The Audit is a deterministic chassis. Same inputs in, same verdict out. This page is the versioned reference: the four assumption tiers, the four verdicts, the chassis route families, the citation policy, the reproducibility protocol, and the right-of-reply protocol.
A verdict without a named tier is unfalsifiable. The Audit pins each route to a single tier and captures verdict sensitivity to each named assumption inside that tier.
Noise-free hardware. One cycle per oracle call. Lower-bound condition estimates. Favorable precision. The first cost driver is the cycle count derived from the QLSA-based Newton solve at minimal iterations. If the candidate cannot beat its strongest classical baseline at this tier, no further analysis closes the gap.
cycle count, oracle calls, precision floor, condition estimate
A named hardware family with cited gate speed. Fault tolerance excluded. Energy modeling required. Quantum-error-correction overhead modeled. Cycle-time-to-runtime conversion is the dominant cost driver here. A Tier B Manifest without modeled energy budget fails FR-01 at the runner.
gate speed, cycle time, QEC overhead, energy budget, classical post-processing
A named NISQ hardware family with cited gate speed. Fault tolerance excluded. QEC overhead not modeled. Explicit scope limit is required. A Tier B Manifest that ships without QEC modeling and without the relabeling fails FR-01.
gate speed, calibration cost, shot budget, classical post-processing, named scope limit
Logical qubits, error correction modeled, magic-state distillation modeled, energy modeling included. The dominant cost drivers are logical-qubit count, code distance, and T-gate budget. A Tier C verdict is a long-horizon engineering bet against named thresholds.
logical qubits, code distance, T-gate count, distillation cost, logical clock rate
A free-text bet by the customer about a specific assumption (a hardware threshold, a precision relaxation, an output-requirement change). The chassis derives the threshold at which the bet would change the verdict. The bet is captured verbatim in the Manifest with the test that would settle it.
customer-stated assumption, named threshold, sensitivity test
A different formulation, output requirement, or baseline configuration could survive. The chassis names the redirect path. Most live workloads land here.
A specific named threshold (logical cycle time below X, condition number below Y, code distance D) would change the answer. The threshold is captured in verdict_sensitivity per route.
No reformulation closes the gap, no named threshold opens a future window. The verdict is bounded by the workload, the baselines, and the assumption envelope.
Output extraction is not obviously fatal. A hardware path is named. The recommended next step is bounded, not a multi-quarter program. GO is provisional pending the assumption envelope.
The verbatim rule, enforced by the runner against every Manifest:
A NO-GO that did not first ask "is there a different formulation that survives" does not leave the chassis. The reformulation log is the audit trail of that question. Any Manifest with a NO-GO verdict carries at least one entry in reformulation_log.
The verbatim clause, enforced by the runner via a const constraint on every Manifest:
Verdicts are bounded by this workload, these baselines, and the stated assumption envelope. They do not generalize to all quantum candidates for this problem class.
Amplitude estimation against Monte Carlo for risk pricing, derivatives valuation, and portfolio expectations. The candidate runs at scalar output, 1e-2 to 1e-3 precision, against at least two classical Monte Carlo baselines.
stamatopoulos_2020_option_pricingchakrabarti_2021_threshold_advantagestamatopoulos_zeng_2024_qspHybrid quantum interior point methods against the strongest classical solver (HiGHS plus an optional commercial alternative). The Newton-system condition number is the dominant cost driver in the lower-bound calculation.
binkowski_2026_lp_qipmabbas_marecek_2024_natrev_optimVariational and QML candidates that the chassis treats as a no-go-screen by default. The screen runs the candidate as a REDIRECT against the strongest classical baseline before any speedup claim is recorded.
tang_2018_dequantizationliu_arunachalam_temme_2021larocca_2025_bp_reviewcerezo_2025_dequantizationqoblib_decathlon_2025Quantum simulation and electronic structure against high-quality classical tensor networks (DMRG, coupled cluster, AFQMC). Verdicts gate on logical qubit count, active-space size, and basis-set choice.
daley_2022_nature_simulationlee_babbush_2021_thc
Each citation_key dereferences to an entry in the Manifest's citations.primary_methodology[*] array. Every route in the chassis points to a citation key via applies_to_routes. The full reference list appears in section V.
The chassis citation backbone is embedded in every Manifest as citations.primary_methodology[*]. Each entry carries a citation_key, a full citation, the role it plays in the chassis, and the licence posture of the source paper or its associated code.
binkowski_2026_lp_qipm chakrabarti_2021_threshold_advantage stamatopoulos_2020_option_pricing stamatopoulos_zeng_2024_qsp tang_2018_dequantization liu_arunachalam_temme_2021 larocca_2025_bp_review cerezo_2025_dequantization abbas_marecek_2024_natrev_optim qoblib_decathlon_2025 daley_2022_nature_simulation lee_babbush_2021_thc Every Manifest carries the verbatim independence statement. The schema enforces verbatim text via a const constraint; deviation by a single character fails validation before the field rules even run.
Informed by published lower-bound analysis of hybrid quantum interior point methods. DeployQuantum did not author that research.
Each citation entry records whether the source paper or its associated code is reusable. The Manifest schema is open-sourced under CC BY 4.0. The public-benchmark runner is open-sourced under MIT.
The Audit's pipeline is reimplemented independently in DQ-owned code; no third-party code is copied into the chassis.
Strip the top-level hash field. Recursively sort keys at every depth. Use a JSON serializer with no insignificant whitespace. Preserve Unicode. Produce a single-line UTF-8 string with no trailing newline. Compute SHA-256 over the result.
The hash is what the board signs on. Any field that should commit to the audit goes inside the hashed body; audit-trail metadata that changes post-signing is held outside the manifest in sibling files.
The Manifest carries a runner_version field. Reproducing a Manifest requires the same runner version (or a newer version that maintains canonicalization invariants), the same canonical-hash policy, and the same instance-set hash.
The runner is open-source; any third party can validate a Manifest against the schema and against the twelve field-level publish blockers.
The Manifest's workload.instance_set_hash field commits to the exact instances run. Different instance sets produce different Manifests; identical instance sets reproduce the runtime distribution within the runner's stated tolerances.
A Manifest at audit_tier: "Confidential" requires a populated external_reviewer with role and reviewer confirmation captured before any public reference. FR-11 blocks any Confidential Manifest without an external reviewer.
Before a public Manifest names a vendor or a candidate-author, the chassis sends a pre-publication notice with the verdict, the reasoning, and the assumption envelope. The notice carries the canonical hash of the candidate Manifest as it stands at the time of notice.
A defined response window applies. The vendor or candidate-author may submit a written response (correction, dispute, new evidence) inside the window. The response is recorded in the Manifest's vendor_response field and is published alongside the verdict.
If the response identifies a factual error in the inputs, the assumption envelope, or the baseline configuration, the chassis re-runs the affected route. The corrected Manifest carries a new hash; the prior Manifest is preserved in the change log with the rationale for the correction.
Every public Manifest carries a versioned change log. Corrections, reformulation passes, and updates to the assumption envelope each appear as a versioned entry with the previous hash, the new hash, and the rationale. The log is part of the artefact; readers can reproduce any prior version from the canonical record.
| VERSION | SCHEMA | RUNNER | CHANGE |
|---|---|---|---|
| v1.0 | v0.2.0 | 0.2.0 | Initial published methodology. JSON Schema (Draft 2020-12), Python runner, twelve field-level publish blockers, chassis citation backbone, four assumption tiers, four verdicts, NO-GO-requires-reformulation rule, forbidden-generalization clause, reproducibility protocol, right-of-reply protocol. |