# Internal Codex pre-execution data review

This is an internal independent implementation review, not external peer review. No storage-capacity outcomes have been calculated in this review. Plan v1 remains preserved; the empirical experiment must not run under its failed gates.

## Binding plan and scope

Plan v1 SHA-256: `6f3df8f2d36d2ac74216174df51430a6c679daa80a40e04ede8b136a7cb41b14`. Its exact balanced-residual formula and cyclic, lossless, unlimited-power minimum-capacity calculation are mathematically valid. Its original gates require a contiguous physical Europe/Paris timeline and nonnegative generation inputs. Both fail on the supplied source exports. This does not invalidate the formula, establish a result about the theory, or make the supplied source unusable for a narrower declared experiment.

## Independently verified input findings

All ten raw CSV hashes/byte counts match acquisition receipts. The exports contain 350,688 rows: 175,344 complete actual triples and 175,344 rows with all three actual fields blank. The complete actuals have integer MW values, positive demand, and definitive-data labels. All ten source years contain every unique nominal `(date, heure)` label once: 96 quarter-hour labels and 48 actual half-hour labels per nominal date, including both daylight-saving transition dates. There are zero absent, extra, or duplicate nominal labels.

Every year nevertheless has two duplicate spring actual UTC instants and two absent autumn actual UTC instants. Every year also has four spring quarter-hour labels inconsistent with Europe/Paris conversion of `date_heure`, including two actual half-hours. For example, the 2021 source labels 02:00 and 03:00 on 28 March both map to 01:00 UTC and contain the identical actual triple (51,860 MW demand, 3,474 MW wind, 1 MW solar). The 02:30 and 03:30 labels likewise map to 01:30 UTC and contain (52,090, 3,537, 1). The absent autumn actual instants are 31 October 2021 00:00 and 00:30 UTC. Annual row totals alone therefore cannot establish complete physical-time coverage.

The solar field contains 36,147 negative reported actual values, all in 2015–2020. Annual minima are respectively -2, -2, -2, -2, -23, and -3 MW. Wind and demand have no negative samples. The numerical source sample sum over negative solar values is -37,704 MW; this is a sum of sample values, not a historical energy integral. No source values have been clipped, interpolated, or dropped by this review. No explanation for the negative values has been established.

## Minimal acceptable revision routes

1. Prefer authoritative source clock documentation or an alternate data export that resolves a physical timeline. Any repair must be documented and justified before computing capacity outcomes. Do not infer missing values or silently collapse duplicate UTC rows.
2. If the physical clock cannot be resolved, a separate explicitly nominal-calendar diagnostic is mathematically valid. Sort unique source date/time labels, assign each actual slot a declared 0.5-hour model weight, and state that all days have 48 slots. The resulting weighted annual energy and storage capacities belong to this constructed model. They are not exact historical French annual energies or actual physical timeline estimates. Both headline and figure captions must carry this scope. A same-length-day control becomes identical to the unrestricted monthly day control in this dataset and must not be presented as an independent DST sensitivity.
3. Retaining signed solar values is acceptable only as an explicit signed reported input series, with positive annual source sums verified. Do not call every sample nonnegative production. Clipping would be a separately declared transformation/sensitivity; do not explain negatives as rounding or net accounting without evidence.

The analytic four-slot counterexample and the cyclic SOC proof remain valid independent of any empirical dataset. They show that variance and peaks alone do not determine energy capacity. The original hypothesis already discusses duration, so this is not a refutation of its whole mechanism and not a new theorem.

## Evidence

- `independent-data-preflight-v1.json`: source hashes, per-year signs and integrity gates.
- `independent-calendar-diagnostic.json`: all nominal-label checks, clock mismatches and UTC duplicate/missing examples.
- `independent-analytic-controls.json`: exact independent analytic controls.
- [RTE original file specification](https://assets.rte-france.com/prod/public/2020-07/%C3%A9CO2mix%20-%20Description%20des%20fichiers%20des%20donn%C3%A9es%20en%20puissance.pdf), pp. 3–5, describes 96 daily rows, MW fields, and half-hour definitive actuals versus quarter-hour forecasts. It does not resolve the observed clock discrepancy or explain negative solar.