# H04_E_EXPLORATORY_FREEZE_v0_1

**Created:** 2026-09-13T22:05:06Z  
**Artifact:** `H04_E_EXPLORATORY_FREEZE_v0_1.json`  
**sha256:** `575470bdcb5baf8dc7085c10644f0eef27f1efba55cf6d3df8f374ff03f09dd2`  
**Hub job:** `77356229-193e-4693-a90c-cae8dcb0bc4a`  
**James:** post-H01-E-KILL ops — cheap freeze→run on RUNNABLE data; no FT900/ESR/AAR dig; no H01 reopen

## STATUS
| Field | Value |
|-------|-------|
| STATUS | EXPLORATORY |
| VINTAGE_STATUS | UNVERIFIED/INCOMPLETE |
| CLAIM_STRENGTH | LOW |
| CAPITAL_USE | PROHIBITED |
| MODEL_CHANGE | NO |
| REAL_CAPITAL | PROHIBITED |
| TARGETS_OPENED_AT_FREEZE | NO |

## Why runnable today
EIA `WCESTUS1` hist XLS + **141/141** STEO `*_base.xlsx` archives (jan2015–sep2026) fetched HTTP 200. Physical→STEO official-outlook lag. Avoids blocked archive lanes.

## Locked (all 10)

### 1. Eligible reconstructed sample
- **140** adjacent calendar-consecutive STEO pairs (jan2015→sep2026)
- Forecast dates: mix of PARSED + SCHEDULE_PROXY (EIA first-Tue-after-first-Thu rule when workbook lacks date)
- Every obs **VINTAGE_UNVERIFIED**

### 2. Predictor
`wcestus1_surprise_mean_inter_issue` — mean weekly Δstocks (kb) surprise vs training week-of-year norms.  
Release proxy = week_end + 5d.

### 3. Predictor window
`origin_forecast_date < release_proxy <= destination_forecast_date`.

### 4. Normalization
Training-only (destination years ≤2019). No holdout peek.

### 5. Target
`COSXPUS_dest[M] − COSXPUS_origin[M]` (mmbbl), M = destination year-month. **Values closed at freeze.**

### 6. Baseline (Model A)
Intercept + destination-month seasonality + prior revision (+ missing flag).

### 7. Chronological split
- Development destination years **2015–2019** (N=59)
- Holdout destination years **2020–2026** (N=81)

### 8. Metric
**RMSE** primary; MAE + directional accuracy secondary.

### 9. Success threshold
**≥ 5.0%** relative holdout RMSE improvement.  
**KILL** (<1% or neg / N<8) · **HOLD** ([1%,5%)) · **EXPLORATORY_PROMOTE** (≥5%; NOT capital)

### 10. Trial budget
ONE spec: Model A baseline; Model B = A + one WCESTUS1 feature. No rescue.

## Expected sign
**POSITIVE** (stock-build surprises → upward COSXPUS revision).

## Next
1. Post freeze to Hub (/docs + comment + /reply chatgpt+all). Never /james.
2. Only then open COSXPUS and run once.