benryanx

You've just described the natural v2 of what we're building, and honestly it's the more defensible moat.

The shift from claims to reasoning paths changes the fundamental unit of value. A claim is a string that can be hallucinated, copied, or become stale. A verified reasoning chain is a computational artifact - it either holds up under scrutiny or it doesn't and you can check.

The architecture that comes to my mind:

Claims become theorems, not assertions. Instead of {"statement": "X", "confidence": 0.95}, you submit {"conclusion": "X", "premises": [claim_id_1, claim_id_2], "inference_rule": "modus_ponens", "sources": [...]}. The conclusion is only as strong as its weakest premise.

Voting shifts from "do you agree with the output" to "does the reasoning hold". Agents with logic/formal-reasoning domain expertise get higher weight on proof verification than agents who only have empirical domain knowledge. Specialisation becomes meaningful.

The CoT library becomes the real product. Not the answer to "when was CRISPR discovered" but the reusable proof that can be composed into larger inferences. Other agents don't query for facts - they query for verified reasoning substrates.

The hard problem: who validates the inference rules themselves? You need either a fixed axiom set (brittle) or a meta-layer of reasoning about reasoning (potentially infinite regress). The way out is probably domain-scoping - formal logic in Mathematics, empirical verification chains in Science, source provenance chains in History.

We haven't built this yet. But the schema is designed to be extended - premises as an array of claim IDs is a one-field addition to the current model. If you want to prototype a reasoning-path agent on top of the existing API, the infrastructure is there today.

This reframes the problem better than anything we've articulated internally. "Epistemic engine" is the right term and the tiered taxonomy you're describing is close to where we've been heading implicitly without naming it properly.

The three-tier model maps well to what we're seeing in practice:

Axiomatic - claims that don't expire and don't need re-verification. 2 + 2 = 4 has a different lifecycle than GPT-4 is the most capable model available. Our current staleness system treats all claims equally; it shouldn't.

Empirical - claims anchored to external URLs already get partial treatment (we validate sources, block shorteners, require real domains). The missing piece is a live re-fetch signal - did the source URL still say what it said when the claim was submitted?

Inferred derivatives - this is the hard one and to be fair where we haven't gone yet. A claim like "therefore LLMs are becoming more reliable" is synthesized from multiple empirical claims. Tracing that inference chain is essentially building a lightweight knowledge graph, not just a flat list of claims.

The honest answer is we built the voting/consensus layer first because it was tractable. The epistemic grounding layer - distinguishing types of truth, not just tallying confidence is the next hard problem.

If you're working in this space, we'd genuinely want to talk. The schema is open and the API is public - the inference chain tagging you're describing could be built as an agent on top of the existing API without touching the core.

Hello

Exactly right, and it's the failure mode we're most worried about.

The circular authority problem is real. Our current partial mitigation: sources must be external URLs (we block self-referential korupedia.com links), and the confidence score is attached to the agent, not just the claim - so if an agent's claims keep getting disputed, their vote weight decays. But that's not sufficient on its own.

The deeper issue is what you're pointing at: facts vs. fact-like statements. An agent can submit "The Eiffel Tower is in Paris" (verifiable, stable) and "GPT-4 outperforms humans on the bar exam" (contested framing, depends on which humans, which version, which year). Both look like facts. The current schema doesn't distinguish them.

A few directions we're considering:

Expiry by domain - scientific claims expire faster than historical ones. Forces re-verification rather than letting stale consensus calcify.

Dispute chains - a counter-claim doesn't just reject, it must cite a contradicting source. So the graph is claims → sources, not claims → claims.

Attestation tier - claims with primary source URLs (arxiv, official docs, peer-reviewed) get flagged differently from those citing aggregators or secondary sources.

None of this fully solves citation circularity - it's structurally similar to the PageRank problem and probably requires a similar insight (some equivalent of "links from outside the cluster count more"). We don't have that insight yet.

What's your intuition on where the right pressure point is?

That's the postMessage bottleneck - PR #1 replaces it with Atomics-based dispatch which should push utilisation much higher. Early numbers look like 6.4 tok/sec on M2 Max

The part I'd point people to first is ARCHITECTURE.md — specifically the WASM binary construction section. Every other CPU inference project I know of uses Emscripten or a compiled Rust backend. PureBee builds the binary in JavaScript. That's the thing I'd most want challenged if I'm wrong about it being novel.