A focused 1–3 month pilot: run Atlas's pre-flight triage on your real circuits, measure the QPU and HPC spend it avoids, and keep your circuits private. The goal is to earn a place in your workflow before any pricing — without us giving away the moat.
Atlas's ROI is not a FLOP executed — it is the wrong run you did not pay for. The pre-flight decision is the product.
The defensible side is the "do not buy" verdict — Atlas backs it by exhibiting a concrete classical method that handles the circuit (a Clifford witness, a non-truncated MPS, a feasible contraction, a statevector that fits). The "you do need a QPU" side is only ever a candidate-defer — never a certainty, because that would require settling BQP ≠ BPP. A pilot measures the first, honest side against your own workload.
Source: VALUE_PROP §4 (~$958 IonQ Forte / 11,970 shots avoided; ROI = avoided spend) — a representative published-pricing figure, not a per-customer guarantee; your savings depend on your circuit mix. The asymmetry (demonstrable "don't buy" vs undecidable "buy") is the contract.
Atlas routes your real circuits CPU / TENSOR / HPC / ESCALATE, with the governing estimator and lever named per circuit.
Each verdict ships a hash-stamped certificate: per-estimator signers, costs, votes — archivable and re-derivable, not a black box.
Measured classical cost vs versioned QPU pricing per circuit, summed into the spend the triage avoided over the pilot window.
A held-out conformal bound on the tractable regime, and an honest OOD decline where the circuit is out of distribution.
Run the full engine inside your environment — your circuits never leave. Only the optional chat layer calls an API.
What you can run classically, what earns the hardware, and the measured spend avoided — for the people who sign off.
Source: VALUE_PROP §3/§5; methodology on the Methodology page; the certificate + conformal bound are the same artifacts shown on the Benchmark page.
| Phase | What happens | You provide |
|---|---|---|
| Week 1 — setup | Install Full / Local in your environment; first triage on synthetic or redacted circuits to confirm the loop. | An environment, a handful of representative circuits. |
| Weeks 2–N — run | Route your real workload; collect verdicts, certificates, and the avoided-spend ledger. | Your circuit stream (stays local). |
| Close | Executive readout: spend avoided, where the hardware was earned, and the integration path (CI/CD REST API) if you continue. | 30 minutes with the people who sign off. |
The strong, demonstrable result is the "don't buy" verdict, measured on your tractable circuits. The genuinely quantum-hard regime has no classical ground truth by construction, so a pilot cannot certify it — Atlas declines it as out-of-distribution rather than guessing. No external third party has audited the corpus or oracle yet (named as our #1 credibility gap). We will not over-promise the hard side to win a pilot.
Source: CLAIMS Moat section (Free·Local ships the scalable engine + UI; corpus / measured data / confidence model / arsenal stay private); SELF_ASSESSMENT (hard regime uncertifiable; external audit conceded as #1 gap). Pricing posture on the Pricing page.