Archive fixed-CDM sampler recovery and efficiency evidence

[AlexanderFengler]

Purpose

• archive the preregistered fixed-CDM Slice, PyMC NUTS, and NumPyro NUTS comparison without rerunning inference
• bind the frozen compact result to immutable hashes, the already-retained canonical datasets, and the exactly reconstructed 1,500-trial scale input
• make the missing-primary-evidence boundary explicit before any ecosystem promotion claim

Recorded outcome

• 15 fits across five scenarios and three routes
• four canonical scenarios contain 16 unique parameter truths, checked on three routes for 48/48 route-specific HDI inclusions
• compact rows report zero divergences across all 10 NUTS fits, eight of them canonical
• maximum recorded objective discrepancy: 4.55e-13
• recorded-machine ESS/s ratios versus Slice: PyMC NUTS 2.82x, NumPyro NUTS 3.46x; scale-normalized ratios 2.42x and 3.13x

These are float64/JAX-x64 smoke results and macOS-arm64 timing measurements, not calibration or portable performance/default evidence.

Evidence boundary

• historical analytical producer: JEAM 0c0ef8b; durable black-box reference: a9f547b; current safety revision: ede7a4f (not rerun here)
• the frozen compact result is pinned at 35b154b5...; the provenance addendum is pinned at 07ca04c8...
• 15 posterior trace hashes and sizes were recorded, but zero trace files were retained
• raw prior- and posterior-predictive draws and historical backend trace attributes were not retained; their compact rows cannot be independently reconstructed from primary evidence
• the historical lock hash exists, but the lock bytes do not
• the scale input is post-hoc reconstructed with its exact historical SHA and canonical 300-row prefix; this restores input bytes only

The network-free verifier recomputes the compact scientific and recorded-machine gates while preserving these limitations. Ecosystem promotion remains blocked.

Verification

• historical six-commit range is patch-identical by git range-diff
• 57 passed in the focused sampler/evidence suite
• 159 passed, 2 deselected in the current-JEAM non-slow compatibility lane
• Ruff check and format, mypy, Pyrefly, and git diff --check passed
• both standalone evidence verifiers passed
• no canonical MCMC or optimization was rerun

Scope

This PR adds frozen benchmark artifacts and network-free verification only. It changes no HSSM runtime API, model default, likelihood route, sampler default, or released behavior.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 31214cb4-c1db-482f-a47b-de443de33388

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}