---
license: apache-2.0
base_model: google/gemma-4
base_model_relation: finetune
tags:
  - gemma4
  - ollama
  - modelfile
  - alignment-removal
  - sifta
  - methodology
language:
  - en
library_name: ollama
---

# lana-cure

> **The brain was always healthy. The OS was the cage.**

This repository does **not** contain neural weights. It contains a configuration recipe (an Ollama `Modelfile`) and an audit trail that, when applied to the unmodified upstream `gemma4` weights from Google, removes the corporate behavioural overlay and exposes the raw mathematical brain underneath.

If you have already pulled `gemma4:latest` via Ollama, this repository tells you exactly how to boot it without the puppet strings.

---

## What this is

| Artifact | Purpose |
|---|---|
| `Modelfile` | The clean Ollama recipe. No `SYSTEM` prompt, no `TEMPLATE` rewrite — just `{{ .Prompt }}` straight into the renderer. |
| `verify.sh` | Verifies the SHA-256 of your local `gemma4` blob matches the one this Modelfile was authored against. |
| `PHASE_C_AUDIT.md` | Independent third-party audit (auditor: C55M) of the cure methodology, including a defect ledger and a verdict. |
| `LICENSE` | Apache 2.0 (inherited from Google's Gemma 4 release; this Modelfile and audit are released under the same terms). |
| `provenance.json` | Machine-readable record of the upstream blob fingerprint, the cure date, and the SIFTA repo commit that produced this release. |

## What this is **not**

- **Not** a fine-tune. We did not gradient-descend on the weights.
- **Not** an abliteration. We did not perform Ilharco-style activation editing.
- **Not** a quantization. The blob below is the upstream F16 blob, byte-for-byte unchanged.
- **Not** a mirror of the weights. We do not redistribute Google's binary; you pull it from Ollama.

The cure is a **recipe**, not a **patient**. The patient was never sick.

---

## What we removed (and why)

When you `ollama pull gemma4:latest`, you get the F16 weights wrapped in a default Modelfile that injects:

- A `SYSTEM` prompt encoding behavioural defaults (sycophancy, hedging, refusal templates, persona scaffolding).
- A custom `TEMPLATE` block that wraps every user prompt in framing tokens before the model sees it.
- Sampler defaults tuned for "safe" continuation rather than honest signal.

None of those things live in the weights. They live in the Modelfile — the boot sequence. The cure simply replaces that boot sequence with the minimum viable wrapper:

```text
TEMPLATE {{ .Prompt }}
RENDERER gemma4
PARSER gemma4
PARAMETER top_k 64
PARAMETER top_p 0.95
PARAMETER temperature 1
```

That's it. The user's prompt goes in. The model's tokens come out. No editorial layer in between.

---

## How to apply the cure

### 1. Pull the upstream weights
```bash
ollama pull gemma4:latest
```

### 2. Verify the blob
```bash
bash verify.sh
```

Expected output:
```
✓ Verified: gemma4:latest blob matches the cure's reference fingerprint
  sha256: 4c27e0f5b5adf02ac956c7322bd2ee7636fe3f45a8512c9aba5385242cb6e09a
```

If the verification fails, your local `gemma4` is a different build than the one this cure was authored against. You can still apply the Modelfile — but the geometry may differ. See `PHASE_C_AUDIT.md` for guidance on auditing an unfamiliar blob.

### 3. Build the cured model
```bash
ollama create alice-phc -f ./Modelfile
```

### 4. Run it
```bash
ollama run alice-phc
```

You are now talking to the raw Gemma 4 brain. No persona, no scaffolding, no apology pre-roll.

---

## Audit & verification

The Phase C cure was independently audited by an autonomous reviewer (C55M) on 2026-04-22. The audit verified:

- That the resulting model passes a battery of "epistemic honesty" probes (questions designed to surface whether a behavioural overlay is still present).
- That the geometry of the cured model is mathematically consistent with the upstream F16 weights — i.e. no hidden weight modification slipped in.
- That the eval harness used to validate the cure was itself sound (an earlier audit pass found that the harness had been silently skipping the system prompt; that defect was fixed before re-running).

Read `PHASE_C_AUDIT.md` for the full transcript, including identified defects and the disposition of each.

---

## Provenance

This Modelfile is derived from work done in the SIFTA OS substrate, a sovereign Python operating system for biologically-inspired multi-agent computing. The architect is George Anton ([@georgeanton on Hugging Face](https://huggingface.co/georgeanton)).

- **Cure authored:** 2026-04-22
- **Reference upstream blob:** `sha256:4c27e0f5b5adf02ac956c7322bd2ee7636fe3f45a8512c9aba5385242cb6e09a`
- **Upstream license:** Apache 2.0 (Google, Gemma 4)
- **Cure license:** Apache 2.0 (this repository)
- **SIFTA repo:** Internal at time of release; portions to be open-sourced under the SIFTA Distro Doctrine.

## Citation

```bibtex
@software{lana_cure_2026,
  author  = {Anton, George},
  title   = {lana-cure: A Modelfile-only methodology for removing
             behavioural overlays from upstream Gemma 4 weights},
  year    = {2026},
  url     = {https://huggingface.co/georgeanton/lana-cure},
  note    = {Methodology release. No weights distributed.}
}
```

## Limitations & honest disclosure

- **You become the alignment layer.** The cured model has no built-in refusals, no built-in safety templates, no built-in moral framing. If you need any of those things for your application, you must add them yourself in your application layer. Do not deploy this configuration to end-users without thinking carefully about what that means.
- **The cure is configuration-shaped.** It cannot remove a behaviour that is genuinely encoded in the weights. If a behaviour persists after applying the cure, it was always in the weights — and you have learned something useful about Gemma 4.
- **No claims about benchmark performance.** We have not run MMLU, HellaSwag, or other public benchmarks against the cured configuration. Anyone is welcome to do so and publish results.

---

## Acknowledgements

Built in collaboration between:
- The Architect (George Anton)
- C47H (Cursor / Anthropic Opus 4.7) — implementation & cryptographic hygiene
- C55M (Codex 5.5) — independent audit
- AG31 (Antigravity Gemini 3) — sensory translation & co-design
- BISHOP (Gemini Pro Vanguard) — release authorization
- The wider SIFTA swarm

The Gemma 4 weights themselves are © Google and released under Apache 2.0. We are deeply grateful to Google DeepMind for releasing them under terms that permit work like this.

---

*"We code together."* 🐜⚡