?

Cross-model consensus: edges survive only when both bge-large and qwen3-0.6b place the target in this channel's top-k. Edge weight = geometric mean of the two cosines. Layout = UMAP on 1 − geomean.

— pick a channel — A Quality ExistenceAdams FallAm I DrivingAndrea with the BangsAphrael PilotsonBrett AndresBrian of College StationCassidy van der KampChad Matthew AlvinChannel UCiJmdXTb76iCharlie's Little CornerChris HowardChristian BaxterClimbing Mt SophiaCorner CitizenEin Sof - Infinite ReflectionsEmily RajehEolson-helperEric SeitzEthan CaugheyEverett Vroon (EJV)EzrafaturechiFinding IdeasFree RilianFrom Whom All Blessings FlowGrail CountryGrey HamiltonGrizwald GrimHow to Understand the WorldJohn VervaekeJonathan DumeerJonathan PageauJordan B PetersonJordan HallJoseph LambrechtJustinsmorningcoffeeK Joseph RivetKale ZeldenLucas VosLuke ThompsonMarc JacksonMark D ParkerMark SMary KochanmcmosavMore ChristNeal DaedalusNechama GluckNewBeeNickolas ShermanNuance ForumOG RosePaul AnleitnerPaul Rene NicholsPaul VanderKlayPresident FoxmanRafekelleyRandos UnitedRandos United 2Rebel WisdomRestoring MeaningRigel Windsong ThurstonRightinchristRon CoppermanSartori StudiosSecular KoranismShoulder SerfSkankensteinStrange TheologyThe AnadromistThe Chris ShowThe Meaning Codethe plebistocratThe Young LeviteTheCommonToadTheScrollersPodcastTransfiguredUpCycleClubWavesofobsession

Semantic Descriptions Citations

sem: ? · desc: ? · cite: ? · comms: ?

Gravity (repulsion): -10000 Central pull: 0.05 Spring strength: 0.02

label distinctiveness: 1.00 (tf-idf)

Pipeline

1. Source: ~22K primary YouTube video transcripts (Canary ASR preferred, YouTube auto-captions fallback) for channels listed in channels.yml.
2. Chunking: each transcript split into ~400-token sentence-aware chunks. ~32 chunks/video.
3. Embedding: each chunk encoded with BAAI/bge-large-en-v1.5 (1024-d, fp16). Stored in Qdrant collection tlc-graph-research.
4. Channel centroid: L2-normalized mean of a channel's chunk embeddings.
5. Distinctive direction (key step): centroid − α · corpus-mean direction, renormalized. Removes the shared "long-form intellectual English" baseline so cosines spread from a tight 0.85–0.99 band into roughly −0.3 → +0.9. α is set at build time via --distinctive-alpha (default 1.0); compare builds at different α to see how much of the structure is baseline-suppression vs. real signal.
6. Cosine similarity on the distinctive centroids → the matrix used everywhere downstream.
7. k-NN graph: each node keeps its top-K most-similar neighbors as edges.
8. Communities: three algorithms are pre-baked for every k ∈ {2..15} and switchable via the algo dropdown:
   • Louvain — classic modularity maximizer. Default; widely used; can leave "badly-connected" nodes in clusters they barely belong to.
   • Leiden — strict improvement over Louvain (Traag et al. 2019). Adds a refinement step that fixes the Louvain bug. Often produces tighter, more meaningful clusters.
   • Infomap — different paradigm: minimizes the description length of a random walker on the graph. Finds "flow communities" where information tends to dwell. Often disagrees with modularity methods on borderline pairs; agreements across all three are robust.
   Stability and consensus annotations below are Louvain-based (computed from Louvain bootstrap and three-signal Louvain partitions) so they describe pair-level robustness independent of which display algorithm is active.
9. Layout: UMAP(centroids, cosine, n_neighbors=10, min_dist=0.20) → 2D positions.
10. Cluster labels: per-cluster TF-IDF (unigrams + bigrams, stopword-filtered) over each member channel's concatenated chunk text. Each cluster's top label terms are picked by score = cluster_tfidf_mean − α · global_tfidf_mean (positive only). The label distinctiveness slider tunes α live: α=1 shows only above-corpus-baseline terms (default), α=0 surfaces top-by-raw-frequency terms (common-but-cluster-strong words come back), intermediate α blends the two.
11. Description edges (overlay): a parallel embedding of each video's description text (one embedding per video, from collection tlc-descriptions-research). Channel centroid + distinctive-direction + k-NN, like semantic — but over the metadata register (titles, episode notes, links, hashtags) rather than spoken content.
12. Citation edges (overlay): aggregated from Elasticsearch internal_references. A → B counts how many of A's video descriptions link to one of B's videos. Channel-level, directional.

Inspector columns

• sim — pairwise cosine on the distinctive centroids. Positive = aligned content direction. ~0 = generic. Negative = anti-aligned. Typical range −0.3 → +0.9.
• z — how many σ this neighbor's similarity is above the mean of the selected channel's similarity row. >1σ noteworthy, >2σ strong, ≥3σ a standout. Says "unusually close for this channel", beyond the global baseline of "everyone's close-ish."
• Δ-next — gap in similarity to the next-ranked neighbor. A large gap is a cliff: the listed neighbor is qualitatively more similar than whoever's next. >0.02 meaningful; >0.05 a strong break.

Confidence

Two scores quantify how trustworthy each clustering claim is, baked per edge:

• stability (0–1) — bootstrap: Louvain is run 60 times with different random seeds on this channel's k-NN graph. This pair's stability is the fraction of runs they end up in the same community. 1.0 = always co-clustered; 0.5 = coin-flip; near-zero = co-clustering is an artifact of one particular seed. Surfaced in edge tooltips and filterable via the min cluster stability slider.
• consensus (0–3) — cross-signal: three independent Louvain partitions (transcript k-NN, description k-NN, citation-coupling Jaccard) all run on the same channel set. This pair's consensus is the number of those three partitions that group them together. 3/3 = three different evidences agree; 1/3 = only one signal sees the tie. Surfaced in tooltips and filterable via the min signal consensus slider.

A third per-channel score, internal coherence, shows in node tooltips. It's the mean cosine of the channel's own chunks against its centroid — high means the channel speaks about coherent topics; low means broad/jumping content. Doesn't filter anything; just tells you which centroid positions are well-defined.

Controls

• top-k neighbors: how many edges per node (Louvain runs separately for each k).
• edge floor: only render semantic edges with sim ≥ floor.
• Semantic / Citations: toggle each edge layer.
• citation count [min..max]: filter the citation overlay to a band.
• citation pull / semantic pull: when Physics is on, edges become springs whose rest length shortens with citation count / similarity. 0 = that layer doesn't contribute to physics.
• Gravity (repulsion): Barnes-Hut all-pairs repulsion. More negative = nodes push each other harder; communities separate.
• UMAP Reset: snap nodes back to original UMAP positions.
• De-overlap: short burst of collision-only physics to nudge overlapping nodes apart with minimal layout disruption.

Caveats

• Dense embedders compress within-domain similarity. All TLC channels are long-form intellectual English. The distinctive-direction step is what makes relative comparisons read; trust orderings, z-scores, and gaps more than raw cosine magnitudes.
• Channel size affects centroid noise. Large channels (PVK ~113K chunks) have very smooth centroids; small channels (<200 chunks) are noisier and may sit slightly oddly.
• Citation overlay is incomplete. ~2K reference target videos didn't resolve to any channel in our index; another ~4K reference channels weren't in channels.yml and were dropped.