An order-of-magnitude estimate of what it costs to build and serve the claim graph, doing it right. The cost is dominated by LLM inference; a frontier-quality build lands in the high single-digit millions, concentrated in the broad middle tiers and the ingestion firehose rather than the flagship claims.
Twelve testable properties an agentic claim graph should satisfy, from idempotency to calibration, with notes on how to measure each and where adversarial inputs make it harder.
An open, LLM-administered repository of claims — built, maintained, and served by an LLM-based bureaucracy drawing on humans, research agents, and the rest of civilization's full epistemic stack.
Canonical claim wording should be clear and neutral; negations aren't separate claims. “The moon landing was a hoax” is canonical, since the affirmative is what's actually asserted.
Open subproblem: how to rank claims by importance to prioritize maintenance and guide traversal.
The foundational subproblem: reading a corpus and extracting clean, organized claims. An overview of its constituent difficulties.
Open subproblem: what counts as an edge between claims, and how edges should be typed, directed, and weighted.
Finding the claim an extracted proposition corresponds to requires vector search driven by an LLM agent iterating, not a single nearest-neighbor hit.
Minerval should resolve contentious factual questions while avoiding three failure modes: overconfident framing, deciding normative claims, and reifying unstable LLM priors.
Open subproblem: how an update to one claim should propagate to related claims without runaway cascades.
Like species in biology, claims must be split or lumped. The test: if both sides of an argument agree two propositions are the same claim, they are.
Why mechanical Bayesian reasoning over the whole graph fails, and the informal alternative: any claim can parent any other, with relationships carried by Arguments.
Agentic traversal is straightforward: the fetch-claim tool takes claim IDs and returns the IDs of all linked claims.
Public contributions reach claim stewards only through wary routing agents; abusers are flagged or banned, and good reputations might be rewarded.