Benchmark and artifact retention review — 2026-09-27

The historical whole-environment retention decision below was revised by the operator-approved 2026-10-07 cleanup. Package inventories and retained installation metadata now replace the reviewed .venv trees. The raw-evidence retention rules remain in force. Prepared-image cache retention was subsequently revised by the operator-selected recipe-only image cleanup. The image availability statements below describe the 2026-09-27 cleanup, not current image storage; reconstruction records now replace the audited Codira images.

Scope and method

Inventory, physical reorganization, and approved cleanup of benchmarks/ and .artifacts/. Existing experiment leaf names and contents were preserved while their parent directories were grouped by benchmark family and artifact role. Sizes are approximate filesystem usage from du; separate directory totals can overlap when files are hard-linked, so they should not be added together as a reclaimed-space estimate.

Findings

• .artifacts/ measured about 222 GiB before cleanup and about 86 GiB after it, a reduction of roughly 136 GiB by du.
• The nineteen completed agent-efficiency attempt environments originally had agent/.benchmark/uv-cache/ trees with a combined du total of about 152 GiB. The runner copies a prewarmed cache into the attempt environment, then uses frozen offline dependency sync; this cache is dependency-install machinery, not task output. The exact attempt-local copies were removed after approval. The prepared image copies remain available.
• Twenty-four agent-efficiency .venv trees account for about 82 GiB when measured separately. Keep these for now: they record the installed runtime used by each attempt and support direct forensic inspection. Cache and environment trees shared hard-linked files, so the measured reduction in .artifacts/ was less than the cache directories' separate du total.
• .artifacts/benchmarks/ contains 1,281 files. Its 616 *.log files total about 3.6 MB. The logs are small and can contain diagnostics not captured by structured results, so retain them with their .exit files.
• Dated performance runs include campaign plans, selection data, Hyperfine traces, phase timings, profile summaries, and profiler outputs. Keep these together: summaries alone do not retain the detailed comparison and reproduction evidence.
• Corrected, recovery, and rerun semantic-pipeline results now sit under .artifacts/benchmarks/semantic-pipeline/. Their distinct identities remain separate because they provide useful provenance.
• Benchmark inputs are grouped under benchmarks/agent-efficiency/, benchmarks/embedding/, benchmarks/retrieval-quality/, benchmarks/performance/, and benchmarks/semantic-pipeline/. Runtime outputs are grouped under .artifacts/agent-efficiency/, .artifacts/benchmarks/, and .artifacts/validation/. Code, documentation, tests, and local-only manifests were migrated to the new input paths.
• Existing run leaf names and evidence remain intact. Immutable generated receipts and manifests retain historical path values as provenance; current scripts and documentation use the grouped layout.

Cleanup decision table

Group	Recommendation	Reason and replacement summary
.artifacts/agent-efficiency/attempts/*/agent/.benchmark/uv-cache/ (19 trees, about 152 GiB by du)	Removed after explicit approval and image-cache verification	The three referenced fixture-image digests were present locally; no-network, read-only checks confirmed the embedded cache in each. Kept this aggregate inventory and the affected attempt identities here.
.artifacts/agent-efficiency/attempts/*/agent/.venv/ (24 trees, about 82 GiB by separate du)	Keep for now	Installed environment from the actual attempt; useful for direct forensic examination and repeatability.
.artifacts/agent-efficiency/attempts/*/state/attempt-work/ other than the cache candidate	Keep	Contains task workspaces, events, state, provider response evidence, protected baselines, and run records. These are primary experiment evidence.
.artifacts/benchmarks/**/logs/ and .artifacts/validation/repo-gates/ logs/exits	Keep	Only about 3.6 MB for benchmark logs; preserve diagnostics and paired terminal exit state.
Hyperfine, phase timing, profile summary, selection, campaign-plan, and profiler files	Keep	Needed for detailed comparison, provenance, or future profiling; a prose summary cannot replace them.
.artifacts/benchmarks/semantic-pipeline/, retrieval-quality/, backend, and dated campaign roots	Keep for now	Distinct historical identities and recovery/correction provenance have not been proven redundant.
.artifacts/worktrees/	Keep until checked	Currently empty by size inspection; verify worktree registrations before removing the directory.
benchmarks/{performance,embedding,semantic-pipeline}/*.local.json	Keep in place, untracked where ignored	Operator-local inputs may contain machine-specific paths. They were moved with their benchmark family and remain ignored.

The approved cleanup removed only the per-attempt uv-cache trees. The directory reorganization moved benchmark inputs and runtime-output groups into the categories above. It did not delete or rewrite run evidence: prepared images, .venv environments, campaign manifests, logs, response bodies, records, workspaces, and exit statuses remain intact. Original run leaf names were retained under their new parents.

Completed cleanup scope

The approved scope was limited to the nineteen directories matching .artifacts/agent-efficiency/attempts/*/agent/.benchmark/uv-cache/ in completed attempts. All sibling files and directories were preserved, especially .venv, campaign manifests, logs, response bodies, records, workspaces, and exit status. The aggregate before-and-after size and affected run identities below summarize the removed caches without replacing experiment results or provider evidence.

Affected attempt identities are:

• 007: patch-001-r01-codira-mcp
• 008: symbols-001-r01-codira-mcp
• 009: patch-001-r01-codira-mcp
• 010: patch-001-r01-baseline, patch-001-r01-codira-mcp, symbols-001-r01-baseline, and symbols-001-r01-codira-mcp
• 012: symbols-001-r01-baseline
• 013 and 014: patch-001-r01-baseline, patch-001-r01-codira-mcp, symbols-001-r01-baseline, and symbols-001-r01-codira-mcp in each run
• c012a: patch-001-calibration-codira-mcp
• c013a: patch-001-calibration-codira-mcp
• e014a: patch-002-calibration-codira-mcp

The campaign specs reference three distinct prepared fixture-image digests. All three remained available locally. A no-network, read-only check confirmed that each image contains the offline fixture cache before cleanup. A post-cleanup inventory found zero attempt-local uv-cache directories and all 24 .venv directories still present.

Privacy review

This report uses repository-relative paths and aggregate counts only. It omits machine paths, user names, credentials, provider payloads, and local dataset contents.

Operator-approved fixture-worktree cleanup — 2026-10-05

The stale registration for .artifacts/ae/011/fixture-sources/codira was pruned after confirming its worktree directory was already absent. Two clean detached fixture-source checkouts were then removed at the operator's request:

• .artifacts/agent-efficiency/executions/fixture-update-20261003-r1/fixture-sources/codira-current-public at 6a3c025e119a4113323bf3eef7b3c2a376a301fe.
• .artifacts/agent-efficiency/executions/representative-preparation-20261002/fixture-sources/codira-current-public at c19490eff963ac78d41b03cb431cef7574343fe9.

Both commits are ancestors of the current branch head 1485219c87af2fc5fd912c5b76a17aec1dd54d97, and neither checkout contained tracked, untracked, or ignored changes. No source patches needed porting. Campaign records and raw execution evidence were not modified. If either exact source snapshot is needed for a future fresh campaign, recreate a detached worktree from its recorded commit under the designated temporary-workspace directory; do not reuse a historical campaign identity.