# Storage chronology diagnostic: independent internal audit

**GO for the sealed v2 nominal-calendar model results. No remaining numerical or method blockers were found.** This is an internal Codex review by a separate agent, not external peer review or independent validation of RTE measurements.

Across ten annual nominal profiles, the median ratio of original-order capacity to the median shuffled-order reference capacity was **1.1074: 10.7% higher**. All ten primary ratios exceeded 1, ranging from 1.0269 to 1.2781. This is the median of ten year-specific ratios, not a pooled estimate of France’s storage needs.

## What was checked

The review verified the plan, execution approval, code and raw-input hash bindings and the order of plan commitment, code approval and execution. The original physical-timeline plan aborted before capacity outcomes. Its replacement was committed and reviewed before the single v2 run.

- All 50 original scenario/year capacities and exact zero energy balances were independently recomputed using reduced rational supply coefficients and direct slot-by-slot cumulative sums.
- Seven permutation cases, fixed before outcomes, were independently regenerated. Exact capacities, seeds and day-order hashes matched.
- All 49,950 retained permutations had the expected unique replicate IDs, seeds and unit conversions. Every condition’s quantiles, medians, ratios, differences, inclusive reference ranks and descriptive Holm adjustments matched independent aggregation.
- All 18,245 daily exact summaries and 18,245 cumulative endpoints matched an independent calculation over 875,760 scenario slots. Exclusions, clipping counts, monthly residual sums, marginal residual hashes, means, variances and peaks matched.
- CSV and JSON copies of all 49,950 permutation rows, 50 condition rows and five across-year summaries agreed. The combined result file agreed with its constituent files.

The two audit scripts recorded **830,790 successful comparisons and zero discrepancies**. The reviewer did not regenerate all 49,950 permutation capacities; seven were replayed directly, while all retained rows and summaries were checked as described above.

## Results and required interpretation

| Scenario | Median original/reference ratio | Annual profiles above reference median |
|---|---:|---:|
| Primary 70% wind / 30% solar | 1.1074 | 10/10 |
| 50% wind / 50% solar | 1.0558 | 10/10 |
| 90% wind / 10% solar | 1.1568 | 9/10 |
| Both DST transition dates removed | 1.0972 | 10/10 |
| Negative reported generation clipped before normalization | 1.1089 | 10/10 |

The 90/10 sensitivity retains an exception: the 2016 ratio was 0.9840. The complete sensitivity set should remain available. Six primary original capacities fell outside the central 95% conditional permutation reference interval; these intervals and reference ranks are descriptive, not calibrated causal or randomization p-values.

Every empirical capacity belongs to the declared model: unique source date/hour labels, 48 slots per nominal day, each assigned a 0.5 nominal-hour weight; hypothetical annual energy-balanced renewable scaling; lossless storage; unlimited power; free initial state of charge; and a separate cyclic boundary for each annual profile. Permutations retain complete within-day matched profiles and month membership. They change between-day order within months, not seasonal order or all temporal structure.

The source’s derived UTC timestamps contain two duplicated spring actual instants and two absent autumn actual instants in every year. They were not repaired. Therefore “nominal MWh-equivalent” and “nominal mean-demand hours” are model accounting units, not verified historical French energy or physical storage requirements. Signed reported solar values remain in the primary input; their cause has not been established. The separately named clipping and date-removal sensitivities do not convert the source into a verified physical timeline.

The established cyclic formula is valid: if cumulative residual energy is D[k] and total residual energy is exactly zero, any feasible capacity must span max(D)−min(D); choosing initial charge max(D) achieves that bound. The independent control [1,1,−1,−1] MW versus [1,−1,1,−1] MW at 0.5 hours gives 1 versus 0.5 MWh, with equal mean, variance and peaks. The empirical experiment illustrates chronology sensitivity at an identical marginal residual distribution. It does not establish a new theorem or refute the original mechanism, which already explicitly includes duration.

## Reproducibility bindings

- Plan v2 SHA-256: `07add03f2e9de805f1170f0aebaa29dfb157443b5d93d83a877f3ffde1ddaa53`
- Runner SHA-256: `2bcb9e611039bb450cb70aa072f03f32d65b9fea9f4d9a95bf99b0fe23fdcaec`
- Execution: 30 October 2026, 17:18:53–17:19:12 UTC; 50 conditions × 999 permutations.
- Ten downloaded annual source CSV hashes and byte counts matched their acquisition receipts and were unchanged through review.

Detailed machine-readable findings are in `storage-independent-audit.json`. Independent arithmetic, preflight, replay selections and the two detailed result audits are preserved in the accompanying research package’s internal-review materials. Public text and figures must retain the nominal-model scope above.

## Reader-facing check

The prepared overview, full report and Blogger text, plus the four SVGs’ text labels, retain the nominal-model scope, the median-of-ratios interpretation and the 90/10 exception. They distinguish descriptive reference comparisons from causal inference and internal review from external peer review. No material wording blockers were found. This was a text/label check, not a rendered-layout or literature-citation audit. The Blogger statement that the files are public must follow verification of the linked public packet.