No LLMs. No generated consensus. Deterministic NLP only.
The project uses lightweight natural language processing to make long technical discussions easier to inspect. It highlights likely evidence, keeps the surrounding sentence, and points back to the original post.
Rule-based signal map
Source sentence
I agree with the direction, but can we clarify the migration risk?
Agreement
agree
Disagreement
risk
Question
clarify?
Progress
revised
Long threads are hard because the important movement is distributed across dozens or hundreds of replies. Kirigami turns the thread into a map: opening context, agreement evidence, disagreement evidence, questions, progress, phases, authors, and source posts.
The reader does not claim that a thread reached consensus. It identifies posts whose language looks like evidence for agreement, disagreement, uncertainty, or progress, then keeps the original source one click away.
The goal is compression for navigation, not replacement. Issue cards list every source post they use, and colored evidence controls filter the review list without removing access to the full source.
How the goals are achieved
The pipeline is intentionally conservative: clean text, identify discourse acts, extract local evidence, rank by interaction, and keep everything tied to source posts.
Post HTML is stripped into plain text, whitespace is normalized, and code blocks plus headings are removed from signal detection. This avoids treating code snippets or section titles as participant intent.
Each post is scanned for phrase families that usually mark agreement, disagreement, questions, progress, concessions, revisions, or resolution. A post can appear in more than one category because real discussion posts often do more than one thing.
When a signal phrase is found, Kirigami extracts the sentence around it. If that sentence is too short to be meaningful, nearby sentences are included so the evidence is readable in context.
Signals are scored by the number of matched terms plus lightweight conversation structure: replies and quotes add weight because they suggest that other participants interacted with that post.
The analysis groups posts by author, counts quoted targets, estimates reading time, groups evidence into issue cards, and divides the thread into phases using post order and large gaps in time.
Issue cards show the complete set of source posts for that issue. Clicking a colored bar segment or matching colored chip filters those posts to agreement, disagreement, questions, progress, or other signal types; reset returns to the full list.
NLP examples
Not an LLM pipeline.
The analysis is transparent string processing and conversation structure: phrase lists, sentence windows, quote counts, author grouping, and time gaps.
This is deliberately simple NLP. The system uses curated phrase lists for discussion acts instead of asking a model to summarize intent. That makes the behavior inspectable and easy to correct.
Kirigami approximates sentence boundaries, finds the sentence containing the matched term, and expands to neighboring sentences when the extracted evidence would otherwise be too thin.
Discourse quote metadata identifies which posts were quoted by later replies. Frequently quoted posts are treated as structurally important, regardless of whether the quote is supportive or critical.
Posts are grouped by author to show participation patterns. The thread is also split into phases using order plus time gaps, which helps distinguish the initial burst from later follow-up.
The dashboard groups related evidence into issue cards. Each card lists all source posts it uses, and the colored signal bar plus matching chips filter that card's source list by evidence type.
The plain text word count is divided by 220 words per minute. This gives a practical sense of source-reading cost before someone opens the full thread.