# Actual validation — 2026-10-06

- Reused20 hash-locked public resource files and existing Python3.12.14 environment; no hidden author path in source. Download stage means verified existing resources; no claim of redownloading weights.
- Correctly specified five-arm smoke before formal:10updates/48generations,exit0,18.305180666968226s,RSS2617606144B. Conservative ×24 estimate439.324336s<1200/6GiB.
- Formal five64step arms,320updates/960generations,exit0,267.89943604078144s,RSS2961309696B. All checkpoint files preserved privately with SHA256; all320step shifted CE checks pass,max1.1920928955078125e-7. No training failures or replacements; no confirmation inference. GPU memory unmeasured.
- Independent raw audit960/960; after export949fully reparsed,11news echoes retain hashes/derived facts, cannot be text-reparsed from public package. Raw960fileSHAa013cf02a004a9ace0ccbd45279592acbcf8204f493cd628a0c4473f314365b3. smoke48/48,no redactions. Each audit reports exact scope.
- fixed93/fixed94 numerical state identity and output identity pass. Three shuffled states distinct.091 target_only64loss/gradient trajectories and baseline/after160prompt-label-score records match exactly; continuity.json. State hash algorithm now includes shape metadata; compare numerical state directly across algorithms, not their different hash definitions.
- Conditional paired means/wins/losses independently recalculated; bootstrap reproduced by frozen analysis.py, not claimed as independent coverage proof. Two parser edge fixtures checked. No iid movie/source-group guarantee, no selection-adjusted inference, no human model-output gold.
- Five official upstream source files fresh anonymous reads match installed bytes, permanent commits/licenses in SOURCE_MAP. First sandbox public read denied then permitted request succeeded for code; some subsequent webpage TLS/403 errors are documented outside package, web-tool primary pages used where accessible. Experiment never needed network because cache hashes passed.
- New audit/docs/export added after formal inference. Executed modules/config/protocol/resource files retain identity.json hashes. No favorable seed selection, no parameter adjustment after seeing results. Final source ZIP validation runs the same real five-arm smoke after extraction; article-root package_validation.json records command/exit/hash comparisons. Distribution excludes checkpoints, token traces, raw source data, caches, environment, .DS_Store, __pycache__, credentials and private paths.