A small experiment
One issue.
Two questions.
Can we tell evidence from guesswork? Give the lab one GitHub issue. Ask what the source says, then something it cannot know.
The point is simple: an answer is easier to trust when you can inspect its source and see where its knowledge ends.
Real actions and responses. Original script, synthetic voice and original music.
Download MP4 ↓The source
A public node-postgres issue proposing a SQL tagged template function.
The check
A source-backed question, an unsupported question, and a repeat query to inspect cache reuse.
The result
Cited evidence, a refusal where evidence is missing, and a reusable answer with its citations attached.
This recording uses keyword retrieval and cited extracts; no language model generates its answers. The graph uses extracted structure and inferred lexical connections. Imports on this unconfigured deployment are temporary. Persistent storage and hybrid/model-backed runs require Postgres and provider configuration.
Read the full experiment transcript
00:00One issue. Two questions.
One issue. Two questions. Can we separate evidence from guesswork?
00:05Start with one real issue.
Start with a GitHub issue. Import it, and keep the search focused on that source.
00:14Find the reason in the source.
Ask how the proposal reuses SQL statements. The answer quotes the source. Follow the citation to check the original.
00:24See where the knowledge lives.
Open the graph. Each clue leads back to a source location. Follow its branches to see how the information connects.
00:33Now test the boundary.
Now ask about Tokyo weather. This source cannot answer that. The lab refuses, instead of inventing evidence.
00:42Ask the first question again.
Repeat the first question. The exact match comes from the cache, with its citations attached. Less repeated work.
00:52A source. A citation. A boundary.
Try your own issue. This demo uses extracts, with temporary imports.
Try the same check on a source you know.
Bring your own issue →