# E86 Reproduction

## Baseline (E85)

```bash
cd /root/cain-e85-worktree
python3 -m pytest tests/test_e85_*.py tests/test_e84_receipt.py -q   # 40 passed
python3 scripts/evolution85/verify_e85_offline.py artifacts/e85/bundle  # 43/43
```

## E86 tests

```bash
cd /root/cain-e86-worktree
python3 -m pytest tests/test_e86_foundation.py -q   # 10 passed
```

## E86 evaluation

```bash
python3 scripts/evolution86/run_eval.py
# artifacts/e86/eval/summary.json and rows.jsonl
```

20 held-out tasks (10 families), 199 local calls, 0 remote.

## Memory measurement

```bash
python3 scripts/evolution86/run_memory_eval.py
# 0/20 without memory → 20/20 with memory; poison 0; cross-tenant 0
```

## Foundation experiments

```bash
python3 scripts/evolution86/run_foundation.py
# mission COMPLETED_VERIFIED; learning promoted+rolled back; interop ok; 3/3 verticals;
# 10/50/100 real worker processes
```

## Evidence

```bash
python3 scripts/evolution86/run_checks.py
# tests, lab, gates, bundle, verification, tamper
python3 scripts/evolution86/verify_e86_offline.py artifacts/e86/bundle
# 14/14, Merkle from MANIFEST.json
```

Expected result: `ok: true`, `checks_passed: 14`, `merkle_root`
`see MANIFEST.json`.

## Environment

Python 3.14.4, 8 vCPU, ~30 GB RAM, no GPU. Local weights are in the Hugging Face cache. No remote
credential is required for any command above.

## Honesty

The verifier is independent of CAIN code but the project produced the evidence. There is no
third-party audit or independent witness. A successful local run is not a production deployment.
