DEVELOPMENT INTERPRETATION BEFORE RESERVED GEOMETRIES
Recorded UTC: 2026-10-14T04:31:45.539493+00:00
Status: local record; not a public preregistration.

The full development bank (7,800 parents; zero failed calculations) and all prespecified development summaries have been examined. No reserved H1/H3 parents have been generated or examined. The original protocol, configuration, primary estimator, k values, targets, sample sizes, replicate counts, five-subset rule and 9,999-bootstrap rule remain unchanged.

Interpretation fixed for the reserved stage:
1. Permutation centring does not generally calibrate information magnitude or between-sample differences. At I=0.4, centred KSG4 Q1x means are 0.072188 (N=78) and 0.154254 (N=500), both far below truth. Its mean constant-truth profile rises by 0.082066 nats.
2. In the declared I_A=0.4, I_B=0.6, N_A=500, N_B=78 comparison, Q1x baseline difference error is about -0.2752 nats, so the mean estimated difference has the opposite sign to the true +0.2. Correct ordering occurs in 13.0% of 200 replicates. Common-78 improves that fraction to 55.0%, but its mean difference error remains about -0.1907 nats. This is a setting-specific result, not a universal rate.
3. At I=0.4, common-78 profiles satisfy the reporting tolerance for all six development designs, while baseline profiles do so only for L1. A flat profile does not establish correct absolute values. For Q1 at N=500, the baseline mean 0.38262 and RMSE 0.04943 become 0.27766 and 0.13389 under common-78. Subsampling can therefore worsen absolute calibration while improving unequal-size comparisons.
4. Equal-information, equal-dimensional geometry changes retain nonzero comparison errors at common sizes. At N_A=N_B=500 and I=0.4, the L3-to-T3 difference is -0.03940 nats and Q1x-to-Q3 difference is +0.04340. Their intervals overlap the tolerance boundary; neither is labelled outside tolerance on that basis. Equality of information does not imply equality of finite-sample bias.
5. GC's inability to represent symmetric Q/H dependence follows from its population covariance target; a finite-sample correction is not a remedy for omitted nonlinear structure.
6. In the 102 development baseline KSG4-centred sensitivity cells, the largest absolute mean B500-minus-B100 change is approximately 0.002866 nats; this limited baseline sensitivity is not a test of P1 at B500.

No data-dependent procedure or k is selected. The reserved stage applies the same fixed procedures to H1/H3. It will assess whether the sample-size, dimensionality and common-size tradeoffs extend to an absolute-value geometry, retaining all outcomes including exceptions or indeterminate units. Claims will remain bounded by the tested settings and will distinguish average calibration from individual ordering. The reserved findings will not be used to retune or rerun the development study.

Numerical references: results/analysis_development/*.csv and results/development_key_results.json. All full-grid summaries are retained; the examples above do not replace them. The initial analysis watcher ended without outputs; the unchanged deterministic analysis was successfully rerun in the pinned runtime, recorded in results/analysis_execution_log.jsonl.
