# Reproducing CARVE Syn-2Cap Results

Follow [`.cursor/BUILD-PLAN.md`](.cursor/BUILD-PLAN.md) Phase B–C for the full commit sequence.  
For commit/report/README rituals after each milestone, see [`.cursor/PROJECT-WORKFLOW.md`](.cursor/PROJECT-WORKFLOW.md).

## Environment

```bash
pip install -e ".[dev]"
python .cursor/scripts/check_env.py
```

## Syn-2Cap construction

1. Train LoRA_A on capability A only → save `delta_a.pt` with key `delta_w`
2. Train LoRA_B on capability B only → save `delta_b.pt`
3. Merge:

```bash
python scripts/synthetic_entangle.py merge \
  --delta-a delta_a.pt \
  --delta-b delta_b.pt \
  --output results/syn2cap/entangled_adapter \
  --config configs/syn2cap_1.5b.yaml
```

## Eigenvalue diagnostic (no cut)

```bash
python scripts/synthetic_entangle.py diagnostic-only \
  --s-f path/to/S_f.pt \
  --s-r path/to/S_r.pt \
  --layer layer0 \
  --output results/diagnostic
```

## CARVE go/no-go (GPU)

Targets: **FE ≥ 0.90**, **RF ≥ 0.95**, CARVE Pareto-dominates Maat-SVD at matched |J|.

```bash
pytest tests/test_synthetic_e2e.py -m gpu -v
```

## Reproduced result — complete proof suite (2026-07-17)

```bash
pip install -e ".[dev]"
python scripts/run_complete_proof.py \
  --adapter results/gate_c_live/entangled_adapter \
  --probes results/gate_c_live/probes
```

Expect `all_behavioral_pass: true` in `results/complete_proof_*/complete_proof_summary.json`.
Full MLP win uses selective soft-γ + short retain repair. See [`report/2026-07-17-complete-proof.md`](report/2026-07-17-complete-proof.md).

---

## Reproduced result — publishable Syn-2Cap suite (2026-07-17)

Decision **PASS**. Figures in `report/assets/syn2cap_*.png`. Summary: `results/publish_live/publish_summary.json`.

```bash
pip install -e ".[dev]"
python scripts/run_publishable_eval.py \
  --adapter results/gate_c_live/entangled_adapter \
  --probes results/gate_c_live/probes \
  --output results/publish_live \
  --lambda-threshold 1.5
```

Expected (λ=1.5 CARVE): FE≈1.0, RF≈1.0; Maat-SVD RF≈0; relearning forget acc rises by step 30.

Full MLP confirmation (slower):

```bash
python scripts/run_publishable_eval.py --full-mlp --skip-pareto --skip-relearn \
  --adapter results/gate_c_live/entangled_adapter \
  --probes results/gate_c_live/probes \
  --output results/publish_full_mlp \
  --lambda-threshold 1.5
```

---

## Reproduced result — Syn-2Cap Gate C (2026-07-16)

**Hardware:** NVIDIA GPU ≥6GB VRAM (tested RTX 4050 Laptop).  
**Decision:** `PASS` — see `results/gate_c_live/gate_c_summary.json` and [`report/2026-07-16-syn2cap-gate-c.md`](report/2026-07-16-syn2cap-gate-c.md).

```bash
pip install -e ".[dev]"
python scripts/run_syn2cap_gate_c.py --steps 100 --rank 8 --lambda-threshold 1.5 \
  --output results/gate_c_live
```

Expected ablation (approximate; small holdout is noisy):

| Method | FE | RF |
|--------|-----|-----|
| CARVE | ≥0.90 | ≥0.95 |
| Maat-SVD | lower FE or much lower RF | — |

If Cap A/B already trained under `results/gate_c_live/lora_{a,b}/adapter`:

```bash
python scripts/run_syn2cap_gate_c.py --merge-only --output results/gate_c_live --lambda-threshold 1.5
```

---

## Reproduced result — separability diagnostic (2026-07-05)

Toy synthetic run; certificate `SURGERY_VIABLE`. Figure committed at `report/assets/synthetic_spectrum_20260705.png`.

```bash
pip install -e ".[dev]"
python scripts/synthetic_entangle.py diagnostic-only \
  --s-f results/full_eval_20260705_083435/S_f.pt \
  --s-r results/full_eval_20260705_083435/S_r.pt \
  --layer synthetic \
  --output results/diagnostic_run
```

Expected: `certificate.json` with `status: SURGERY_VIABLE`, `synthetic_spectrum.png` in output dir.

> Re-run `scripts/run_full_eval.py` first if `S_f.pt` / `S_r.pt` are missing locally (`results/` is gitignored).

---

## Full pipeline CLI

```bash
python scripts/run_carve.py \
  --base-model Qwen/Qwen2.5-1.5B-Instruct \
  --adapter results/syn2cap/entangled_adapter \
  --forget-data data/probes/D_f_train.jsonl \
  --retain-data data/probes/D_r_train.jsonl \
  --forget-holdout data/probes/D_f_holdout.jsonl \
  --retain-holdout data/probes/D_r_holdout.jsonl \
  --output results/carve_out
```

Results land in `results/<run_id>/` with `preregistration.yaml`, `metrics.yaml`, `report.md`.