Add Claude Haiku 5.5 evaluation artifacts for 99 CPU tasks
Adds the latest recorded result for each of 99 selected CPU tasks under claude-haiku-5-5/, with all 102 recorded attempts retained. GPU task tb3/exam-pdf-eval was not run.
Raw statuses: 11 pass, 49 fail, 4 invalid, 18 infrastructure_error, 17 calibration_error. Valid pass/fail records: 60/99; recorded coverage is not valid completion.
Includes summary.json, scores.json, report.json, notes.txt, retest-calibration.json, and task-artifacts.tar.gz with an artifact index and SHA-256 manifest. Extract the archive inside the model directory; summary paths identify archive member directories. Configuration files, credentials, email addresses and local host paths were removed/redacted. Workspace/checkpoint copies and binary/cache files were excluded. Embedded source hashes refer to pre-redaction originals; manifest/index hashes cover exported bytes.
This is a mixed-runtime latest-result snapshot: the original Haiku run used 5933dd2, and selective retests used a3dde57 (main 842e5a6 plus macOS compatibility patch). Changed scored retests replace older results regardless of improvement, including the previous pass-to-fail change. Calibration-blocked Together retests retain the earlier scored result and remain in history. Raw error classifications and timeout/API caveats are preserved. Prior report usage estimates may exclude retests and have missing-usage/cache-pricing limitations.
Validation: 99 unique task IDs, 102 attempt records, status totals, 6,328 archive file hashes, 8 top-level file hashes, archive links, and sensitive-pattern scans passed locally.