Spaces:
Running
Running
File size: 18,800 Bytes
cd19b51 b083097 cd19b51 dad6879 72b06db cd19b51 a1e7884 cd19b51 dad6879 cd19b51 b083097 cd19b51 2ad3161 72b06db 8f3dc81 2ad3161 b083097 72b06db 8f3dc81 72b06db 0f708fe 8f3dc81 a1e7884 e542d1d 72b06db e542d1d 72b06db e542d1d 72b06db e542d1d b083097 8f3dc81 b083097 a1e7884 b083097 cd19b51 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 | {
"headers": [
"Method",
"True drifts found",
"Findings",
"Precision",
"Tokens"
],
"rows": [
[
"Symbol check alone",
"24 of 45 (19 of 33 on the 9-repo union)",
43,
"58.1%",
"0 (no model calls)"
],
[
"Whole-repo Beam alone",
"11 of 45 (11 of 33 on the 9-repo union)",
19,
"57.9%",
"4,424,575"
],
[
"Hybrid (symbol + whole-repo Beam triage)",
"34 of 45 (22 of 33 on the 9-repo union)",
60,
"75%",
"~9.2M tokens (two whole-repo passes over 12 libraries)"
],
[
"Hybrid + repro-confirmed (strict)",
"30 of 45 (19 of 33 on the 9-repo union)",
41,
"92.7%",
"~9.3M tokens (two whole-repo passes plus repro scripts)"
],
[
"Chunked Beam (~32k chunks) (9-repo union)",
"15 of 33",
25,
"60%",
"3,529,945"
],
[
"Chunked Beam (~128k chunks) (9-repo union)",
"9 of 33",
13,
"69.2%",
"3,342,201"
]
],
"notes_markdown": "- The **45-drift reference set** is the union of what the methods found, so it favours them. Labels were checked against the source by an AI reviewer. The stored field is still `hand_verdict`.\n- Beam triage judged **56** candidates and rejected **0** of 32 true candidates, while removing **15** of 24 false or unclear ones. 11 of 12 libraries were triaged; structlog failed; cattrs needed 2 calls.\n- **2 false + 1 unclear (werkzeug S38, S39 false; S41 unclear)** received a strict repro badge.\n- structlog: Beam's triage reply had no JSON, so its 6 candidates are shown untriaged.\n- Chunked Beam alone found **11** local docstring/prose drifts the hybrid misses.\n- The strict repro rule was set after the first repro run.\n- No prompt caching was observed on triage calls (each billed the full ~300–400k pack).\n- Both whole-repo passes (first sweep + hybrid triage) cost about **9.2M** tokens (4,424,575 + 4,739,392 = 9,163,967). End-to-end including repro is **9,302,105 (~9.3M)** (4,424,575 + 4,739,392 + 138,138). The cheaper symbol-only-candidate variant was not run.\n- The local-file vs whole-repo ablation spent **1,948,469** extra tokens. That ledger is not part of the 9.3M hybrid end-to-end.\n- The docs-pass + 32k docstring-chunk extra pass spent **7,713,453** extra tokens. That ledger is not part of the 9.3M hybrid end-to-end.\n- The matched-cost retrieval control spent **3,957,434** extra tokens. That ledger is not part of the 9.3M hybrid end-to-end.\n- The tomlkit `docdrift run` recording spent **185,842** extra tokens. Matched-cost controls spent 3,957,434 + 7,528,878 (extension). Together with the recording that is **11,672,154**, not part of the 9.3M hybrid end-to-end.",
"funnel": [
{
"step": "Symbol check",
"true": 24,
"detail": "0 Beam tokens"
},
{
"step": "Beam triage",
"true": 34,
"detail": "rejected 0 of 32 true; removed 15 of 24 false/unclear"
},
{
"step": "Strict repro",
"true": 30,
"detail": "38 of 41 badges were real; 2 false + 1 unclear (werkzeug S38, S39 false; S41 unclear)"
}
],
"recall": {
"max_probe_tokens": 947446,
"far_hits": [
5,
6
],
"chunk_far": [
0,
6
],
"focused_tokens": 4755853,
"chunk_baseline_tokens": 1000174,
"focused_multiplier": 4.8,
"docstring_pooled": "1/32 pooled over 2 seeds (0/18 + 1/14)",
"chart": "assets/recall_vs_distance.png",
"source": "results_live/recall/",
"buckets": [
"~1k",
"~10k",
"~50k",
"~100k",
"~200k",
"~400k",
"700k+"
],
"default_series": [
{
"id": "focused_seed1",
"label": "focused whole-context",
"points": [
0.25,
0.0,
0.25,
0.75,
0.0,
0.25,
0.8333333333333334
]
},
{
"id": "whole_seed1",
"label": "whole-repo",
"points": [
0.0,
0.0,
0.0,
0.5,
0.0,
0.0,
0.5
]
},
{
"id": "chunk128k_seed1",
"label": "chunked 128k",
"points": [
0.25,
0.0,
0.5,
0.5,
0.0,
0.0,
0.0
]
}
]
},
"chart_present": true,
"ablation": {
"available": true,
"headers": [
"Metric",
"Whole-repo (reused hybrid)",
"Local-file only",
"Whole-repo, subset (27)",
"Local, subset (27)",
"Whole-repo rerun, subset (27)"
],
"rows": [
[
"True candidates kept",
"32/32",
"32/32",
"13/13",
"13/13",
"13/13"
],
[
"False/unclear candidates rejected",
"15/24",
"15/24",
"9/14",
"12/14",
"10/14"
],
[
"Precision (kept that are true)",
"0.78 (32/41)",
"0.78 (32/41)",
"0.722 (13/18)",
"0.867 (13/15)",
"0.765 (13/17)"
],
[
"Verdict accuracy",
"0.839",
"0.839",
"0.815",
"0.926",
"0.852"
],
[
"Citation hits gold file:line (gradable)",
"41/52",
"34/52",
"19/26",
"19/26",
"22/26"
],
[
"Right verdict + gold citation",
"39/52",
"33/52",
"17/26",
"19/26",
"20/26"
],
[
"Citation verified (quote matches source)",
"54/56",
"52/56",
"25/27",
"26/27",
"27/27"
],
[
"Cross-file cases: right verdict",
"26/30",
"26/30",
"12/13",
"12/13",
"12/13"
],
[
"Cross-file cases: right verdict + citation",
"25/30",
"14/30",
"12/13",
"8/13",
"12/13"
],
[
"Evidence outside local context: right verdict",
"12/13",
"12/13",
"4/4",
"4/4",
"4/4"
],
[
"Evidence outside local context: right verdict + citation",
"11/13",
"0/13",
"4/4",
"0/4",
"4/4"
],
[
"Beam tokens",
"4,022,532 (11 calls; 4,331,995 as spent incl. a cattrs re-send)",
"830,527 (56 calls; 848,654 as spent incl. 1 re-send)",
"—",
"—",
"1,099,815"
]
],
"claim": "Long context didn't make verdicts more accurate here; it made them cite the real evidence across files, which is what we verify and show maintainers.",
"caveats": [
"n=56 judged candidates (structlog excluded; its hybrid triage failed).",
"Labels and gold file:line were checked against the source by an AI reviewer, not a human maintainer.",
"Whole-repo is one batched call per repo; local is one call per candidate.",
"Local was given every file the candidate already named.",
"6 of 56 disagreements split 3-3. Stability rerun: 26/27 verdicts matched.",
"Whole-repo used 4,022,532 tokens vs 830,527 for local (~4.8×). This ablation run's own ledger is 1,948,469 tokens over 60 calls; that is extra experimental cost, not part of the 9.3M hybrid end-to-end.",
"Matched-cost retrieval control: 3,888,241 attributable / 3,957,434 as spent (57 calls; same 56 candidates). Extra experimental cost, not part of the 9.3M hybrid end-to-end.",
"Pre-registered matched-cost extension: 7,528,878 tokens (63 more items). Extra experimental cost, not part of the 9.3M hybrid end-to-end."
],
"n": 56,
"disagreements": 6,
"stability_agree": 26,
"stability_n": 27,
"whole_tokens": 4022532,
"local_tokens": 830527,
"token_multiplier": 4.8,
"ablation_ledger_tokens": 1948469,
"ablation_ledger_calls": 60,
"cross_file_cite_whole": "25/30",
"cross_file_cite_local": "14/30",
"outside_cite_whole": "11/13",
"outside_cite_local": "0/13",
"accuracy_whole": "0.839",
"accuracy_local": "0.839",
"accuracy_whole_subset": "0.815",
"accuracy_local_subset": "0.926",
"outside_cite_whole_subset": "4/4",
"outside_cite_local_subset": "0/4",
"subset_note": "On the 27-item rerun subset, local-only scored higher on verdict accuracy (0.926 vs 0.815). Whole-repo's advantage is citation: 4/4 vs 0/4 when the proof lives outside the candidate's files.",
"matched_available": true,
"why_matched_cost_note": "**Is it the whole repo, or just more tokens?** We gave a retrieval baseline the **same** Beam budget as whole-repo triage. That came to 3,957,434 against 4,022,532 tokens as spent (3,888,241 attributable; 96.7% of whole-repo) over the same 56 candidates. For each candidate, it got the local files plus the symbol definitions, imports and BM25 matches that fit in about 56k tokens.\n\nVerdict accuracy came out the same (46/56 against 47/56). Grounding still favoured the whole repo: the right verdict **and** the right cross-file line in 25/30 cases, against 19/30 for retrieval. Where the evidence lives outside the files the doc points at, the figures are 11/13 against 6/13. Retrieval pulled in importers or re-exports rather than the defining file in 5 of 13 outside-local items.\n\nThat gap is suggestive, not proven, at this sample size (exact McNemar p = 0.070). We pre-registered the test and report it as \"not shown at this n\".\n\n**Pre-registered extension.** We then added 63 items (57 planted renames, 6 real structlog candidates), frozen before any Beam call (design sha256 42f6b6fe…c996). Pooled over 78 cross-file items, the right verdict **and** the right line came out 52/78 for the whole repo against 38/78 for retrieval (exact McNemar p = 0.0094). Verdict accuracy still did not differ (108/119 against 105/118). The new items alone give p = 0.096 (\"not shown at this n\"), and the pooled result relies on one pre-registered re-send of a whole-repo call that failed to parse (p = 0.18 without it). The honest takeaway: whole-repo context doesn't make Beam **judge** better, but it helps Beam **point at the right line** more often, for no more spend than retrieval.",
"cross_file_cite_matched": "19/30",
"outside_cite_matched": "6/13",
"accuracy_matched": "0.821",
"matched_tokens": 3888241,
"matched_ledger_tokens": 3957434,
"matched_ledger_calls": 57,
"matched_mcnemar_p": "0.070",
"gold_file_retrieved": "8/13",
"gold_file_missed": 5,
"accuracy_frac_matched": "46/56",
"accuracy_frac_whole": "47/56",
"matched_ext_available": true,
"matched_ext_cross_whole": "52/78",
"matched_ext_cross_matched": "38/78",
"matched_ext_p": "0.0094",
"matched_ext_tokens": 7528878,
"matched_ext_outside_whole": "38/61",
"matched_ext_outside_matched": "25/61",
"matched_ext_outside_p": "0.0146"
},
"beamv3": {
"available": true,
"title": "Catching what one whole-repo call misses",
"claim": "32k docstring chunks recovered 7 of the 11 drifts hybrid missed; the docs-focused whole-context pass recovered 1; 3 prose-doc drifts are still missed. Union coverage 34 → 42 of 45. Marginal cost ~140k tokens per extra drift for 32k chunks vs ~1.5M for the docs pass. Precision of the new findings is a lower bound: 18.5% for hybrid + 32k-chunk new items (60.0% among labelled); most are unlabelled and are not counted as true. 7 new true drifts sit outside the labelled 45 (checked against the source by an AI reviewer). Stability repeat: 18/19 verdicts agreed.",
"recipe": "Recommended: symbol check + whole-repo Beam triage, plus 32k docstring chunks if you want more coverage at lower precision.",
"headers": [
"mode",
"findings",
"true",
"false",
"unclear",
"unlabelled",
"precision LB",
"among labelled",
"of 45",
"extra of 45 vs hybrid",
"new (AI)",
"tokens",
"tokens / distinct true",
"marginal tokens / extra drift"
],
"rows": [
[
"Hybrid (existing, for reference)",
"60",
"45",
"-",
"-",
"-",
"75.0%",
"-",
"34",
"-",
"-",
"4,739,392 (triage pass only)",
"139,394",
"-"
],
[
"Hybrid + repro strict (existing)",
"41",
"38",
"-",
"-",
"-",
"92.7%",
"-",
"30",
"-",
"-",
"4,877,530",
"-",
"-"
],
[
"(a) docs-focused pass alone",
"43",
"17",
"8",
"2",
"16",
"39.5%",
"63.0%",
"15",
"1",
"2",
"4,485,924",
"263,878",
"-"
],
[
"(b) py32k docstring chunks alone",
"72",
"18",
"3",
"5",
"46",
"25.0%",
"69.2%",
"13",
"7",
"5",
"1,680,568",
"93,365",
"-"
],
[
"Hybrid + (a): new items only",
"26",
"4",
"4",
"2",
"16",
"15.4%",
"40.0%",
"35",
"1",
"2",
"9,225,316",
"249,333",
"1,495,308"
],
[
"Hybrid + (b): new items only",
"65",
"12",
"3",
"5",
"45",
"18.5%",
"60.0%",
"41",
"7",
"5",
"6,419,960",
"139,564",
"140,047"
],
[
"Hybrid + (a) + (b): new items only",
"91",
"16",
"7",
"7",
"61",
"17.6%",
"53.3%",
"42",
"8",
"7",
"10,905,884",
"222,569",
"411,099"
],
[
"Hybrid + v3 new items strict-confirmed by repro",
"17",
"4",
"0",
"0",
"13",
"23.5%",
"100.0%",
"36",
"2",
"2",
"11,344,203",
"298,532",
"1,651,203"
]
],
"caveats": [
"This is a Benchmark-tab experiment, not a headline number. Headline coverage stays 34 of 45.",
"Precision LB = labelled-true / all findings. Unlabelled findings are not counted as true.",
"'Among labelled' is biased upward: the labelled items are the ones that hit known drifts.",
"15 new-item labels were checked against the source by an AI reviewer at the pack ref, not a human maintainer. Unclear counts as not-true.",
"3 still-missed drifts are prose-doc: C:attrs-py2-reprs, C:httpx-sslcontext, C:mm-quotes-400.",
"Whole-repo is one call per repo; the extra pass is one call per 32k Python-docstring chunk; docs is one focused whole-context call per repo.",
"This run's own ledger is 7,713,453 tokens over 168 calls. That is extra experimental cost, not part of the 9.3M hybrid end-to-end.",
"The shipped CLI is still symbol check + one whole-repo triage. 32k docstring chunks are an experimental extra pass (`scripts/beamv3.py`), not a `--docstring-chunks` flag."
],
"recovered": 8,
"missed": 11,
"py32k_extra_of_45": 7,
"docs_extra_of_45": 1,
"still_missed": [
"C:attrs-py2-reprs",
"C:httpx-sslcontext",
"C:mm-quotes-400"
],
"union_of_45": 42,
"hybrid_of_45": 34,
"new_ai_verified": 7,
"py32k_precision_lb": "18.5%",
"py32k_precision_labelled": "60.0%",
"py32k_marginal": 140047,
"docs_marginal": 1495308,
"stability_agree": 18,
"stability_n": 19,
"ledger_tokens": 7713453,
"ledger_calls": 168
},
"token_footnote": "~9.2M tokens (two whole-repo passes over 12 libraries). 4,424,575 first sweep + 4,739,392 hybrid triage. +138,138 for repro scripts = 9,302,105 end to end. Extra experimental runs, not in the 9.3M: ablation 1,948,469; beam v3 7,713,453; matched-cost control 3,957,434; matched-cost extension 7,528,878; tomlkit CLI run 185,842.",
"pack_coverage": {
"available": true,
"headers": [
"Library",
"Files packed",
"Tokens packed",
"Cap",
"Truncated",
"Skipped (binary/undecodable)",
"Last packed file"
],
"rows": [
[
"starlette",
"119",
"327,836",
"400,000",
"yes",
"0",
"tests/middleware/test_middleware.py"
],
[
"httpx",
"89",
"327,864",
"400,000",
"yes",
"0",
".github/PULL_REQUEST_TEMPLATE.md"
],
[
"click",
"73",
"327,831",
"400,000",
"yes",
"0",
"tests/test_types/test_Path.py"
],
[
"structlog",
"115",
"318,657",
"400,000",
"no",
"3",
"tests/processors/__init__.py"
],
[
"nox",
"103",
"294,382",
"400,000",
"no",
"0",
"tests/resources/sym_dir/sym_file.txt"
],
[
"arrow",
"50",
"280,937",
"400,000",
"no",
"0",
"tests/__init__.py"
],
[
"cattrs",
"170",
"262,894",
"400,000",
"no",
"0",
"tests/strategies/__init__.py"
],
[
"marshmallow",
"96",
"247,224",
"400,000",
"no",
"0",
"tests/__init__.py"
],
[
"attrs",
"116",
"327,857",
"400,000",
"yes",
"1",
".github/workflows/codspeed.yml"
],
[
"jinja",
"105",
"300,733",
"400,000",
"no",
"0",
"tests/res/__init__.py"
],
[
"werkzeug",
"138",
"327,823",
"400,000",
"yes",
"6",
".github/pull_request_template.md"
],
[
"pydantic-settings",
"56",
"311,093",
"400,000",
"no",
"0",
"tests/example_test_config.json"
]
],
"note": "truncated: true means the packer hit the remaining-budget cutoff while files were still in the queue. skipped_paths only records binary or undecodable files, so the dropped remainder is not listed. 5 of 12 first-sweep packs hit the remaining-budget cutoff (starlette, httpx, click, attrs, werkzeug). The packer puts tests, CI and examples last; exact dropped paths were not recorded in the traces. Starlette packed 119 files (327,836 of a 400,000 token cap), including test files; the last packed file was tests/middleware/test_middleware.py."
},
"recall_extra": true
}
|