Historical benchmark and failure reports¶
The 29 reports below preserve authored findings previously held only in ignored artifact trees. They are historical records, not current launch instructions or performance claims. Original reports and raw evidence remain untouched. Known directory moves are repaired in these copies; absent evidence is explicitly identified in each report rather than replaced with a different run.
Reports¶
Remaining evidence gap¶
21 reports reference historical evidence that could not be located in this checkout. These include early backend runs, the embedding matrix, final embedding campaigns, ONNX sweeps, split-engine experiments, the July 14 retrieval-quality run, and earlier vector-store candidate runs. Recover those exact identities from an external backup before attempting to revalidate their numbers. A rerun is new evidence and cannot replace an old identity.
The July 20 vector-store final campaign and later agent-efficiency failure evidence remain available. Generated CSV tables and raw diagnostics remain ignored; a Git-only checkout contains authored findings, not all reproduction data. Their absence must not be presented as current verification.
Current retention policy is in the dated review.