nx_patch_attribution.nx

nx_patch_attribution.nx -- the COMPOST primitive. User: "did this patch improve things if so then lets find out why and test and integrate like how we are using a lora right now that improves our elder ai realism image gen but i couldnt tell you why and neither can the system". THE answer. Given a population of "before-patch" samples and a population of "after-patch" samples, each with per-axis measurements, this primitive reports: - WHICH measurement axes the patch significantly moved - HOW MUCH each axis moved (Cohen's d effect size) - WHETHER the movement is statistically significant after family-wise alpha correction (Bonferroni) The substrate's bits-up answer to the user's concrete pain: don't ship LoRAs nobody understands; ship a decomposition that turns the LoRA's effect into measurable named-axis changes that can then be COMPOSTED into substrate architecture per the patches-as-manure cardinal. ===== Workflow =================================================== 1. Run N generations WITHOUT the patch; collect per-axis measurements via caller's measurement_fn. Call this sample population B (before). 2. Run N generations WITH the patch; collect same measurements. Call this population A (after). 3. nx_patch_attribute(B_data, A_data, n_axes, n_samples) 4. Result lists per-axis: mean_before, mean_after, effect_size, significant? Tells caller EXACTLY what the patch is doing. 5. Caller then chooses: is this a useful axis movement? If yes: ship a substrate-side primitive that achieves the same axis movement without the LoRA. COMPOSTED. ===== Statistical methods ======================================= Cohen's d (Cohen 1988): standardised mean difference. d = (mean_A - mean_B) / pooled_std