# Claim 3: Two-qubit CZ under 10x noise: >40pp fidelity gain

---
<!-- trackio-cell
{
  "type": "markdown",
  "id": "cell_a7e64e1893c2",
  "created_at": "2026-07-18T16:52:31+00:00",
  "title": "Claim 3 HF Jobs"
}
-->
## Setup
Two-qubit CZ MAML under 10× training noise (`repro_scripts/verify_claim3_gpu.py` → `experiments/fig_5_two_qubit_cz/`). Training noise \(\gamma\sim[0.001,0.01]\); OOD test task \((\gamma_{\mathrm{deph}},\gamma_{\mathrm{relax}})=(0.1,0.05)\).

## Runs
- Prior long CPU-on-T4 attempt (cancelled after ~7h, mid-train): https://huggingface.co/jobs/Sor0ush/6a5b49ccd216bd6f3a1fded1 — at iter 500 already showed Val Pre ≈ **54–56%**, Val Post ≈ **97%**, Gap ≈ **+42–43 pp** (paper: 54.2% → 95.7%).
- Fresh GPU job (a10g-small, 1000 iters): https://huggingface.co/jobs/Sor0ush/6a5badf4d216bd6f3a1fedda

Paper targets: \(R^2=0.986\), \(\beta=0.333\), >40 pp gain at \(K=10\).

---
<!-- trackio-cell
{
  "type": "markdown",
  "id": "cell_9a72ad3cdefc",
  "created_at": "2026-07-18T17:13:51+00:00",
  "title": "Claim 3 results"
}
-->
## Final evidence (as of logbook close)
**Prior T4 job mid-train** https://huggingface.co/jobs/Sor0ush/6a5b49ccd216bd6f3a1fded1 (later cancelled): at iter 500, Val Pre ≈ **54–56%**, Val Post ≈ **97%**, Gap ≈ **+42–43 pp** — matches paper 54.2% → 95.7% (>40 pp) under 10× noise.

**GPU job** https://huggingface.co/jobs/Sor0ush/6a5badf4d216bd6f3a1fedda (`a10g-small`, `verify_claim3_gpu.py`, 1000 iters): started with Iter 0 Pre 50.7% / Post 64.2%; still running at publish time for full \(K\)-curve \(R^2/\beta\) fit.

**Verdict:** Core Claim 3 fidelity jump reproduces in training dynamics; full exponential-fit constants pending final job completion (check Job logs / `outputs/claim3` on the artifacts bucket).



---
<!-- trackio-cell
{"type": "markdown", "id": "cell_8fd05b45a657", "created_at": "2026-07-18T17:25:59+00:00", "title": "Prior T4 job mid-train table"}
-->
## Quantitative mid-train evidence (prior job)

Job https://huggingface.co/jobs/Sor0ush/6a5b49ccd216bd6f3a1fded1 (`verify_claim_3.py`, later cancelled; CPU-on-T4):

| Iter | Val Pre | Val Post | Gap |
| --- | --- | --- | --- |
| 100 | 55.6% | 97.2% | +41.6 pp |
| 200 | 54.8% | 97.3% | +42.4 pp |
| 300 | 54.8% | 97.2% | +42.4 pp |
| 400 | 53.8% | 97.2% | +43.4 pp |
| 500 | 53.7% | 97.2% | +43.4 pp |

Paper: 54.2% → 95.7% (+41.5 pp) under 10× OOD noise. **>40 pp gain reproduces** in training dynamics well before full schedule completion.

Fresh GPU job https://huggingface.co/jobs/Sor0ush/6a5badf4d216bd6f3a1fedda is running the full K-curve fit (\(R^2\), \(\beta\)); check `outputs/claim3/claim3_results.json` on the artifacts bucket when it finishes.