Files
Joshua MoonandClaude Fable 5.1 7d7782aed0 feat(semantic): add cross-encoder rerank stage
Rerank the top 32 fused files with a pinned bge-reranker-v2-m3
cross-encoder, opt-in via --rerank or [semantic] rerank = true.
The shared worker runs the reranker on its own executor thread and
GPU (rerank_model_dir, rerank_gpu_uuid) with a 1500 ms deadline;
failures keep fused order and report rerank_status in coverage.
Adds the cuda_rerank evaluation arm and the measured development
replay: file recall@5 0.542 and recall@10 0.625 against 0.292 and
0.542 for file-first lexical on the same day.

Co-Authored-By: Claude Fable 5.1 <[email protected]>
2026-09-15 14:35:31 -04:00

763 lines
22 KiB
JSON

{
"schema_version": 1,
"split": "development",
"held_out_claim": false,
"measured_commit": "3835a69",
"default_semantic_enabled": false,
"arms": {
"oldlexical": {
"binary_sha256": "727057b4771a19ffa2c2b68273fb316e3c8991d3b790df248db32f814d7400e6",
"page_budget_tokens": 600,
"summary": {
"tasks": 12,
"snapshot_invalid": 0,
"unavailable": 0,
"errors": 0,
"truncated_searches": 11,
"misses": 8,
"mean_file_recall_at_5": 0.25,
"mean_file_recall_at_10": 0.25,
"mean_latency_seconds": 0.6679960291100239,
"usage": {
"tool_calls": 36,
"stdout_bytes": 52077,
"stderr_bytes": 0,
"emitted_tokens": 16801,
"tokenizer": "o200k_base",
"latency_seconds": 24.04785704796086,
"token_cost_censored": false
}
},
"tasks": [
{
"task_id": "lunatic-document-append-guard",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 2,
"page_complete": false
}
},
{
"task_id": "lunatic-legacy-airlock-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 2,
"page_complete": false
}
},
{
"task_id": "lunatic-cycle-link-validation",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 2,
"page_complete": false
}
},
{
"task_id": "lunatic-script-permission-interface",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 2,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-permission-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.5,
"file_recall_at_10": 0.5,
"hits_emitted": 2,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-cycle-state",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 2,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-rebuild-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 2,
"page_complete": false
}
},
{
"task_id": "tfs-writing-commit",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.5,
"file_recall_at_10": 0.5,
"hits_emitted": 2,
"page_complete": true
}
},
{
"task_id": "tg-airlock-controller-cycle",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 2,
"page_complete": false
}
},
{
"task_id": "tg-airlock-pump-pressure",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 2,
"page_complete": false
}
},
{
"task_id": "tg-airlock-secure-open",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 2,
"page_complete": false
}
},
{
"task_id": "tg-airlock-wire-permissions",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 2,
"page_complete": false
}
}
]
},
"newfilefirstlexical": {
"binary_sha256": "92865aa758bdcd5e552e7d1e5aaef69365b1c8a11bc2f59cfbdb6d79df5b01db",
"page_budget_tokens": 600,
"summary": {
"tasks": 12,
"snapshot_invalid": 0,
"unavailable": 0,
"errors": 0,
"truncated_searches": 10,
"misses": 6,
"mean_file_recall_at_5": 0.2916666666666667,
"mean_file_recall_at_10": 0.4583333333333333,
"mean_latency_seconds": 0.4801186444632809,
"usage": {
"tool_calls": 84,
"stdout_bytes": 215074,
"stderr_bytes": 0,
"emitted_tokens": 67921,
"tokenizer": "o200k_base",
"latency_seconds": 40.329966134915594,
"token_cost_censored": false
}
},
"tasks": [
{
"task_id": "lunatic-document-append-guard",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": true
}
},
{
"task_id": "lunatic-legacy-airlock-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "lunatic-cycle-link-validation",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "lunatic-script-permission-interface",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-permission-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.5,
"file_recall_at_10": 0.5,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-cycle-state",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-rebuild-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tfs-writing-commit",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": true
}
},
{
"task_id": "tg-airlock-controller-cycle",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tg-airlock-pump-pressure",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 5,
"page_complete": false
}
},
{
"task_id": "tg-airlock-secure-open",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tg-airlock-wire-permissions",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 1.0,
"hits_emitted": 7,
"page_complete": false
}
}
]
},
"cuda": {
"binary_sha256": "e5f671a87bef34bdec417f126f87dede2755fcf59d36530f4c2553ad8f73dd3f",
"page_budget_tokens": 600,
"summary": {
"tasks": 12,
"snapshot_invalid": 0,
"unavailable": 0,
"errors": 0,
"truncated_searches": 12,
"misses": 8,
"mean_file_recall_at_5": 0.3333333333333333,
"mean_file_recall_at_10": 0.3333333333333333,
"mean_latency_seconds": 0.4304255027680306,
"usage": {
"tool_calls": 82,
"stdout_bytes": 214106,
"stderr_bytes": 0,
"emitted_tokens": 67137,
"tokenizer": "o200k_base",
"latency_seconds": 35.29489122697851,
"token_cost_censored": false
}
},
"tasks": [
{
"task_id": "lunatic-document-append-guard",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "lunatic-legacy-airlock-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "lunatic-cycle-link-validation",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "lunatic-script-permission-interface",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-permission-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-cycle-state",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-rebuild-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tfs-writing-commit",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tg-airlock-controller-cycle",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 5,
"page_complete": false
}
},
{
"task_id": "tg-airlock-pump-pressure",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 5,
"page_complete": false
}
},
{
"task_id": "tg-airlock-secure-open",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tg-airlock-wire-permissions",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
}
]
},
"cuda_rerank": {
"binary_sha256": "907b07c43e4c6e22702ad0e3d5e9ea4ec621f16d41346b6003f143a23551cd7c",
"page_budget_tokens": 600,
"summary": {
"tasks": 12,
"snapshot_invalid": 0,
"unavailable": 0,
"errors": 0,
"truncated_searches": 12,
"misses": 4,
"mean_file_recall_at_5": 0.5416666666666666,
"mean_file_recall_at_10": 0.625,
"mean_latency_seconds": 0.6553997579626967,
"usage": {
"tool_calls": 82,
"stdout_bytes": 218870,
"stderr_bytes": 0,
"emitted_tokens": 68316,
"tokenizer": "o200k_base",
"latency_seconds": 53.74278015294112,
"token_cost_censored": false
}
},
"tasks": [
{
"task_id": "lunatic-document-append-guard",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "lunatic-legacy-airlock-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "lunatic-cycle-link-validation",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "lunatic-script-permission-interface",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-permission-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.5,
"file_recall_at_10": 0.5,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-cycle-state",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 5,
"page_complete": false
}
},
{
"task_id": "tfs-airlock-rebuild-fixture",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tfs-writing-commit",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tg-airlock-controller-cycle",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tg-airlock-pump-pressure",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tg-airlock-secure-open",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 1.0,
"file_recall_at_10": 1.0,
"hits_emitted": 6,
"page_complete": false
}
},
{
"task_id": "tg-airlock-wire-permissions",
"snapshot_valid": true,
"page": {
"file_recall_at_5": 0.0,
"file_recall_at_10": 0.0,
"hits_emitted": 5,
"page_complete": false
}
}
]
}
},
"latency": {
"serial": {
"requests": 100,
"p50_seconds": 1.0717706709983759,
"p95_seconds": 1.4897110809979495,
"maximum_seconds": 4.224220871998114,
"errors": 0,
"nonready_responses": 6,
"over_budget_responses": 0,
"p95_limit_seconds": 2.0,
"passed": false
},
"concurrent": {
"requests": 120,
"p50_seconds": 3.3592287749925163,
"p95_seconds": 6.08325262699509,
"maximum_seconds": 6.460270354000386,
"errors": 0,
"nonready_responses": 0,
"over_budget_responses": 0,
"p95_limit_seconds": 5.0,
"passed": false
},
"clients": 6,
"waves": 20,
"warmup_requests_excluded": 12,
"tokenizer": "o200k_base",
"passed": false
},
"limitations": [
"Twelve development tasks, no held-out evaluation or statistical significance claim.",
"Selected source hashes validated; complete repositories were not frozen and Lunatic changed during measurement.",
"First-page file recall does not establish that source bytes were read or understood.",
"Tool output tokens are observed; model input, reasoning and billed tokens are unknown.",
"CUDA retrieval and latency use the measured commit; subsequent recovery-only fixes require separate regression checks."
],
"recovery_fix_commit": "8ffa665",
"raw_artifact_archive": "/home/josh/.cache/trufflepig/evaluation/gpu-search",
"verification": {
"cuda_library_checkpoint_passed": 289,
"final_preparation_cuda_tests_passed": 15,
"final_worker_semantic_tests_passed": 18,
"final_preparation_no_default_tests_passed": 15,
"python_tests_passed": 59,
"all_targets_cuda_check": "passed"
},
"rerank_measurement": {
"measured_tree": "uncommitted rerank change on 6c50a74",
"binary_sha256": "907b07c43e4c6e22702ad0e3d5e9ea4ec621f16d41346b6003f143a23551cd7c",
"reranker": "rozgo/bge-reranker-v2-m3@fbd57b17b4db111a9d16813bb08b4c804fac18e9",
"embedding_gpu": "NVIDIA GeForce RTX 4090",
"rerank_gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
"note": "All three arms were re-measured in one session on the current working trees, so the same-day baselines below are the comparison for cuda_rerank; the arms recorded above were measured at 3835a69 on earlier trees.",
"same_day_baselines": {
"newfilefirstlexical": {
"binary_sha256": "907b07c43e4c6e22702ad0e3d5e9ea4ec621f16d41346b6003f143a23551cd7c",
"summary": {
"tasks": 12,
"snapshot_invalid": 0,
"unavailable": 0,
"errors": 0,
"truncated_searches": 10,
"misses": 5,
"mean_file_recall_at_5": 0.2916666666666667,
"mean_file_recall_at_10": 0.5416666666666666,
"mean_latency_seconds": 0.3370037924220977,
"usage": {
"tool_calls": 85,
"stdout_bytes": 218053,
"stderr_bytes": 0,
"emitted_tokens": 68941,
"tokenizer": "o200k_base",
"latency_seconds": 28.645322355878307,
"token_cost_censored": false
}
}
},
"cuda": {
"binary_sha256": "907b07c43e4c6e22702ad0e3d5e9ea4ec621f16d41346b6003f143a23551cd7c",
"summary": {
"tasks": 12,
"snapshot_invalid": 0,
"unavailable": 0,
"errors": 0,
"truncated_searches": 11,
"misses": 8,
"mean_file_recall_at_5": 0.3333333333333333,
"mean_file_recall_at_10": 0.3333333333333333,
"mean_latency_seconds": 0.41867398462051003,
"usage": {
"tool_calls": 82,
"stdout_bytes": 214534,
"stderr_bytes": 0,
"emitted_tokens": 67192,
"tokenizer": "o200k_base",
"latency_seconds": 34.33126673888182,
"token_cost_censored": false
}
}
}
},
"per_task_file_recall_at_5": {
"lunatic-document-append-guard": {
"newfilefirstlexical": 1.0,
"cuda": 1.0,
"cuda_rerank": 1.0
},
"lunatic-legacy-airlock-fixture": {
"newfilefirstlexical": 0.0,
"cuda": 0.0,
"cuda_rerank": 0.0
},
"lunatic-cycle-link-validation": {
"newfilefirstlexical": 0.0,
"cuda": 0.0,
"cuda_rerank": 1.0
},
"lunatic-script-permission-interface": {
"newfilefirstlexical": 0.0,
"cuda": 1.0,
"cuda_rerank": 1.0
},
"tfs-airlock-permission-fixture": {
"newfilefirstlexical": 0.5,
"cuda": 0.0,
"cuda_rerank": 0.5
},
"tfs-airlock-cycle-state": {
"newfilefirstlexical": 1.0,
"cuda": 0.0,
"cuda_rerank": 0.0
},
"tfs-airlock-rebuild-fixture": {
"newfilefirstlexical": 0.0,
"cuda": 0.0,
"cuda_rerank": 1.0
},
"tfs-writing-commit": {
"newfilefirstlexical": 0.0,
"cuda": 0.0,
"cuda_rerank": 0.0
},
"tg-airlock-controller-cycle": {
"newfilefirstlexical": 1.0,
"cuda": 1.0,
"cuda_rerank": 1.0
},
"tg-airlock-pump-pressure": {
"newfilefirstlexical": 0.0,
"cuda": 1.0,
"cuda_rerank": 0.0
},
"tg-airlock-secure-open": {
"newfilefirstlexical": 0.0,
"cuda": 0.0,
"cuda_rerank": 1.0
},
"tg-airlock-wire-permissions": {
"newfilefirstlexical": 0.0,
"cuda": 0.0,
"cuda_rerank": 0.0
}
},
"limitations": [
"Twelve development tasks, no held-out evaluation or statistical significance claim.",
"Workspace searches rerank each of the three members sequentially; mean search latency rose from about 1.1 s (cuda) to about 2.7 s (cuda_rerank) at a 600-token page.",
"Lunatic and tales-from-space working trees have moved since the recorded parent revisions; selected label hashes still validated for all twelve tasks."
]
}
}