REPRODUCIBILITY AMENDMENT: 6 OCTOBER 2026

PURPOSE
Make the existing final scientific evidence portable and accurately document verification. No new site, model, outcome definition, optimization objective, parameter search or scientific result is introduced here. The original final ZIP remains unchanged.

SOURCE AND CHAIN OF CUSTODY
Original: Supplementary_Data_1_Core_Reproducibility_Final.zip
Original size: 44,282,069 bytes
Original SHA-256: 4d55eae3e6be9a8f9089dda5d70c30542ab27c250991983ff045fab2c7b8cf80
All 179 original manifest entries were checked successfully. The repaired working copy had 190 manifest entries, also successfully checked before staging. Both historical manifests are retained under verification_20261006. See original_core_integrity.json and release_amendment_inventory.json. Current release contents are governed by the top-level MANIFEST_SHA256.csv.

CHANGES
1. robustness/round6_latency_build.py and round6_resource_driver.py resolve the core directory relative to the package instead of historical absolute paths. Their scientific logic and parameters are unchanged. Original and repaired hashes are recorded in evidence_audit_portability_changes.json.
2. Top-level guidance now distinguishes normalized-input controller reproduction from raw AEMO reparse. The former uses the bundled normalized CSV and does not require raw archives; the latter uses Supplementary Data 2.
3. This amendment adds a scientific evidence ledger, pinned dependency list, manifest verifier, clean-location analysis check, compact audit records and a current manifest. Bytecode caches are excluded.
4. The four historical auxiliary construction scripts are retained unchanged for provenance only. They contain historical paths/upstream dependencies and are not a promise of turnkey final manuscript/figure regeneration. New publication figure assets are supplied separately when applicable.

FRESHLY VERIFIED IN THE OCTOBER 6 AUDIT
- Statistical recomputation of all three archived analysis groups: no differences in scientific endpoints at absolute tolerance 1e-9 / relative tolerance 1e-8
- Independent physical replay of each controlled step and valid paired label for UK and Australian cohorts, including shared first commands and finite sampled scenario-target checks
- Independent reconstruction of the declared UK and Australian causal input pipelines under the simulation's availability assumptions
- Offline reparse of all 30 AEMO daily archives; normalized power, proxy power and snapshot/pair provenance outputs byte-identical to archived outputs
- Event-scoring sensitivity: every numeric table field reproduced, with 34,788 baseline label checks and zero mismatches
- Four nonbaseline DMPC obligation-scale reruns (two cohorts, scales 0.9 and 1.1), with no solver fallback and all 12 table rows matching at archived three-decimal precision
- Repaired 20-minute latency builder: checked arrays match, maximum absolute difference 3.50e-15
- Repaired resource driver at the Australian baseline: controller arrays match except measured solve time; the only differing ledger column is solve_seconds

WHAT WAS NOT FRESHLY RERUN
The entire altered latency/resource sensitivity grid was not independently rerun. Its complete reported grids remain archived results, with repaired baseline-equivalence checks. The main full --recompute workflow was not rerun end-to-end during this packaging amendment. Independent replay is a check of stored commands and physics, not a claim to have reoptimized every original controller run. Scenario-target replay checks sample 54 UK and 18 Australian origins per controller, with 32 scenarios and two branches; they are not an exhaustive scenario check.

PACKAGE-ACCEPTANCE CHECK
The portable package was exercised at a fresh location. All three analysis-only groups passed the frozen-hash and scientific comparison checks. Details, versions and checked fields are in verification_20261006/clean_location_analysis_verification.json and clean_location_analysis_comparison.json. The final ZIP was checked for integrity; split download parts reconstruct the exact ZIP and are governed by the separate restoration README and checksums.

INTERPRETATION LIMITS
The study concerns continuation-specific assessment in a declared simulation. UK development data were previously used. Australia is one farm over 18 days, frozen after a prior-exposure review, not pristine prospective blinding. The UK raw-turbine-to-aggregate provenance chain was not fully re-audited. Public archive creation times are not live delivery/receipt measurements. Synthetic obligations are not commercial-product validation. Paired moving-block ranges condition on the frozen model and observed cohorts; they do not include site, season, training or model-selection uncertainty. Mean probability agreement and Brier scores alone do not prove full calibration.

DOCUMENT CORRECTION
The Australian event-threshold gap range must include the frozen 0.01-pu baseline: -12.34 to -4.76 percentage points. The earlier manuscript's -12.10 lower endpoint was a prose error; the underlying scientific table is unchanged.
