# Research checkpoint — September 21, 2026

These archives preserve the complete measurement audit, comparison migration, experiment results and bounded threshold comparison, including preliminary attempts and verification logs. Original source directories remain on disk; the compressed copies avoid duplicating large generated tables in Git. `manifest.json` records every original file hash and archive hash. Every archive was reopened and its contents verified byte for byte before committing.

Extract a selected archive from the repository root to an empty checkout with `tar -xzf results/research_checkpoint_2026_09_21/<archive>.tar.gz`. The stored paths restore the original `results/` layout referenced by the reports and verification tools. Do not extract over newer evidence.

The threshold comparison's canonical evidence is its `final/` directory. No configuration passed all gates; no promotions occurred. Historical inputs remain subject to the limitations described in the reports. Raw TradingView data remains separately managed under the repository's existing ignore rules.

The subsequent fragility diagnostic and higher-low entry-gate replay are preserved
in `mlp_revisit_2026_09_21.tar.xz`, with byte-verified member hashes in
`revisit_manifest.json`. Extract this archive with
`tar -xJf results/research_checkpoint_2026_09_21/mlp_revisit_2026_09_21.tar.xz`.
It restores `results/mlp_experiments/revisit_2026_09_21/`, including complete traces,
both audits, the local review and test log. Findings are in
`docs/mlp_revisit_results_2026_09_21.md`; no perturbation was selected or promoted.