# CARVE Progress Report — Complete Proof Suite

**Date:** 2026-07-17  
**Run ID:** `complete_proof_20260717_152414`  
**Claim:** Entangled LoRA capability removal **without full retraining**

---

## Executive summary

`all_behavioral_pass: true` across four experiments:

| Check | Decision | FE | RF | Notes |
|-------|----------|-----|-----|-------|
| Stronger holdout | PASS | 1.0 | 1.0 | n=8 paraphrases |
| Full MLP | PASS | 1.0 | 1.0 | λ=2.0, soft_γ=1.0 on forget dirs, repair=20 |
| Founder OS sample path | PASS | 1.0 | 1.0 | Probe builder E2E; Syn-2Cap stand-in adapter |
| TOFU mini | PASS* | 1.0 | 1.0* | *weak pre-retain; forget wipe real |

Figure: `report/assets/proof_completion_bars.png`

---

## Full MLP fix

Hard cut on all MLP modules over-cut retain. Fixed by:
1. Selective soft-shrink **only** on high-λ directions (`apply_soft_cut(..., cut_indices=J)`)
2. Short retain-only LoRA repair (20 steps) — **not** full adapter retrain

Winning config: λ=2.0, soft_γ=1.0, repair=20 → forget 0.5→0.0, retain 0.75→1.0; Maat RF=0.

---

## Artifacts

- `results/complete_proof_20260717_152414/complete_proof_summary.json`
- `scripts/run_complete_proof.py`

## Remaining (optional / production)

- Real Founder OS adapter weights
- Official TOFU/WMDP evaluation protocol + larger N
- Multi-seed confidence intervals, 7B

---

*Report 2026-07-17*