The system turns a broad question into testable claims, researches both sides, constructs one typed reasoning graph, and derives the answer and visual from that same artifact.
Three web-enabled models independently map current debates. Each keeps questions separate from candidate answers and links every hypothesis to a question.
Canonicalize questions without seeing hypotheses. Distinguish object questions from higher-order questions about whether an effect matters, then map their dependencies.
Independent models receive the questions and neutral context, but not the discourse hypotheses or web search. They add nulls, rival mechanisms, boundary cases, and proxy critiques.
A fixed policy ranks questions by likely importance, significance, leverage, answerability, novelty, and diversity. Deferred and specialized questions remain visible.
Combine discourse-grounded and blind alternatives for each active question. Preserve provenance and disagreement; do not truth-rank or select a winner.
Only after the question–hypothesis sets are stable does the system create atomic, contestable research targets and candidate inference routes.
Each load-bearing claim receives separate support and challenge searches. Grounds preserve source, scope, extraction, limitations, and origin path.
A verifier and falsifier independently build the strongest cases. A comparator checks scope, warrants, undercutters, and source dependence.
The completed claim graph, arguments, claim-backed defeaters, grounds, assessments, provenance, and terminal claims become one durable object.
Code rejects malformed topology, clusters dependent sources, evaluates defeasible routes, and tests strict formal candidates with Z3 and a truth table.
The writer compresses terminal verdicts into prose without adding or strengthening claims. A separate model audits every material sentence.
Claims become cards; terminal claims become answers; arguments become support lines; claim-backed defeaters become challenge lines.
Signed valence from −1 to +1. Positive displays as supported; negative as unsupported. Zero is exact equipoise.
A separate five-part evidence-depth measure. It does not determine direction.
Coverage records directions searched, queries, blind spots, inaccessible material, saturation, and stopping reason.
Sources sharing an observation, dataset, press release, or harness are clustered before aggregation.
Current Kimi artifact: migrated from the earlier inquiry pipeline; 16 claims, 16 arguments, 7 claim-backed defeaters, 12/12 artifact checks passed, and zero empirical arguments misrepresented as formal proofs.