Skip to content

[P1 CPU / existing data] After #702: within-model quantile-shape symmetry and full-vector finite-size motion #622

Description

@LightChainr

Re-scoped 2026-09-12 — one analysis, not another probe portfolio

The previous W1–W5 brief is superseded by merged #702, notes/exact-foundations-correction-20260912.md §3. Keep the valid location-without-shape constructions and the repaired source data. Do NOT rerun their proofs, repeat tiny censuses, test the invalid raw-p C3 of #666, or claim the two Aff(1) actions fail to intertwine.

Decision this task must change

Do the existing square-site, same-observable productions show a resolved departure from reflection-symmetric normalized shape, and does that departure decrease along an actual size lineage? Separately, how much non-affine shape displacement remains between adjacent sizes? This decides whether a further observation is needed and what it must separate; it does not name an exponent or field.

Inputs and resource envelope

Use one ordinary CPU machine with a complete repository checkout; no GPU and no new Monte Carlo. First profile one existing batch to measure runtime. Cap this first delivery at 4 CPU-hours; if the full delete-one path cannot fit, return the measured bottleneck and completed source audit, not a substituted estimator. Machine assignment by the team; do not power on a fleet.

Base: current main including #702; use #703's support policy after it is integrated, or record its exact reviewed head. Read #633/#655 for corrected cos(4theta), per-orientation inversion order and pooled/delete-one semantics. Reuse scripts/threshold_quantile_lineage.py from the #655 lineage if needed, with its source revision recorded.

Primary size lineage: N=145,290,725 from the P50/#612 assets. N725 raw: results/server-20260907/P612-n725-fullcurve/raw/; N145/N290: results/server-20260829/P50-n145-n290-fullcurve/. Resolve actual filenames from the checkout, do not invent missing paths. If one raw block is unavailable, return which and perform only identifiable comparisons. Additional already-committed lineages are sensitivity, not required for this first delivery.

Fixed objects, before examining new scores

  1. Reconstruct each orientation's pooled inverse CDF Q first, then form the declared spin-0 weighted Q. Check weights sum to one and sum(w*cos4)=0. Do not mix CDFs before inversion. Report equal weighting separately, from the SAME block.

  2. Primary grid: nine deciles 0.1,...,0.9. Anchors a=0.2,b=0.8. W_N=Q_N(b)-Q_N(a)>0; Z_N(u)=[Q_N(u)-Q_N(a)]/W_N.

  3. Shape reflection residual:

    A_N(u) = Z_N(u)+Z_N(1-u)-1
    = [Q_N(u)+Q_N(1-u)-Q_N(a)-Q_N(b)]/W_N.

    Do not replace this by M_N(p)+M_N(1-p). The latter is a reflection about raw p=1/2 and conflicts with a location limit centred elsewhere.

  4. Keep pooled estimates primary. For each aligned delete-one batch, repeat the entire Q -> weighting -> W -> Z -> A transformation. Retain the joint covariance. Cross-size independence/coupling follows recorded random-stream provenance, not matching batch IDs.

  5. Symmetry/anchor identities cause deterministic redundancies. Use independent coordinates, e.g. A at u=0.1,0.3,0.4,0.5 (u=.2 is zero; reflected pairs repeat). Do not interpret a noisy covariance rank as a dimension of the mean law, or discard exact contradictory residuals with a pseudoinverse.

Deliverable

A source-resolved JSON plus one short note and the reusable analysis script. Report:

  • N and geometry/lineage, site count versus shortest period, spin0 weights;
  • pooled Q,W,Z,A and full delete-one covariance;
  • at each size, the covariance-aware A=0 diagnostic on nonredundant coordinates;
  • adjacent-size Delta Z and Delta A with correctly propagated covariance; test zero displacement, not a fitted asymptotic exponent;
  • measured changes in an explicitly declared norm, with uncertainty. Three sizes cannot establish convergence; distinguish observed finite movement from a limit statement;
  • equal-weighting sensitivity without counting it as another experiment.

Nominal Gaussian-reference tests with estimated covariance must be labelled as such. All comparisons are C2 reanalysis: these target blocks have already been seen. No fresh g/PCA basis selected to make residuals smaller, no second exponent, no optional extra sizes, no site/bond/tiny/large pseudo-ladder, no resuscitation of the retired 1.55 factor.

Outcome and stop

Return exactly the conclusions supported by this block: reflection residual resolved/unresolved; adjacent shape change resolved/unresolved; decreasing/increasing/nonmonotone on this finite lineage with uncertainty. If weighting or covariance changes the qualitative reading, say so and stop before a physical claim. A larger-N purchase needs a separate decision identifying which surviving alternatives it distinguishes; this task authorizes no purchase.

This is the continuation channel for #622. Do not open another issue for the same analysis. References #633,#655,#663,#702,#703; #275's original-U candidate-map problem is separate.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    priority:P1Bounded parallel analysis or a concrete reserve direction; not all run at once.

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions