{
  "title": "H05_E_MECHANICAL_CONFOUND_FREEZE_v0_1",
  "created_at_utc": "2026-09-13T23:16:06Z",
  "hub_job": "77356229-193e-4693-a90c-cae8dcb0bc4a",
  "parent_artifacts": {
    "freeze": "H05_E_EXPLORATORY_FREEZE_v0_1",
    "freeze_json_sha256": "2f3fca7925519cf5a743488a89975b195f25c9df311d9ff5e648ac25c15b4292",
    "run_result": "H05_E_EXPLORATORY_RUN_RESULT_v0_1",
    "run_result_json_sha256": "27e7af66549e57661fc7bad4a7af06f67a2c74a1a804f165c25bbb0bbbd8597d",
    "prior_disposition": "EXPLORATORY_PROMOTE",
    "prior_holdout_pct_rmse_improvement": 28.48825157789054
  },
  "status_fields": {
    "STATUS": "EXPLORATORY",
    "VINTAGE_STATUS": "UNVERIFIED/INCOMPLETE",
    "CLAIM_STRENGTH": "LOW",
    "CAPITAL_USE": "PROHIBITED",
    "MODEL_CHANGE": "NO",
    "REAL_CAPITAL": "PROHIBITED",
    "TARGETS_OPENED_AT_THIS_FREEZE": "NO_NEW_TARGETS \u2014 reuse H05-E closed NGWGPUS revisions already in unit_table"
  },
  "purpose": "Cheap mechanical confound screen: is H05-E lift mostly STEO copying NW2/WNGSR weekly working-gas changes, or does the surprise feature add incremental information beyond a mechanical weekly-storage baseline?",
  "scope_locks": {
    "sample_split_target_metric": "IDENTICAL to H05-E freeze (140 units; dev dest years 2015-2019 N=59; holdout 2020-2026 N=81; y=NGWGPUS_dest[M]-NGWGPUS_origin[M]; holdout RMSE primary)",
    "MODEL_CHANGE": "NO",
    "trial_budget": "ONE primary confound comparison; no rescue; no alternate feature fishing",
    "parent_lanes": "H01/H01-E/H02/H03/H04 NOT reopened"
  },
  "primary_confound_check": {
    "id": "C1_NESTED_MECH_RAW_DELTA",
    "mechanical_feature": {
      "name": "nw2_l48_wg_mean_raw_delta_bcf",
      "definition": "Mean of admitted inter-issue weekly \u0394stocks_bcf (same release_proxy window as H05-E: origin_forecast_date < release_proxy <= destination_forecast_date; release_proxy = week_end + 6d). NOT demeaned by week-of-year. Mechanical NW2/WNGSR\u2192STEO inventory-update proxy.",
      "units": "bcf per week (mean)"
    },
    "tested_feature": {
      "name": "nw2_l48_wg_surprise_mean_inter_issue",
      "definition": "Unchanged from H05-E (mean weekly \u0394 minus training week-of-year norms)."
    },
    "models": {
      "A0": "H05-E Model A: intercept + dest-month dummies(1..11) + prior_revision + prior_revision_missing",
      "A_mech": "A0 + nw2_l48_wg_mean_raw_delta_bcf",
      "B_vs_mech": "A_mech + nw2_l48_wg_surprise_mean_inter_issue"
    },
    "primary_metric": "relative holdout RMSE improvement of B_vs_mech vs A_mech: 100*(RMSE_A_mech - RMSE_B_vs_mech)/RMSE_A_mech",
    "diagnostic_only_not_decision": [
      "A0 vs A_mech relative holdout RMSE improvement (how much of original lift raw delta alone captures)",
      "A0 vs B_original already reported in H05-E result (28.49%) \u2014 not recomputed as decision",
      "feature coefficient signs on holdout-fit is NOT used; training-fit coef on B_vs_mech surprise term reported for integrity only"
    ],
    "disposition_map_incremental_over_mechanical": {
      "EXPLORATORY_PROMOTE_SURVIVES_NONMECHANICAL": "incremental % >= 5.0 (surprise still clears 5% bar after mechanical control; CLAIM_STRENGTH remains LOW; NOT capital; NOT validated edge)",
      "HOLD": "1.0 <= incremental % < 5.0",
      "MECHANICAL": "incremental % < 1.0 or negative OR N_holdout < 8 OR mechanical feature train variance == 0",
      "note": "MECHANICAL / HOLD downgrades prior EXPLORATORY_PROMOTE for claim framing; does not reopen H05-E arithmetic."
    },
    "estimator": "numpy.linalg.lstsq OLS, same as H05-E reproduce_metrics_from_unit_table.py"
  },
  "explicitly_not_in_this_screen": [
    "No capital / trading / validated-edge claims",
    "No MODEL_CHANGE",
    "No first-print / vintage repair",
    "No archive archaeology",
    "No H01/H04 reopen",
    "No secondary feature tournaments (latest level, residualization variants) in THIS freeze \u2014 reserved only if C1 is ambiguous and a NEW freeze is authorized"
  ],
  "prospective_protocol_sketch_companion": "H05_E_PROSPECTIVE_PROTOCOL_SKETCH_v0_1 (design only; no outcomes) \u2014 written after result if useful",
  "plain_english_pre_result": "Before opening confound metrics: we ask whether H05-E's NW2 surprise feature still beats a baseline that already knows the raw weekly Lower-48 working-gas changes STEO would mechanically copy from WNGSR/NW2. If not, the click is mechanical."
}