# Input inventory: HackerTube-inspired astronomy sample

## Scope and completed work

All five supplied attachments were read with Python in the session notebook on 2026-11-03. JSONL was parsed in the kernel; the full comment corpus was not printed. This inventory checks input structure, joins, access states and provenance. It does not assign recommendation stances, select a watchlist, verify video contents, or evaluate HackerTube's actual catalogue. No external source was re-fetched. No figures were necessary for this inventory; any later figures should use English and PNG dimensions no greater than 2400 pixels on the longest side.

## File inventory

| Input | Exact notebook path | Size (bytes) | Scope and role | Main limitation |
|---|---|---:|---|---|
| analysis_notes.md | inputs/analysis_notes-f3f21ece6fce.md | 5863 | Analysis contract: pair-level stance/topic judgments, evidence review, selection and reproducibility rules | Instructions, not observed recommendation labels or a gold standard |
| source_brief.md | inputs/source_brief-5c883d784200.md | 3972 | Sampling description, provenance and declared limitations | Collection claims are supplied documentation; original 40-hit response and excluded hits are not separate inputs |
| hn_comments.jsonl | inputs/hn_comments-78c23616d7b5.jsonl | 146388 | 23 selected comment records; official HTML, convenience text, search snapshot, links, timestamps, nearest parents and root story | Bounded context, not full discussions; official HTML is authoritative wording |
| comment_video_links.csv | inputs/comment_video_links-f9ff64257870.csv | 10756 | 38 rows × 9 columns; one distinct comment–video mention per row | Mentions are not recommendations; different links in one comment can have different stances |
| video_metadata.csv | inputs/video_metadata-dcc11572ab87.csv | 13908 | 36 rows × 14 columns; one observed YouTube oEmbed result per video ID | Titles/authors only where retrieved; no transcripts, full descriptions or content validation |

Exact input Version IDs, in the table order above: bd9286fa-a257-45a4-83fe-42f0365f6e7c; 6af857fa-c399-40d6-b115-de4694fb8f22; 46fee011-ffdd-45cd-8ba3-f83e6d448cf6; 56eccfa1-372f-4840-8922-1c02ae0eb962; 7f4562ce-f658-44d7-bc4a-ad0a095fbf6d.

## Provenance and sampling boundary

The source brief attributes the inspiration to [Karpathy's HackerTube post](https://karpathy.bearblog.dev/hacker-tube/) and its [HN discussion](https://news.ycombinator.com/item?id=50319510). Its cited whole-site figures (roughly 40,000 recommending comments, 6,454 videos and 150 topics) are attributed claims, not measurements verified in this inventory.

The documented [Algolia query](https://hn.algolia.com/api/v1/search?query=astronomy+youtube.com&tags=comment&hitsPerPage=40) is `astronomy youtube.com`, `tags=comment`, `hitsPerPage=40`, frozen at 2026-11-03T05:26:08.015008+00:00. The brief says the first 40 relevance-ranked hits were filtered mechanically for specific 11-character YouTube IDs in watch or short links, retaining 23 comments and excluding channel-only hits. The package permits checking the retained data, not independently reproducing the exclusion of the other 17 hits or assessing all HN comments. Keyword matching and relevance rank are not astronomy labels, endorsement labels, random sampling, or quality rankings.

Selected wording is attributed to the official HN Firebase API, with direct source-comment URLs. Recorded comment retrieval spans 2026-11-03T05:26:08.937987+00:00 to 05:26:20.281826+00:00. Selected comment creation dates span 2009-12-18T16:03:36+00:00 to 2026-10-06T17:31:06+00:00. Thus this is a recent retrieval of comments across many years, not a sample of only recent comments.

## Python checks and actual counts

- 23 selected comments, 38 distinct comment–video pairs, 36 unique video IDs.
- No duplicate pair rows, no duplicate metadata IDs, and no invalid 11-character video IDs.
- CSV pair set exactly matches the IDs listed in the JSONL selected comments; metadata ID set exactly matches the linked video ID set. This is internal consistency, not an independent reconstruction of all URLs from HTML.
- All 23 selected comment records and all 23 root-story records report available. All 35 stored parent-item occurrences report available; this occurrence count is not a count of 35 distinct additional comments and may include story items.
- All 23 stored search-text-match flags are true. These flags report agreement with the official snapshot; this inventory did not independently authenticate upstream responses.
- 18 comments mention one video; one comment each mentions 2, 3, 4, 5 and 6 videos. These reconcile to 23 comments and 38 pairs.
- Video IDs `9IlmekWsEAQ` and `dhbAEx2le7o` each occur in two comments; the remaining 34 videos occur once. Repeated mentions are not votes for quality.

## Missing context

Three parent chains reached the three-item cap: [32605650](https://news.ycombinator.com/item?id=32605650), [47936513](https://news.ycombinator.com/item?id=47936513), and [46315093](https://news.ycombinator.com/item?id=46315093). A separately retrieved root story does not restore omitted intermediate comments. Even untruncated chains do not supply every sibling, reply or surrounding discussion. Parent wording must retain its own attribution rather than being assigned to the selected author. External story URLs do not constitute retrieved article contents.

The inputs contain no pre-generated semantic judgments, independent human reference labels, or inspected HackerTube watchlist against which to measure agreement. A later audit must interpret each link in its own comment context, distinguish recommendation from reference or criticism, and assess astronomy fit separately. An astronomy example inside a software discussion need not make an entire video an astronomy video.

## Metadata coverage and limits

Recorded oEmbed checks span 2026-11-03T05:26:34.187365+00:00 to 05:26:40.625509+00:00. Metadata is available for 33/36 videos (91.7%); three rows report HTTP errors and have blank titles and authors:

| Video ID | Recorded HTTP response | Interpretation |
|---|---|---|
| QCIRkD71do4 | 403 Forbidden | Metadata unavailable at retrieval |
| tJOJ6B2XOyA | 404 Not Found | Metadata unavailable at retrieval |
| ubaX1Smg6pY | 404 Not Found | Metadata unavailable at retrieval |

These responses do not by themselves establish deletion, permanent unavailability or regional playback restrictions. Preserve blanks rather than filling from memory. All 36 rows record `playability_checked=false` and `content_viewed=false`. Successful oEmbed supplies a public title and author attribution, not verified content, scientific accuracy, educational quality, full description, transcript, duration, or playback evidence. Title-based topic screening remains provisional.

## What this inventory supports next

The retained sources can support an evidence-backed audit of this bounded sample. Before selecting up to five unique videos, freeze all 38 actual pair-level judgments with separate stance and topic-fit labels and attributed short evidence. Review against original HTML and available context, retain the first pass, and log only evidence-supported revisions. Deduplicate by video ID for selection while preserving separate comments and any conflicting stances; retain omitted/deferred candidates and their reasons. Fewer than five choices is valid. AI review agreement is not human accuracy, and this convenience sample does not justify population accuracy or whole-site claims.

## Execution record

Python notebook runs: `notebook-run-1793684410580-1` (read all inputs and initial inspection), `notebook-run-1793684453043-2` (source brief and schema inspection), and `notebook-run-1793684470866-3` (aggregate and integrity checks). No semantic-input judgment tables were created or used. Assistant runtime identifies itself as Codex based on GPT-6; an exact deployment/build version is not exposed. The inventory prose synthesizes the observed checks and supplied documentation; semantic judgments are not deterministic outputs of Python.
