Validation results

The recorded runs through 25 October 2026 show how the classifier handles evidence, applies eligible recommendations, and leaves exceptions for review. Each apply performs fresh inference rather than replaying the preceding dry run. Confidence is the model's assessment; these results do not establish a universal accuracy rate.

Batch	Dry run	Apply
1	20260927T030125Z	20260927T031250Z
2	20261024T082708Z	20261024T103009Z
3	20261024T163649Z	20261025T002722Z
4	20261025T025418Z	—

Current Batch 4 dry runs (oldest to newest): 20261025T005345Z; 20261025T025418Z.

The three completed applies recorded 19, 17, and 14 successful bookmark updates, respectively: 50 distinct bookmarks in total. They left one, three, and six medium- or low-confidence results unchanged, with no failed classifications or writes reported. Skipped bookmarks can recur in subsequent batches, so selected and skipped counts must not be added as if each run used a disjoint sample.

The second Batch 4 dry run recorded 13 applicable high-confidence recommendations, six medium-confidence recommendations, one blocked high-confidence taxonomy gap and no classification errors. All five eligible metadata-only reassessments remained medium. Neither Batch 4 snapshot records a write.

Demonstrated behavior and limits

One bookmark shows why a skip can be worth revisiting. In Batch 2, the direct request for Advanced AI Guide returned HTTP 200, but the classifier rejected the content it received and had only metadata to work with. The apply left it at medium confidence and made no change. The next batch's dry run used saved archive text and recommended the AI collection with high confidence. The separate apply recorded a successful write to that collection. Batch 2 apply · Batch 3 dry run · Batch 3 apply

Other completed writes used complete one-pass text, evidence processed across multiple requests and text retrieved directly from an original page. The processing mode alone did not determine whether a recommendation could be applied.

In the second Batch 4 dry run, repairing the structure of a model response left its taxonomy gap intact. The bookmark stayed in Unorganized with a collection suggestion for human review. Missing content, challenge pages, unavailable videos and mismatched page content remain visible in the individual reports.

High-confidence recommendations can also change between runs. The two retained Batch 4 dry runs choose different collections for the Data Industry Primer and How to be a -10x Engineer bookmarks despite unchanged saved-text hashes. These snapshots preserve the observed decisions without treating either choice as independently verified.

The public copies omit operational links. Machine-readable records, rollback state, page bodies, credentials, and runtime telemetry remain unpublished.