Research

The evidence discipline behind every claim.

Counterfactual systems make it easy to generate scenarios and hard to know which ones deserve trust. Our research program keeps that line sharp: every public claim is scoped, tiered, and — where the outcome arrives — scored.

Tiers

Four comparative-evidence tiers, declared everywhere

Because the unchosen action is usually not observed, every Decision Fidelity result declares one comparative-evidence tier. Tiers B–D are never presented as direct causal validation.

  1. Acontrolled: randomized, exhaustive, or directly observed alternatives.
  2. Bquasi-experimental: matched, natural-experiment, or off-policy estimates with stated assumptions.
  3. Cproxy: simulator, retrospective analogue, or blinded expert comparison.
  4. Dunresolved: no valid comparative resolution yet.

Provenance

Where a branch came from is a separate question

Expert-authored, retrieved, rule-generated, simulator-generated, learned, synthetic, permissioned, and observed are provenance labels, not evidence tiers. Both source provenance and the A–D comparative-evidence tier remain visible so a realistic-looking branch cannot borrow causal credibility it has not earned.

Scope

Robust means robust within the declared family

A robustness statement here is scoped four ways: to the branch family we declared, the assumptions we shear-tested, the constraints we enforced, and the evidence tier underneath. A Shear Test that was never run is a risk we have not measured, and we say so. Unscoped robustness is marketing, and we do not publish it.

Resolution

Decision Fidelity closes the loop

Decision Fidelity is a declared evaluation protocol and metric vector, not a decorative confidence score. It can report action-ranking agreement, decision regret, coverage, calibration, and stability. Observing the chosen action resolves its consequences; it usually does not reveal what unchosen actions would have done. Every result therefore carries an A–D evidence tier and remains Tier D until a valid comparative evaluation exists. The outcome-resolution system is in development.

Compute research

Pilot Wave: quantum methods, measured honestly

Pilot Wave is our provider-neutral research program benchmarking annealing, hybrid, and gate-model methods against strong classical baselines on selected discrete workloads — measured end to end, including wall time, overheads, and cost. It makes no quantum-advantage claim.

Inside Pilot Wave

The same discipline, end to end

Pilot Wave applies the evidence discipline to the loudest hardware question in the field.