Evidence before synthesis.
YouTube Intelligence is a free provenance-first methodology demo: Stage 0 and Stage 1 foundation work passed, Stage 2 evaluation remains provisional, and graph work has not started.
A transparent demonstration of how a knowledge system can preserve sources, exact locators, changing state, and disagreement instead of presenting every generated summary as truth.
What has passed
Internal checks established a reproducible corpus foundation, honest locator handling, a valid SQLite database, FTS5/BM25 indexing, real pinned embeddings, deterministic retrieval paths, and retrieval-run records. The latest independent review verified 9,036 transcript segments, 954 valid chunks, 954 embeddings, database integrity, schema migration 8, and focused Stage 0/1 plus Stage 2 tests.
Missing timestamps must remain explicitly missing. The public method never renders 00:00:00 as if it were valid evidence.
What remains provisional
The frozen 50-question evaluation uses provisional, not-human-reviewed relevance labels generated from a hard-coded expected source followed by an in-source lexical locator. That construction likely favours the token-overlap baseline.
Frozen Recall@10 results
- Token overlap
- 0.58 baseline
- FTS
- 0.12
- Embeddings
- 0.34
- RRF
- 0.36
- RRF + reranking
- 0.38
The best Stage 2 result is 0.20 below baseline. These metrics are diagnostic evidence, not a public quality guarantee.
Why graph work has not started
The operating doctrine requires retrieval and provenance gates before graph extraction. Stage 3 remains stopped until the frozen evaluation is honestly adjudicated, ranking and compatibility issues are remediated, and the predefined quality criteria pass without relabelling evidence to improve a score.
There is no completed knowledge graph, contradiction engine, agent runtime, or automated truth-resolution system to demonstrate publicly.
The public demo method
- Select a small rights-checked set of public sources.
- Preserve each source and exact timestamp or honest file/chunk locator.
- Extract only claims supported by the located material.
- Separate shared conclusions, scope differences, direct contradictions, and temporal updates.
- Label synthesis as an inference and show its premises.
- Publish limitations beside findings rather than hiding them in a technical appendix.
A sample synthesis pattern
Convergence
Reliable agent systems repeatedly benefit from provenance, controlled change, explicit checks, and selective human approval.
Scope tension
Vectors, graphs, and structure-aware retrieval solve different question types; one method should not be marketed as replacing all others.
Current-state rule
Historical events remain immutable while a newer assertion may supersede stale current state without deleting the evidence trail.
Rights and privacy
This page summarizes methods and short evidence points; it does not redistribute full transcripts, audio, or video. Original public sources should be linked in any published evidence brief. No email is required to read this methodology page.