| Conf. | Type | Category | Lesson | ✓/✗ | Last confirmed |
|---|---|---|---|---|---|
| 0.8894 | fact | methodology | Lexical (ts_rank) demand retrieval fabricates evidence (matched 45Z clean-fuel credit to a SaaS support job ad); semantic embedding retrieval at cosine>=0.72 discriminates correctly. | 1/0 | 07-10 |
| 0.84 | fact | methodology | Connected-components clustering on semantic edges chains transitively into hairballs; mutual-kNN plus hairball-splitting yields tight, useful clusters. | 1/0 | 07-10 |
| 0.7953 | heuristic | methodology | Federal Register agency feeds must be filtered to type=RULE/PRORULE; Notices (exemption applications, airworthiness directives) are not filing mandates and waste reasoning budget. | 1/0 | 07-10 |