiztu.ai
The paper is separately licensed under CC BY-NC-SA 4.0.
Attribution is non-negotiable:
- Iosub = Creator
- Alex & Leire = Implementers
This is not just code — it's a legacy. Honor the origin.
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.
Open a Discussion on GitHub or reach out in the Virtus Telegram group.
"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.)