Lattensepp Analysis

Lattensepp / IT-Mario Replication — V2.5

Rebuilds the YouTube-analysis pipeline from IT-Mario's video "Ich wurde wissenschaftlich auseinandergenommen" three different ways on the same 150-video corpus, then scores each method against his declared winners. The goal is a legible, side-by-side comparison of a hand-written keyword script, three parallel AI agents, and a hybrid that mixes deterministic math with extended keyword lists.

Headline numbers

• 10 of 13 metrics (77%) agree with IT-Mario's declared winners across all three methods.
• Corpus: 150 videos (50 per creator: apored, maithinkx, papaplatte) plus 18,146 YouTube comments.
• Total compute cost: $0.27 ($0.24 for the three parallel AI agents plus $0.03 for the Claude-via-OpenRouter censorship-density analysis). The deterministic methods are free on a laptop.

Source: `video_charts/05_bottom_line_summary.json`.

The three methods

1. IT-Mario replica. His exact 5-6 word keyword lists as stated on camera. Deterministic. Reproduces what he would have computed with his own declared inputs.
2. AI agents. Three parallel Claude agents, one per creator, each given the metric spec and asked to count and normalize. No shared state between agents.
3. Hybrid. Extended keyword lists plus simple stemming, a 234k-word English dictionary for the denglisch metric, and pure math for TTR / word length / token count. This is our best-of- breed internal diagnostic, NOT a truth claim; the public accuracy number is each method's agreement with IT-Mario's declared winners. See docs/CURRENT_STATE.md for the full primary/ secondary reference split.

A fourth deterministic "script" method (wider keyword lists, no stemming) lives in the repo as an internal baseline but does not appear in the video narrative.

Dataset

• 150 German YouTube videos: 50 each from apored, maithinkx, and papaplatte.
• 18,146 comments scraped from a subset of those videos.
• All data is regen