Round 3 transparent causal input pipeline

Evidence boundary
This is a newly specified synthetic replay pipeline. It does not recover the missing legacy forecast generator, actual market forecast publication logs, or actual commitment issuance logs. Observation availability is explicitly assumed to equal the supplied ten-minute observation label plus 20 minutes. Synthetic commitments are issued one hour before delivery and use the power observation labelled 80 minutes before delivery. If original telemetry arrived later than this assumption, the replay would not establish real operational causality.

Frozen model
The original frozen model file is retained without modification.
SHA256 dc3d2569c73df9d87356a106cd9d0bbcc52985ac7a0185b9b47ccd2498c4f374
Fit protocol SHA256 16362cadd9843874889bd2901f7fe7d7d4caab46f55b094747df5918bfa9604d
Fit cutoff: 2021-09-09 16:00 UTC. Residual bank freeze: 2021-10-07 16:00 UTC.
Training origins: 3,995. Frozen bank origins: 4,020, exactly 670 per clock-hour phase.
No legacy point predictions, commitments or residual bank are read by the fit routine.

Information at decision time
Each origin t uses wind lags 2 through 25 (24 measurements), plus two UTC time-of-day features. All observation labels plus the 20-minute assumed delay are no later than t. Frozen train-only scaling and ridge alpha 10 generate 12 targets t through t+110 minutes, clipped to [0,1]. The 32 nearest residual paths are selected within the same clock-hour phase. The entire bank is frozen before development starts. A delivery hour h uses clip(power[h-80 minutes],0,1); its synthetic issuance is h-60 minutes. The current and next-hour commitments are therefore known at each decision origin.

Eligibility and missingness
Dispatch eligibility uses only same-episode, finite, already-available lag observations and synthetic commitment sources, plus the fixed calendar gate after bank freeze. Current or future outcome completeness never removes an eligible decision retrospectively. Each source episode requires 25 history rows before its first eligible origin. Missing observations are never interpolated or filled. Gaps create new source episodes. Retained origins are regrouped into contiguous dispatch episodes for state resets. An episode can start at a nonzero hour phase. Terminal truth_future entries remain NaN and horizon_mask is strictly retrospective scoring metadata. scoring_hour_complete identifies complete six-step clock hours; it does not gate dispatch. Fixed origin-start/end arguments select a cohort by time alone; prior measurements remain available as history, while only selected origins enter the exported controller input. Future truth outside the selected end is masked from scoring.

Data schema
The CSV apply input requires timestamp,power and accepts an optional episode column. Timestamp is UTC on a ten-minute grid; power is normalized but is not silently clipped. The output NPZ is pickle-free and retains point, plan, phase, episode, neighbors, residuals, truth, truth_future, horizon_mask and timestamps. Additional fields include source_indices, source_episode, scoring_hour_complete, dispatch_eligible, forecast_issue_ns, feature_latest_available_ns and commitment_issue_ns. All exported dispatch_eligible entries are true because the output contains only selected eligible origins. The dispatch ledger also lists excluded source rows and the separate calendar cohort selection flag.

Run from the experiment bundle root after arranging the referenced source input:
python sr_round3/causal_inputs/causal_pipeline.py develop --source-npz sr_reproduction/Wind_Literature_Baseline_Stage6/inputs/calibration_2021.npz --model-dir sr_round3/causal_inputs --output-dir sr_round3/causal_inputs/development

Apply frozen inputs to the authorized external temporal confirmation cohort (run only after the overarching confirmation protocol is frozen):
python sr_round3/causal_inputs/causal_pipeline.py apply --model-dir sr_round3/causal_inputs --csv sr_round3/confirmation/COOPGWF1_10min_normalized.csv --output-dir sr_round3/causal_inputs/confirmation --role confirmation --input-role external_temporal_confirmation --split confirmation_aemo_20260911_28 --name causal_confirmation_aemo_20260911_28.npz --origin-start 2026-09-10T14:00:00Z --origin-end 2026-09-28T14:00:00Z

Verification
Development validation changes observations unavailable by each cutoff and checks that earlier forecasts, commitments, residual-neighbor indices and eligibility do not change. It separately verifies fit and bank label availability. Independent protocol refitting matches the stored model to 1e-12 tolerance (maximum observed coefficient difference 9.14e-16; residual difference 2.73e-15); integer indices and timestamps are compared exactly. The stored frozen model is never overwritten by develop/apply. Apply performs additional information-flow mutations without computing control or forecast performance metrics. The 144-row time-window self-test uses already-exposed 2021 data, starts at phase 1, retains terminal partial truth horizons and passes. These checks establish code-level information flow under the declared availability assumption, not telemetry arrival evidence.

Implementation repairs before confirmation freeze
1. Use int64 for phase-to-nanosecond multiplication to prevent overflow.
2. Store timestamps as fixed Unicode rather than object arrays, permitting allow_pickle=False.
3. Add fixed time-window selection and causal apply validation.
No forecast hyperparameter, fitted coefficient, scaler or residual-bank value changed.

Post-freeze technical amendment 01
The first external apply attempt stopped during CSV timestamp parsing, before feature construction, forecasting or any outcome scoring. The frozen-v1 implementation called .asi8 on a pandas Series. The repaired implementation explicitly wraps parsed timestamps in a DatetimeIndex. The original code is preserved as causal_pipeline_frozen_v1.py. confirmation_amendment_01.json records the original confirmation freeze hash, old/new code hashes, unchanged scientific parameters/model/bank and a previously exposed 2021 array-to-CSV validation. Timestamp/episode/phase/lag-index checks are exact; feature numerical difference is at most 2.22e-16. The amendment was recorded before the successful external apply.

Completed external apply
The authorized cohort was applied once successfully after amendment 01; the earlier failed attempt generated no input or scores. confirmation_apply_execution.json records start/end UTC, freeze/amendment/source/model/code hashes. The output contains 2,592 origins in one episode, 432 complete clock hours, and six terminal origins with incomplete active future horizons retained for dispatch but excluded from complete-horizon scoring. Information-flow checks pass. The pipeline itself computes no forecast error or control outcome score. Controls and statistical interpretation are separate stages.

External source semantics
AEMO normalization and exposure evidence are in the adjacent confirmation directory. The normalized signal is a ten-minute proxy formed from two five-minute instantaneous SCADA snapshots, not measured ten-minute average energy. The generic input-manifest timestamp_semantics_verified=false deliberately does not certify telemetry arrival or the assumed +20-minute availability delay. The source documentation can establish snapshot time-label meaning; it cannot by itself verify that each observation was publicly available by the assumed decision deadline. The cohort is a bounded temporal confirmation identified after reviewing available prior artifacts, not proof that the researcher has never encountered any related station data.
