Clock-asymmetry and gamma-CDM diagnostics: two frozen negatives, Paper 1 robustness confirmed#1
Open
dmobius3 wants to merge 12 commits into
Open
Clock-asymmetry and gamma-CDM diagnostics: two frozen negatives, Paper 1 robustness confirmed#1dmobius3 wants to merge 12 commits into
dmobius3 wants to merge 12 commits into
Conversation
fit_clock_asymmetry.py implements the frozen registration (mode-identity-theory clock-asymmetry-fit.md, commits dc8f9f6 and ff875be) as the continuous closure of the Appendix A clock-exponent machinery: n(eps) = -1/2 - eps threaded through the mirrored model functions, nothing else new. Modes: gates (slice reproduction against tables/clock_exponent_appendix_A_fits.csv, tolerance frozen at max(0.5, 0.1% of published), tiers refuse to run until gates pass), tier1 (profile Delta-chi2(eps) on the frozen grid, both references reported: vs LCDM 1772.456 and vs eps = 0), tier2 (joint MCMC over the four sampled parameters s0, H0rd, M_B, eps; M_B sampled as in the baseline; flat eps prior [-1.6, 1.1]; seed 12345), and manifest (seed, grids, tolerances, commit hash, package versions, data checksums, registration cross-references). Omega_Lambda = 0.685 fixed; the free-Omega_Lambda scan is robustness, not authority, and is not implemented here. This commit is implementation only: no epsilon point has been evaluated. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Slice reproduction against tables/clock_exponent_appendix_A_fits.csv, deterministic optimizer, frozen tolerance max(0.5, 0.1% of published): C (eps = -3/2): published +7276.63, reproduced +7276.63, diff -0.000 A (eps = -1/2): published +218.69, reproduced +218.69, diff +0.001 D (eps = 0): published +0.11, reproduced +0.11, diff -0.002 B (eps = +1/2): published +1821.93, reproduced +1821.93, diff -0.003 The epsilon-family code path has earned the right to interpolate between the walls. No epsilon point off the published slices has been evaluated; tier1 and tier2 remain unrun pending redline and tag. Manifest committed alongside (seed, grids, tolerances, data checksums, registration cross-references dc8f9f6 / ff875be). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three redline hardenings, no change to family, priors, data, expectation, or verdict authority. (1) The gate record is now bound to the code and data that produced it: a provenance file stores the script sha256 and data md5s at gate time, and the tiers refuse unless both match at run time, so "gates passed" is mechanically true, not procedurally true; the hardening itself invalidated the old record and the gates were re-earned (C -0.000, A +0.001, D -0.002, B -0.003 against the published anchors, same as before). (2) Tier 2 no longer assumes the symmetric point: it refuses to run before Tier 1, seeds most walkers at the Tier 1 minimum eps_hat with its profiled parameters, and stratifies a minority across the full eps prior, each walker seated on its nearest Tier 1 curve row, so a displaced minimum, a second mode, or an excluded zero can be found without luck. (3) Tier 2 reports convergence diagnostics: integrated autocorrelation time, effective sample count, and split-half eps-quantile stability, informational rather than a physics gate. Manifest now carries the script sha256. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Frozen sequence tier1 then tier2, run uninterrupted from the tag (351938c), transcript in results/clock_asymmetry_run.log (force-added past the *.log ignore rule: the run transcript is a registered artifact of this execution, not build noise), seed 12345. Generated files only; no code changed. Tier 1 (profile): minimum at eps_hat = -0.106 with chi2 = 1764.714, delta vs LCDM = -7.742, delta vs eps=0 = -7.849; 68% interval [-0.139, -0.069], 95% [-0.171, -0.031]; both 68% half-widths (0.033, 0.037) inside the 0.05 bar; zero excluded at 95%. The profiled s0 slams to the prior floor for eps > -0.008 (the published Model D regime, reproduced at +0.108) and climbs a degeneracy valley to s0 = 0.467 at the minimum. The eps = +1 slice lands at +6395.5, discharging the registered S^(-3/2) completing fit as expected dead. Tier 2 (joint posterior): eps median -0.092, 68% [-0.131, -0.051], 95% [-0.620, -0.019], corr(s0, eps) = -0.822; best point (s0 = 0.467, eps = -0.106) coincides with the Tier 1 minimum. Diagnostics: acceptance 0.469, tau_max = 131 (chain length ~38 tau, marginal; reported as informational per the spec), effective samples ~974, split-half eps-quantile max shift 0.003 (68% region stable; the deep negative 95% tail rides the s0 degeneracy valley and is the least-resolved feature). Scoring happens against the frozen verdict table in the registration (mode-identity-theory clock-asymmetry-fit.md), not here. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Implements the adjudicated sequence with the questions posed in the script header before any diagnostic runs: audit (recompute headline numbers from the committed raw outputs; halts everything on mismatch, technical not physics), omegaL (Omega_Lambda-free profile over the epsilon grid: valley evaporation = fixed-OL tension absorber, survival = family feature), split (SN-only and BAO-only epsilon profiles: independent negative leanings are harder to dismiss; disagreement suggests the dial is not tick physics). Deterministic profiles only, no MCMC, mirrored numerics; the only new freedoms are the ones each diagnostic names. Nothing here can alter the frozen row-3 scoring. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Audit: all seven headline numbers recomputed from the committed raw outputs match exactly; the row-3 scoring stands on clean arithmetic. Omega_Lambda-free scan: the valley survives in depth (-8.67 vs -7.85 at fixed OL) but RELOCATES: at the free-OL minimum (eps = -0.08 to -0.10) the profiled s0 sits on the prior floor (0.001) with Omega_Lambda = 0.61-0.63. At s0 = 0 the family reduces to a pure matter-exponent deformation of LCDM, H^2 = Om (1+z)^(3+2eps) + OL, in which the budget clock is inert: the data preference is for a softer matter exponent (~2.84) with more matter (~0.375), expressed equally well with the tick structure switched off. The fixed-OL valley (s0 = 0.47) and the free-OL minimum (s0 = 0, OL = 0.625) are two corners of one degenerate ridge. SN/BAO split: eps_SN runs to the grid edge (-0.40, 68% region eps <= -0.23, unbounded below within the grid); eps_BAO = -0.03 with 68% [-0.083, -0.011]. The two 68% regions are DISJOINT; the joint -0.106 is a compromise neither dataset wants. Per the question posed before the run: disagreement, so probably not tick physics. Interpretation stays with review; these are the mechanical readings of the pre-posed questions. Diagnosis, not authority. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Implements the numerical spec frozen in the Desktop gameplan (sections 3-4): roster LCDM / gamma-CDM / wCDM / CPL / BA / JBP plus the oLCDM curvature diagnostic; data variants all / z>0.01 / z>0.1 by a single index mask on the Pantheon+ vector and covariance together, with N_all, N(z>0.01), N(z>0.1) reported everywhere; priors frozen (gamma [2,4] primary and [0,5] robustness, zoo w0 [-3,1] wa [-3,2], Ok [-0.5,0.5]); Gate B evaluated on the z>0.01 sample; evidence by Savage-Dickey at the nested interior point gamma = 3 from the gamma-CDM chain under both priors with a bandwidth-scan stability check, quantifying the prior-fairness caveat exactly; Gate D template image via a noise-free BAO mock at the 13 DR2 points fitted by CPL with DESI covariance weighting (Paper 1's template-bias pattern) plus the model-native effective-density positivity check to z = 1.5; numerical gate thresholds and the curvature-costume 70% flag exactly as in the gameplan. Reproduction gates: LCDM(all) must hit 1772.456 within 0.5 and gamma-CDM(all) must reproduce the prior finding (about 1763.8) within 1.0, since it is the identical model family reparameterized; fits, mcmc, and verdict refuse to run before the reproduction gates pass. Single execution to follow in one results commit. No external-framework language anywhere. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Verdict per the frozen tree: paper = FALSE. Gates A and B fail;
Gate C passes; Gate D fails; no curvature costume.
The three counts explain everything: N_all = 1701, N(z>0.01) = 1590,
N(z>0.1) = 960. The shipped Pantheon+ file carries 111 light curves
below z = 0.01 that the published cosmology cut excludes, and they
were carrying roughly two thirds of the signal: Delta chi2 vs LCDM
goes -8.75 (all) -> -2.97 (z>0.01) -> -1.89 (z>0.1). Gate A's bar
was -4 after the standard cut; -2.97 fails it. The reconciliation
with the literature's mild Pantheon+ preferences (1.7-1.9) is
achieved in one cut. Post-cut gamma_hat = 2.90 with 68% [2.836,
2.956]: a mild residual preference survives, consistent in scale
with published late-time analyses, but it is not a paper.
Gate B: dAIC = -0.97 (needs < -2); Savage-Dickey evidence favors
LCDM under both priors (Delta lnZ = -1.09 primary [2,4], -2.00
robust [0,5]): the posterior excludes gamma = 3 at 68% yet the
evidence still prefers the nested model, the classic prior-volume
lesson, quantified exactly as the fairness caveat anticipated.
Gate C passes POST-CUT: the SN-only and BAO-only 68% intervals
([2.31, 3.15] and [2.85, 3.02]) overlap once the low-z subset is
removed. The earlier disjointness was itself low-z-driven: the
"tension absorber" was specifically absorbing the z < 0.01 subset.
Gate D fails: the post-cut CPL image is (w0 = -0.97, wa = +0.36),
wrong wa sign for the community direction; after the cut the
deformation no longer even points at the DESI quintom-B pull.
oLCDM captures 47% of the improvement: no curvature costume.
Deliverable per the frozen tree: a short note recording the negative
("the soft-exponent preference was low-z driven"), not a paper.
Paper 1 untouched, as always.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Not a falsification test: the gamma-gate killed the separate soft-exponent paper idea, not Paper 1. Paper 1's claim is that Model D+Lambda is statistically equivalent to LCDM (published Delta chi2 = +0.11, s0 < 0.19 at 95%); this asks only whether that near-null status is stable under the standard z>0.01 Pantheon+ cut the published cosmology and the external late-time literature both adopt. Expectation stated in the script before running: stands, s0 already near the LCDM limit. Frozen bars: PASS (dchi2 in [-2,+3] and s0 95% UL <= 0.25), ROBUSTNESS NOTE (dchi2 up to +9 or UL up to 0.45), else REASSESS. The all-variant row doubles as the Paper 1 reproduction anchor (should land near +0.11). One results commit to follow. No framework language. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Result commit. The check the review pass asked for: does Paper 1's near-null Lambda-cos result survive the standard z>0.01 Pantheon+ cut. It does, on every criterion, with the pre-frozen bars. Reproduction anchor (full file): Lambda-cos vs free-Om LCDM Delta chi2 = +0.109, reproducing the published +0.11 exactly, so the pipeline still speaks for Paper 1. Under the cut: best-fit s0 pins to the LCDM limit (0.001) on both the z>0.01 (N=1590) and z>0.1 (N=960) samples, with Delta chi2 = +1.34 and +1.90 (Lambda-cos marginally worse, the expected behavior of a model whose extra freedom the data do not want). s0 profile-likelihood 95% UL tightens slightly to 0.149 (vs 0.19 published on the full file): the low-z SNe were loosening, not driving, the bound; removing them leaves the result more LCDM-like, not less. Frozen verdict bars: PASS = Delta chi2 in [-2,+3] AND s0 95% UL <=0.25. Both met (+1.34, 0.149). Paper 1 stands cleanly; no robustness note or retraction warranted. The dataset-wording footnote (Pantheon+ 1701 vs the z>0.01 cosmology sample of 1590) is optional and cosmetic. Method note: the s0 UL is by profile likelihood (Delta chi2 = 2.706, one-sided 95%), the honest interval for a boundary-limited near-null parameter, adopted after emcee 3.1.6's init condition-number guard rejected the raw-scale (H0rd ~1e4 vs s0 ~0.1) walker ball twice. A technical estimator choice, deterministic, not a physics change; the frozen numeric bar (0.25) is untouched. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…les) Refactor the profile-likelihood UL into a helper and run it on the full file and the z>0.01 cut, so the "tightens" comparison is profile-vs-profile rather than profile-vs-posterior. Result: full-file 95% UL s0 < 0.177, cut 95% UL s0 < 0.149 (both Delta chi2 = 2.706, one-sided); the full-file profile bound sits just inside the published posterior UL of 0.19, and the cut tightens it. Deterministic fits unchanged (+0.109 / +1.34 / +1.90; best fit at the LCDM limit on every variant). Verdict unchanged: PASS. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A two-page standalone note recording the negative result from the gamma gate (results commit 69c4604). The one-parameter matter-exponent deformation H^2/H0^2 = Om (1+z)^gamma + OmegaLambda (gamma = 3 recovers LCDM) is presented as a diagnostic, not a candidate cosmology. The full Pantheon+ vector gives gamma < 3 at Delta chi2 = -8.75 vs LCDM; the standard z > 0.01 cosmology cut reduces this to -2.97 and the z > 0.1 cut to -1.89 (counts 1701 / 1590 / 960). Savage-Dickey evidence favors LCDM under both frozen priors (Delta lnZ = -1.09 primary [2,4], -2.00 robust [0,5]). The z > 0.01 cut heals the earlier SN/BAO split and flips the post-cut CPL template image to wa = +0.36. The apparent preference is low-redshift driven; a mild residual (gamma ~ 2.90, Delta chi2 ~ 3.0) consistent with published mild Pantheon+ preferences remains. Framework-free, house style, no new fitting (every number is from 69c4604). Paper 1 untouched. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
This branch adds a self-contained investigation into two one-parameter deformations of the flat ΛCDM expansion history, run against Pantheon+ (1701 SNe Ia, full covariance) + DESI DR2 BAO (13 points), all under a freeze-before-fit protocol with the git history as the audit trail. It does not modify the published Λcos paper or pipeline; every new script mirrors the existing likelihood machinery and adds one named degree of freedom at a time.
Two questions were asked and answered, plus a robustness check on the published result:
1. Clock-asymmetry (ε) family. The Appendix A clock-exponent models (A/B/C/D) are the discrete slices of a continuous family
dt/dτ = S^(−1/2−ε). The design was frozen and tagged (eps-family-v1.0), the gates reproduced the published slice χ² to ~1e-3, then a single run measured ε. Result: the frozen verdict fired its "challenge" row (ε = 0 excluded at 95%, ε̂ ≈ −0.11), but the post-run diagnostics (Ω_Λ-free scan + SN/BAO split) showed the preference relocates off the model's own structure and the SN and BAO subsets disagree, i.e. the ε dial is contaminated as a clean probe, not evidence of new physics.2. γ-CDM head-to-head. The contaminated preference was re-cast as a standalone phenomenology model,
H² = Ω_m(1+z)^γ + Ω_Λ, and put through a fully numerical frozen gate (six-model roster, frozen priors, low-z vetting, Savage-Dickey evidence, curvature diagnostic). Verdict:paper: false. The apparent Δχ² ≈ −8.75 collapses to −2.97 after the standardz > 0.01cut; the 111 low-redshift light curves carried ~2/3 of the signal. The deliverable is a two-page diagnostic note (paper2/), not a cosmology claim.3. Λcos (Paper 1) robustness. Confirmed the published near-null result survives the same
z > 0.01cut: best fit stays at the ΛCDM limit and the profile-likelihood 95% upper limit tightens from s₀ < 0.177 to s₀ < 0.149 (apples-to-apples). Paper 1 stands; this fed a one-line robustness addition to the framework page (separate repo).What a reviewer should know
Commits
2ddbdddimplement ε family →2406ab5gates pass →351938charden + re-earn gates →6e95fedε run →3085ca4/16094d7diagnostics (dial contaminated) →343c9d8/69c4604γ-gate (paper: false) →117c307/7a52982/a9d7a18Paper 1 robustness (stands) →7f6aa07diagnostic note.🤖 Generated with Claude Code