|
| 1 | +# CPU knowledge retrieval evidence |
| 2 | + |
| 3 | +Acceptance scope and ownership: [InferRSI #190](https://github.com/InferMatrixLab/InferRSI/issues/190). |
| 4 | +This packet tests a manifest implementation that preserves retrieval behavior. |
| 5 | +The frozen quality/parity and paired CPU performance gates passed. Campaign |
| 6 | +acceptance, merge and issue closure remain with the authorized owner. |
| 7 | + |
| 8 | +| Segment | Contents | |
| 9 | +| --- | --- | |
| 10 | +| `plan/` | Prospective protocol, current freeze, preserved first freeze and pre-measurement review amendment | |
| 11 | +| `validator/` | Standard-library runner, 16 development and 8 withheld synthetic cases with source-grounded partial relevance labels | |
| 12 | +| `results/` | Complete quality comparison, paired raw API/CLI measurements, baseline profile, #129 delivery trace, editable EN/ZH architecture artifacts and qualified local software/environment summaries | |
| 13 | +| `patch/` | Fixed runtime candidate and exact source identity | |
| 14 | + |
| 15 | +The current prospective freeze is |
| 16 | +`a465367319f3058398eb1f8579ee5d4174b8cb90457793e90788fb5796151057`. |
| 17 | +Independent review corrected completeness checks and removable assertions before |
| 18 | +formal candidate measurement; `plan/amendment-1.md` records the change. The |
| 19 | +withheld split was released only after the runtime candidate was fixed. Fixture |
| 20 | +text and labels did not change between freezes. Root `.gitattributes` preserves |
| 21 | +evidence bytes; copying or committing must retain the recorded hashes. |
| 22 | + |
| 23 | +The full quality result has **96 outputs per revision, exact parity and unchanged |
| 24 | +source/corpus identities**. For 20 labelled answerable cases, search/default |
| 25 | +collect score Success@5/MRR@5 = 0.20/0.20; relevance/context-item ranking scores |
| 26 | +1.00/0.9625. Four labelled negatives return empty items in 4/4 search/default |
| 27 | +cases and 2/4 relevance/context cases. These are unchanged existing #129 |
| 28 | +behaviors, not a quality improvement from this runtime patch. |
| 29 | + |
| 30 | +Context-item ranking does not establish correct body delivery. A post-hoc |
| 31 | +inspection found the labelled owner's context delivered in **19/20** positive |
| 32 | +cases. `withheld06` finds its correct item first but requires 170,782 context |
| 33 | +bytes against a 65,536-byte budget, so that owner's context is omitted. This |
| 34 | +diagnostic does not change the frozen quality gate. Synthetic, AI-authored, |
| 35 | +partial source labels are not human gold judgements or complete Recall; no |
| 36 | +agent-task efficacy or production-distribution claim follows. |
| 37 | + |
| 38 | +The 20-pair CPU result observed **40.01%** hot-relevance paired-median latency |
| 39 | +reduction, 95% interval **[38.38%, 41.06%]**, with all 400 output pairs and |
| 40 | +80 CLI stdout pairs equal. [performance.md](results/performance.md) retains |
| 41 | +all per-mode/query p50/p95 rows, including search regressions and CLI relevance |
| 42 | +decode p95 **1.810 → 3.046 ms**. Independent read-only recomputation passed; |
| 43 | +reviewer model/effort is unknown and no human/maintainer acceptance is recorded. |
| 44 | + |
| 45 | +See the [audit](../../docs/knowledge-retrieval-audit.md) for source contracts, |
| 46 | +per-query examples, failure qualifications and follow-up priorities. Local |
| 47 | +software verification includes a failed whole-suite run whose initial failures |
| 48 | +were reproduced on baseline, successful focused manifest checks and successful |
| 49 | +isolated wheel/sdist construction; it must not be described as a green full |
| 50 | +suite. No model, GPU or inference-service benchmark ran. |
0 commit comments