Archived experiments
September 18, 2026 ยท View on GitHub
Not the current results. Use the consolidated report.
This directory preserves older methodology, calibration, pilots, regressions and
previous generated presentations. Names such as final-validation, final,
focused-results or BENCHMARK-RESULTS are original historical names, not a
currentness guarantee. Every archived Markdown report is labelled accordingly.
| Collection | Why retained here |
|---|---|
Root experiment/results documents, final-validation/ | Early deterministic and mixed-mechanism experiments |
jev-results/ | Tool-only policy and threshold calibration |
context-results/ | Earlier conversation-wide policy trials |
operations-results/ | Repeated argument/report experiment with an earlier version |
focused-results/ | Previous Claude reference presentation; selected evidence is registered in benchmarks/results/ |
codex-results/ | Previous Codex reference presentation and development notes; selected evidence is registered separately |
Numbers in different collections must not be pooled. Raw private development
runs and probes now live in .yoshi/archive/ with a relocation index. The four
raw runs referenced by the consolidated report remain at their original paths.
No raw trial or source snapshot was deleted during organization, and no new
provider benchmarks were run to replace inconvenient results.