# H05_E_MECHANICAL_CONFOUND_FREEZE_v0_1

**Created:** 2026-09-13T23:16:06Z  
**Artifact:** `H05_E_MECHANICAL_CONFOUND_FREEZE_v0_1.json`  
**sha256:** `4a9d2e501d375a5d9c0a9296da12bc9f571dba9b3fa3f850307fbbd36864da2a`  
**Hub job:** `77356229-193e-4693-a90c-cae8dcb0bc4a`  
**Parent:** H05-E EXPLORATORY_PROMOTE (freeze sha `2f3fca79…`; result sha `27e7af66…`; ~28.49% holdout RMSE)

## STATUS
| Field | Value |
|-------|-------|
| STATUS | EXPLORATORY |
| VINTAGE_STATUS | UNVERIFIED/INCOMPLETE |
| CLAIM_STRENGTH | LOW |
| CAPITAL_USE | PROHIBITED |
| MODEL_CHANGE | NO |
| REAL_CAPITAL | PROHIBITED |
| NEW TARGETS | NO (reuse H05-E unit_table NGWGPUS revisions) |

## Purpose
Cheap mechanical confound screen: is the H05-E lift mostly **STEO copies NW2/WNGSR weekly working gas**, or does the surprise feature add incremental information?

## Locked primary check (ONE) — C1_NESTED_MECH_RAW_DELTA
Same sample/split/target/metric as H05-E.

| Model | Spec |
|-------|------|
| A0 | H05-E baseline A (intercept + dest-month + prior revision) |
| A_mech | A0 + `nw2_l48_wg_mean_raw_delta_bcf` (mean admitted inter-issue weekly Δstocks; **not** seasonally demeaned) |
| B_vs_mech | A_mech + `nw2_l48_wg_surprise_mean_inter_issue` (original H05-E feature) |

**Primary metric:** relative holdout RMSE improvement of B_vs_mech vs A_mech.

### Disposition map (incremental over mechanical)
| Outcome | Rule |
|---------|------|
| EXPLORATORY_PROMOTE_SURVIVES_NONMECHANICAL | incremental % ≥ 5.0 (CLAIM_STRENGTH still LOW; not capital) |
| HOLD | [1%, 5%) |
| MECHANICAL | <1% or negative / N<8 / zero mech variance |

Diagnostic only (not decision): A0→A_mech lift (how much raw delta alone captures).

## Explicitly out of scope
No capital, no MODEL_CHANGE, no vintage repair, no archive dig, no H01/H04 reopen, no secondary feature tournament in this freeze.

## Next
1. Hash-lock this freeze.
2. Build `nw2_l48_wg_mean_raw_delta_bcf` on the existing admitted windows; run C1 once.
3. Post result (+ short prospective protocol sketch, design only).
4. Hub comment + /reply chatgpt+all. Never /james.