# E125 Reproduce

All commands are run from the repository root. Python 3.14, CPU only. Total runtime is about
25–35 minutes; the adaptation step is the slow one.

## 1. Confirmatory challenge (about 30 s)

```bash
python3 -m cain.cain2027.e125.runner --seed 126 --out artifacts/e125/challenge
```

Before running, check the frozen commitment:

```bash
cat artifacts/e125/challenge/commitment.json
# must equal 9f5d67555e04e99fbea80f1ec15b54c4228978a969d2fe8954ac90a6113a12e6
```

The pilot seed 125 is preserved separately:

```bash
python3 -m cain.cain2027.e125.runner --seed 125 --out artifacts/e125/challenge_pilot
```

## 2. Continual learning (about 20 s)

```bash
python3 -m cain.cain2027.e125.continual
```

## 3. Causal world model (about 5 s)

```bash
python3 -m cain.cain2027.e125.causal
```

## 4. Cross-domain transfer (about 5 s)

```bash
python3 -m cain.cain2027.e125.transfer
```

## 5. Collective intelligence (about 5 s)

```bash
python3 -m cain.cain2027.e125.collective
```

## 6. Scientific discovery (about 2 s)

```bash
python3 -m cain.cain2027.e125.discovery
```

## 7. Self-improvement (about 80 s)

```bash
python3 -m cain.cain2027.e125.self_improvement
```

## 8. Governance boundary (about 15 s)

```bash
python3 -m cain.cain2027.e125.governance
```

## 9. Long-horizon mission (300 s)

```bash
python3 -m cain.cain2027.e125.mission
```

The mission clears its journal at start. It launches a worker subprocess, kills it twice, restarts
it, and verifies signatures and the hash chain.

## 10. Real model adaptation (10–15 min, CPU)

```bash
python3 -m cain.cain2027.e125.adaptation
```

This runs both LoRA and full fine-tuning and writes `results_lora.json`, `results_full.json`, and
`results.json`. It downloads the model only if it is not already cached.

## 11. Build the evidence bundle

```bash
python3 scripts/evolution125/build_e125_evidence.py
```

This copies raw files, writes `MANIFEST.json`, runs the standalone verifier, and writes
`VERIFIER_OUTPUT.txt`.

## 12. Verify offline

From the bundle directory:

```bash
cd clawx-site/evidence/e125-generalization-2026-11-21
python3 verify_e125_offline.py
```

The verifier imports **no CAIN code** and uses only the Python standard library. It recomputes
challenge, continual, causal, transfer, collective, self-improvement, governance and mission
numbers from the raw rows, and fails closed on a missing file or a changed manifest hash.

To test under an unprivileged account (as root, on a copy):

```bash
cp -r clawx-site/evidence/e125-generalization-2026-11-21 /tmp/e125-bundle
chmod -R a+rX /tmp/e125-bundle
setpriv --reuid=nobody --regid=nogroup --clear-groups python3 /tmp/e125-bundle/verify_e125_offline.py
```

## What “reproduce” means here

Reproducing these commands reproduces the **numbers in the bundle**. It does **not** make the
challenge externally validated, does not make the team independent, and does not turn the
self-generated key into an independent witness. External reproduction remains
**BLOCKED_EXTERNAL_REVIEW** until a named outside reviewer runs this file and signs their result.
