A method for making nonconsensus claims auditable.

AI has made broad explanation cheap. The scarce work is no longer summarizing a market. The scarce work is saying what the market believes, where that belief may be wrong, what evidence supports the departure, and what would change the conclusion.

Most generated research fails by sounding complete too early. It covers the surface area, fills every section, and hides the unknowns. The Δ Protocol reverses that incentive. It rewards sparse but decision-relevant claims, confidence calibration, and visible unresolved questions.

The method borrows from three older disciplines. The Rachleff 2x2 names the target cell: non-consensus and right. Variant perception from public markets asks what one sees that the market has mispriced. Falsifiability, in the Popperian sense, requires a claim to expose itself to being wrong.

Nonconsensus adds product mechanics around those ideas: evidence snapshots, citation validation, challenge affordances, forked counter-theses, and versioned public memos.

Evidence is tiered before it is polished. Tier A and B sources can support high confidence. Tier D and E sources usually cap confidence at medium unless independently confirmed. Conflicting, stale, or unverified material is low confidence and should create an unknown rather than a stronger sentence.

The reference implementation validates citations, deduplicates repeated claims, recalibrates suspicious confidence distributions, and rejects charts when the evidence fails the gate. A missing chart is acceptable. A wrong chart is a brand failure.

Deep-tech markets often fail through timing, certification, distribution, and trust before they fail through model capability. The protocol forces those constraints into the memo instead of letting the research drift toward capability theater.

The Augustus worked example, Physical AI & Autonomous Industry, treats the fundable layer in physical AI as a claim to test. The memo separates robot narratives from deployment control points: permits, industrial certification, fleet reliability, insurance, simulation compute, and operator economics.

That example is useful because it is not merely positive or negative on robotics. It names what would flip the conclusion: audited humanoid reliability, standardized fleet failure data, federal regulatory preemption, and real RaaS margin disclosure.

The protocol does not remove judgment. It only makes judgment inspectable. Source availability can bias the corpus toward companies that publish more. Private edge can be difficult to express. Some markets move faster than the resolution loop. Some important truths are qualitative before they are measurable.

The method also cannot guarantee that a divergent claim is valuable. It can only force the claim to carry evidence, stakes, a clock, and a failure condition. That is enough to make debate productive, but not enough to make the thesis true.