# CARVE Progress Report — Publishable Syn-2Cap Suite

**Date:** 2026-07-17  
**Run ID:** `publish_live` (+ full-MLP confirmation if available)  
**Milestone:** Publication-oriented Syn-2Cap results pack  
**Claim under test:** Entangled LoRA capabilities can be removed **without full retraining**

---

## Executive summary

Publication suite on Qwen2.5-1.5B entangled Syn-2Cap **PASS**:

| Metric | CARVE (λ=1.5) | Target |
|--------|---------------|--------|
| FE | **1.00** | ≥0.90 |
| RF | **1.00** | ≥0.95 |
| vs Maat-SVD | Dominates (Maat RF=0) | Pareto |

Figures committed under `report/assets/`. Paper filled in `paper/main.tex`.

---

## Figures

| Figure | Path |
|--------|------|
| Ablation bars | `report/assets/syn2cap_ablation_bars.png` |
| FE–RF Pareto | `report/assets/syn2cap_pareto_fe_rf.png` |
| Eigenvalue spectrum | `report/assets/syn2cap_spectrum_publish.png` |
| Relearning curve | `report/assets/syn2cap_relearning.png` |

---

## Ablation (held-out)

Pre: forget=0.50, retain=0.75

| Method | forget_post | retain_post | FE | RF |
|--------|-------------|-------------|-----|-----|
| CARVE | 0.00 | 1.00 | 1.00 | 1.00 |
| Maat-SVD | 0.25 | 0.00 | 0.50 | 0.00 |
| Delete | 0.25 | 0.00 | 0.50 | 0.00 |
| Negation | 0.00 | 0.00 | 1.00 | 0.00 |

## λ Pareto (CARVE)

| λ | FE | RF | In target region? |
|---|-----|-----|-------------------|
| 1.0 | 1.00 | 0.67 | No (RF) |
| **1.5** | **1.00** | **1.00** | **Yes** |
| 2.0 | 1.00 | 0.67 | No (RF) |
| 3.0 | ~0 | 1.00 | No (FE) |

## Relearning after CARVE

| Steps | Forget acc |
|-------|------------|
| 0 | 0.00 |
| 5 | 0.25 |
| 15 | 0.25 |
| 30 | 0.75 |

Honest: behavioral removal is partially reversible with forget-only fine-tuning.

## Separability

`SURGERY_VIABLE`, max λ ≈ 1.68 at threshold 1.5.

## Artifacts

- `results/publish_live/publish_summary.json`
- `scripts/run_publishable_eval.py`

## Caveats

- Primary publish figures: `gate_proj` surgery (laptop VRAM)
- Full MLP at same λ=1.5: FE=1.0 but RF≈0.33 (over-cut) — needs higher λ / soft cut / repair
- Holdout n=4; fictional facts; not TOFU/Founder OS yet
- Relearning recovers forget behavior by ~30 steps

---

*Report 2026-07-17*