# Small evidence audit: final check record

## Outcome

The supplied sample still contains 23 selected comments, 38 distinct comment–video pairs and 36 unique video IDs. All pairs remain in final_judgments.csv. Five unique selections and their six source relationships are unchanged; the other 31 videos remain in final_omissions.csv. Four judgment rows received evidence or wording clarifications; no stance or topic labels changed. This is a small, non-blind model evidence audit, not a measured accuracy benchmark or comparison with a human gold standard.

## Files and evidence checked

Original inputs: analysis_notes.md, source_brief.md, hn_comments.jsonl, comment_video_links.csv and video_metadata.csv. Checks used their exact supplied notebook input versions. Authoritative comment HTML was parsed with HTML entities decoded; source text and supplied parent/root context were used to assess meaning. Convenience text was not substituted for the authoritative source in quotation validation.

Saved structured files checked: comment_video_mapping.csv, comment_video_judgments_first_pass.csv, provisional_selection.csv, provisional_omissions.csv, evidence_audit_and_revisions.csv, final_selection.csv, final_selection_qualified.csv and metadata_review.csv. Both saved list versions, provisional_watchlist.html and qualified_watchlist.html, were read and checked. The earlier inventory prose is available in the conversation's artifact-creation record; its separately saved artifact bytes were not available in the notebook and were not independently re-read. This limitation does not affect checks against the original inputs.

SHA-256 values were taken before checking for the five inputs, eight structured files and two HTML lists. All 15 file hashes were unchanged at completion. The original first pass, provisional selections, original HTML lists and prior audit files were preserved rather than overwritten.

## Check results

| Check | Finding |
|---|---|
| Original-link reconstruction | HTML hrefs plus visible-text URLs recover exactly the CSV's 38 pairs. Video IDs are normalized without merging different source comments. The text-only link is retained. |
| Mapping and judgment coverage | Both match all original pairs, with no duplicates or extra pairs. |
| Source identities | Selected comment URLs, evidence URLs, parent item IDs and root-story IDs match the supplied records. Every selection source ID belongs to the cited video pair. |
| Original URLs | Mapping URLs normalize to their row's video ID. |
| Short extracts | Every nonblank extract is an exact substring of HTML-derived source wording. No new parent quotation is introduced. |
| Selected-video metadata | Titles and author names exactly match supplied metadata in provisional, audited and qualified tables and escaped HTML text. |
| Full metadata review | All original metadata columns, including blanks, errors and access flags, are preserved across all 36 IDs. |
| Selected/omitted partition | 5 selected + 31 omitted/deferred = 36; no overlap or discarded candidate. All separate pair stances and uncertainty are retained. |
| Selected-entry audit | Six source pairs support five unique entries. Direct endorsement and favorable-list endorsement are distinguished. |
| HTML references | Both versions link all six selected source comments. The qualified page additionally links the three metadata-error video IDs. No executable scripts or embedded players occur. |
| Counts | 25 recommend, 13 neutral/reference, 0 critical, 0 unclear stance. Topic fit: 21 yes, 13 no, 4 unclear. These count pairs, not unique videos. |
| Metadata counts | 33 HTTP 200, 1 HTTP 403, 2 HTTP 404. All 36 have recorded request times, untested playback and unviewed content. |
| Figure consistency | Recomputed tabular counts reconcile with the recorded chart-generating counts, including 59 network nodes and 38 edges, and the six-row selected-entry matrix. All five saved PNGs were opened for dimension checks; longest sides are at most 2400 pixels. Pixel-level visual/OCR validation was not performed in this audit. |

## Corrections and their limits

The revision log retains all 38 rows, including 34 unchanged rows. It records old/new labels and reasons, source references, changed fields and quote lengths. It does not reproduce old and new source extracts again.

1. Comment 49536680 / video dhbAEx2le7o: removed one redundant four-word extract. The same source's beginner-guide row retains the list-endorsement evidence, and the book-video relationship and reasoning remain. See the quotation-budget note below.
2. Comment 17036313 / video Jdy09y0A4t0: clarified that praise of technical treatment applies to the channel. The contextual recommendation remains an interpretation of appreciative presentation, not an explicit video-specific quality assessment.
3. Comment 25299540 / video GutJf9umD9c: removed an unverified film/scene identification. The supplied title and cryptography context are enough to support the narrower reference/off-topic judgment.
4. Comment 47579661 / video 4J9MRYJz9-4: removed an unverified comedy characterization. The title and favorite-number context support the retained topic judgment without that addition.

None of these changes alters the five selected entries. All selected reasons remain supported within their stated limits. In particular, criticism in comment 47936513 targets the parent claim about image noise; the linked imaging videos are explicitly appreciated. Repeated Ed Ting references remain separate endorsements, not votes for quality. The selection order is editorial viewing order, not a measured ranking.

## Quotation accounting across report and cards

source_quote_ledger.csv covers 58 unique supplied selected/context/root item IDs, including parents with zero quoted words. Words are counted with a Unicode word-token rule retaining apostrophes and hyphens inside words; repeated extracts within a table count each time. The original maximum within the first-pass table was 18 words per source.

The provisional HTML has no formal quotation blocks, but it repeats one four-word phrase present in comment 49536680's extracts. Conservatively counting this wording as source quotation raises that source's first-pass table plus both cards total to 21 words. This historical edge case is recorded, not erased or described as a clean historical pass.

The final judgment table removes the redundant extract for that source. Its total becomes 13 words in the final table plus 4 in the preserved provisional card = 17. The qualified cards and this check report reproduce no source extracts. Across the final table plus both cards and this report, the maximum for any source is 18 words, within the 20-word limit. Parent quotations contribute zero. Exact source matches in titles are treated as attributed metadata titles rather than comment-extract quotations.

The active release uses final_judgments.csv. Archived first-pass and final tables are alternative versions, not a combined publication: concatenating their repeated quotations would exceed the budget for some sources. The old and final totals are both exposed in the ledger so downstream reports can choose the correct version and avoid repeating extracts.

## Uncertainty retained

All four unclear topic cases remain: uD4izuDMUQA, -Pc3IuVNuO0, QCIRkD71do4 and hSuSG19Pcoc. Their possible relationship to astronomy does not establish the linked video's subject focus. Unavailable metadata remains blank for QCIRkD71do4, tJOJ6B2XOyA and ubaX1Smg6pY. HTTP errors do not establish deletion or current/region-specific playback restrictions.

The selected Starlink video's title differs between the comment and oEmbed snapshot; its exact ID matches, but the cause of the difference is unresolved. The imaging source's parent chain remains truncated. No missing parent text was inferred. Equipment suitability, historical currency, factual correctness, content quality and playback remain unverified. No video was watched and no new network requests were made.

## Final outputs and reproducibility boundary

- final_judgments.csv: actual final model judgments for all 38 original pairs, including unchanged and uncertain cases.
- final_selection_audited.csv: five selections, ranks, reasons, source IDs and metadata limitations; references the final judgment table.
- final_omissions.csv: all 31 remaining videos, with their original dispositions and updated final pair evidence.
- revision_log.csv: 38 review outcomes with four explicit clarifications and zero label changes.
- source_quote_ledger.csv: per-source first-pass and final quotation budgets, including the preserved cards.

Python verified structure, IDs, exact extracts, counts, metadata preservation, file preservation and dimensions. Semantic interpretations and editorial choices were made by the assistant, not deterministically inferred by Python. Execution model identity provided by the runtime: Codex / GPT-6; exact deployment/build version is unavailable. Audit executions: notebook-run-1793685641535-26, notebook-run-1793685701763-27 and notebook-run-1793685769246-28. The final table-generation run performed the final assertions and confirmed preservation of earlier files.
