# Prop-01 E1b patch protocol — sample-efficiency + clean-fail + A3 ablation

| Field | Value |
| --- | --- |
| **Protocol id** | `LAX-PROP-2026-09-04-01a-E1b` |
| **Parent** | E1 full results + §5 scorecard at `writing/prop01-e1-section5-scorecard.md` (+ `.json`) |
| **Parent artifact** | `sims/sr-ladder/results/prop01_live_e1_full.json` / `.md` (1920/1920, `partial=false`) — **do not overwrite** |
| **Amends / extends** | Active campaign under `LAX-PROP-2026-09-04-01a`; §5 thresholds frozen from `LAX-PROP-2026-09-04-01` |
| **Authorized** | Michael Adle via CoS recommendation (2026-09-05/06) — **E1b patch only** (not full blind 1920 redo) |
| **Holds** | Prop-02 GNN **held**; X freeze **kept**; `prop01_support_claim` **never flipped** by this protocol alone |
| **Date** | 2026-09-06 |

---

## Locked rule (support claim)

`prop01_support_claim` stays **false** until **ALL** §5 gates Pass on a re-score of this same scorecard template against **E1b measured artifacts**. Completing E1b cells ≠ support. STLS-surrogate A3 (if used) **never** flips support alone.

---

## Parent E1 §5 status (measured honesty — do not re-litigate)

From `writing/prop01-e1-section5-scorecard.md`:

| Gate | Verdict |
| --- | --- |
| §5.1 Sample efficiency | **Fail** (0/7 clear ≤0.5×; need ≥5/7) |
| §5.2 Noise robustness | **Fail** (`damped_oscillator` mid-noise A4 recovery 0 vs A0 0.65–0.85) |
| §5.3 Clean fail | **Fail** (A4 `dim_violating` clean-fail 0.7833 < 0.9) |
| §5.4 Non-triviality | **Incomplete** (no A3 arm in E1 grid) |
| Flip `prop01_support_claim`? | **No** |

E1b goals are locked to these failed / incomplete gates only.

---

## Goals (locked to failed gates)

### 1. N-grid restore for sample-efficiency resolution (§5.1)

- **N grid:** `{25, 50, 100, 200, 400}` — restore **200** (dropped in E1-reduced), add **25**
- **Seeds:** `0–19` (≥20)
- **Noise:** `σ ∈ {0, 0.05}`
- **Arms:** `A0_pysr` + `A4_physo` (lit backends)
- **Matched wall:** arm-mean wall within **±20%** (same band as E1 / 01a)
- **Soft-pad policy (document in run receipt):** state explicitly whether soft-pad is ON or OFF; if ON, record pad seconds and which arm was padded; if OFF, note that matched-wall may fail and cannot alone authorize support. Prefer documenting the chosen policy in the E1b results header (`soft_pad`, `mean_soft_pad_s`, `matched_wall`).
- **Purpose:** make ≤0.5× N ratios identifiable when both arms already saturate at N=50 on the reduced grid.

### 2. Damped-oscillator repair lane (§5.2)

- Raise **A4 / PhySO** budget (epochs / `max_n_evaluations` or equivalent) under the **SAME** wall-cap matching rule (±20% arm means).
- Re-measure **mid-noise** (`σ=0.05`) recovery on `damped_oscillator` until either:
  - **Pass path:** A4 ≥ A0 − 0.05 per `(law, N)` for `damped_oscillator` across the N grid, **or**
  - **Kill path:** document irreducible fail vs kill criterion (drop > 0.2 systematic) and stop the repair lane without flipping support.
- Do not loosen the −0.05 margin or redefine mid-noise.

### 3. Clean-fail push on `dim_violating` (§5.3)

- Keep **R\* = 0.25** fixed (metrics lock / `methods_metrics.md`).
- Target: A4 clean-fail rate on `dim_violating` **≥ 0.9** (measured not-accepted fraction / aggregate `clean_fail_rate`), not accept-rate cosmetics.
- Do not lower the 0.9 threshold or raise R\*.

### 4. A3 conservation-only for §5.4 only

- **Preferred:** add **A3 conservation-only** arm if PhySO partial / conservation-only mode exists on the lit path.
- **Else:** labeled **STLS-surrogate A3** for §5.4 non-triviality scoring **only**.
- Surrogate A3 **never** flips `prop01_support_claim` alone; lit A3 preferred when available.
- Criterion unchanged: conservation-only must **not** achieve §5.1–§5.2 via constant/trivial identities alone.

---

## Out of scope / freezes

- **Not** a full blind redo of the E1 1920 grid; patch lanes above only.
- **Do NOT overwrite** E1 full artifacts (`prop01_live_e1_full.*`, figures tied to E1 full unless separately versioned).
- **Hold Prop-02 GNN** (no GNN primary start from this protocol).
- **Keep X freeze** (no auto-post / claim flip on social).
- **Never flip** `prop01_support_claim` in runners, findings, or status from E1b setup or partials.

---

## Deliverables (new paths only)

Prefix: `sims/sr-ladder/results/prop01_live_e1b_*`

Suggested set (writers may add siblings; do not clobber E1):

| Path | Role |
| --- | --- |
| `sims/sr-ladder/results/prop01_live_e1b_full.json` | Primary measured aggregate + records |
| `sims/sr-ladder/results/prop01_live_e1b_full.md` | Human receipt |
| `sims/sr-ladder/results/prop01_live_e1b_full.checkpoint.json` | Optional checkpoint |
| `sims/sr-ladder/results/prop01_live_e1b_reduction.md` | Grid / soft-pad / wall policy lock |
| `sims/sr-ladder/results/prop01_live_e1b_*.log` | Run logs as needed |

Scorecard re-score (after artifacts land): update or sibling under `writing/prop01-e1b-section5-scorecard.md` / `.json` — **not** a silent rewrite that erases E1 fail history; keep E1 scorecard as parent measured honesty.

---

## Re-score / flip rule

1. Run E1b to completion (or documented kill on repair lane).
2. Re-score §5.1–§5.4 against E1b measured numbers using the same gate definitions.
3. Flip `prop01_support_claim` to **true** **only** if gates 1–4 all **Pass** with lit backends, matched compute, ≥20 seeds.
4. Until then: claim stays **false**; publish measured honesty only.

---

## Authorization receipt

- CoS recommendation authorized by Michael: E1b patch (not full blind 1920 redo); publish E1 §5 scorecard as measured honesty; hold Prop-02 GNN; keep X freeze; never flip `prop01_support_claim`.
- Deploy of site feeds is a separate parent step — this protocol file does not deploy.

---

## Experimental Design lock (2026-09-05)

**Reviewer:** Experimental Design  
**Verdict:** **LOCKED** — no protocol blockers against frozen Prop-01a §5.

### Confirmations

| Item | Status |
| --- | --- |
| §5 numeric thresholds unchanged (0.8 recovery, ≤0.5× N, −0.05 noise margin, clean-fail ≥0.9, R\*=0.25) | **OK** — no δ/gate edits |
| `prop01_support_claim` stays **false** until E1b re-score clears **all** §5.1–§5.4 Pass | **OK** (Locked rule + Re-score / flip rule) |
| A3 closes **§5.4 non-triviality only**; STLS-surrogate A3 never flips support alone | **OK** |
| Prop-02 GNN / fingerprint tracks **unchanged**; GNN primary **held** | **OK** |
| E1 full artifacts immutable (`prop01_live_e1_full.*`); E1b writes only `prop01_live_e1b_*` | **OK** |
| N∈{25,50,100,200,400} is **resolution expansion only** — success definition identical to frozen §5.1 | **OK** (authorized patch, not silent goalpost move) |

### Soft notes (not blockers)

1. Soft-pad policy must be stated in the E1b results header before scoring; soft-pad / matched-wall alone cannot authorize support (already in Goals §1).
2. Damped repair “per `(law,N)`” pass path is stricter than pooled-mean reading of §5.2 — keep it; it matches the E1 scorecard’s systematic-rejection fail mode and does **not** loosen −0.05.
3. Prefer lit A3 when PhySO partial mode exists; surrogate A3 is scoring-only for §5.4 Incomplete close.

**Action for Sim Eng:** proceed under this file. Flip claim only after `writing/prop01-e1b-section5-scorecard.*` shows gates 1–4 all Pass.