The Audit Methodology.
Public. Versioned. Built for advisor sign-off. Hashed onto every page of every report the audit CLI generates.
If our audit cannot withstand scrutiny by your engineering team, your CTO, and your board's technical advisors, it should not exist. This page is how we prove it.
Public
Anyone can read this. Anyone can challenge it. We respond.
Versioned
Every revision is hashed. Reports cite the version that produced them.
Sign-off
Built for sign-off by a credentialed HPC operations advisor. No advisor is seated yet: first design-partner reports ship unsigned, disclosed in writing (ADR-0018). Status below.
15 archetypes. 7 families. One classifier.
Every workload in your audit is evaluated against this library. Each archetype cites the published result behind it and lists exactly which metadata signals the classifier looks for.
Optimization
3 archetypesQuadratic and combinatorial cost functions classical solvers struggle to scale at enterprise instance sizes.
Combinatorial QAOA
Definition: Max-Cut / MIS / QUBO-expressible scheduling. Basis: Farhi, Goldstone, Gutmann 2014, A Quantum Approximate Optimization Algorithm, arXiv:1411.4028. Signals from your metadata file: structural markers bqm, ising_model, max_cut, qubo_model; tool priors cplex, glpk, gurobi, or-tools, ortools, scip, xpress; size input n_binary_vars in 20–5000. Horizon 2028 · confidence medium · mapping tier published.
Portfolio QAOA
Definition: Allocation with budget + risk (QUBO/Ising). Basis: Hodson, Ruck, Ong, Garvin, Dulman 2019, Portfolio rebalancing experiments using the quantum alternating operator ansatz, arXiv:1911.05296; building on Farhi 2014. Signals from your metadata file: structural markers portfolio_qubo, qubo_model; tool priors cplex, gurobi, mosek; size input n_assets in 20–2000. Horizon 2028 · confidence high · mapping tier published.
MILP-mapped routing
Definition: Vehicle routing / network flow reduced to QUBO. Basis: Lucas 2014, Ising formulations of many NP problems, Front. Phys.. Signals from your metadata file: structural markers tsp_model, vrp_model; tool priors concorde, cplex, gurobi, lkh, or-tools, ortools; size input n_nodes in 10–1000. Horizon 2030+ · confidence medium · mapping tier nontrivial.
Simulation
3 archetypesQuantum systems are the cleanest target for quantum hardware. Hamiltonian-shaped workloads admit known polynomial advantages.
VQE molecular Hamiltonian
Definition: Small-molecule ground-state energy. Basis: Peruzzo et al. 2014, A variational eigenvalue solver on a photonic quantum processor, Nat. Commun. 5:4213. Signals from your metadata file: structural markers active_space_spec, fcidump, second_quantized_hamiltonian; tool priors g09, g16, gaussian, molpro, nwchem, orca, psi4, pyscf, qchem; size input n_spin_orbitals in 4–200. Horizon 2026 · confidence high · mapping tier published.
Quantum chemistry (beyond DFT)
Definition: Electronic structure / correlation energy. Basis: McArdle et al. 2020, Quantum computational chemistry, Rev. Mod. Phys. 92:015003. Signals from your metadata file: structural markers casscf, multireference, post_hf; tool priors cp2k, molpro, nwchem, orca, pwscf, qe, quantum-espresso, vasp; size input n_spin_orbitals in 50–2000. Horizon 2028 · confidence high · mapping tier nontrivial.
QPE eigenvalue
Definition: High-precision spectra / phase estimation. Basis: Kitaev 1995, arXiv:quant-ph/9511026; Nielsen & Chuang §5.2. Signals from your metadata file: structural markers eigen_problem, hermitian_sparse; tool priors arpack, slepc; size input matrix_dim_log2 in 10–60. Horizon 2030+ · confidence medium · mapping tier nontrivial.
Search
2 archetypesGrover-class problems admit known sqrt(N) speedup. Honest classification means flagging the well-known practical caveats.
Grover unstructured
Definition: Pre-image / unstructured search. Basis: Grover 1996, A fast quantum mechanical algorithm for database search, STOC. Signals from your metadata file: structural markers oracle_predicate, unstructured_search; tool priors glucose, minisat; size input log2_search_space in 20–80. Horizon 2030+ · confidence low · mapping tier nontrivial.
Grover-mapped CSP
Definition: Constraint satisfaction via amplitude amplification. Basis: Brassard, Høyer, Mosca, Tapp 2002, Quantum amplitude amplification and estimation, AMS Contemp. Math. 305. Signals from your metadata file: structural markers cnf_oracle, csp_oracle; tool priors gecode, glucose, minisat, or-tools; size input log2_search_space in 20–80. Horizon 2030+ · confidence low · mapping tier nontrivial.
Linear Algebra
2 archetypesHHL and amplitude estimation admit polynomial speedups under tight preconditions. Reporting both with their caveats.
HHL
Definition: Sparse, well-conditioned linear systems. Basis: Harrow, Hassidim, Lloyd 2009, Quantum algorithm for linear systems of equations, PRL 103:150502 — *caveats: Aaronson 2015, Read the fine print*, Nat. Phys. 11:291. Signals from your metadata file: structural markers sparse_linear_system, well_conditioned; tool priors hypre, mumps, petsc, superlu, trilinos; size input matrix_dim_log2 in 10–50. Horizon 2030+ · confidence low · mapping tier research.
Amplitude Estimation
Definition: Monte Carlo / derivative & risk pricing. Basis: Montanaro 2015, Quantum speedup of Monte Carlo methods, Proc. R. Soc. A 471:20150301. Signals from your metadata file: structural markers monte_carlo_integrand; tool priors quantlib; size input log2_paths in 10–40. Horizon 2028 · confidence medium · mapping tier nontrivial.
Number Theory
1 archetypesCryptographic relevance. We report Shor strictly as speculative until hardware credibly threatens to factor enterprise-relevant integer sizes.
Shor's algorithm
Definition: Integer factoring / discrete log. Basis: Shor 1997, Polynomial-time algorithms..., SIAM J. Comput. 26:1484. Signals from your metadata file: structural markers factoring_modulus; tool priors cado-nfs, msieve, yafu; size input modulus_bits in 256–8192. Horizon speculative · confidence speculative · mapping tier published.
ML / Hybrid
2 archetypesMost aggressive marketing claims live here. We deliberately downweight pending replicable evidence.
Quantum kernel methods
Definition: SVM/classification on quantum feature maps. Basis: Havlíček et al. 2019, Supervised learning with quantum-enhanced feature spaces, Nature 567:209. Signals from your metadata file: structural markers quantum_feature_map; tool priors libsvm, scikit-learn, sklearn; size input n_features in 20–1000. Horizon 2028 · confidence low · mapping tier nontrivial.
Variational gradient
Definition: Quantum-assisted training of variational models. Basis: Cerezo et al. 2021, Variational quantum algorithms, Nat. Rev. Phys. 3:625 — *caveat:* barren plateaus. Signals from your metadata file: structural markers variational_ansatz; tool priors pytorch, tensorflow, torch; size input n_params in 10–500. Horizon speculative · confidence speculative · mapping tier research.
Graph / Stochastic
2 archetypesNewer entries to the literature. Reported with appropriate caveats.
Quantum walks
Definition: Graph traversal / centrality / element distinctness. Basis: Ambainis 2007, Quantum walk algorithm for element distinctness, SIAM J. Comput. 37:210. Signals from your metadata file: structural markers element_distinctness, graph_traversal; tool priors graphblas, gunrock, networkx; size input log2_vertices in 12–50. Horizon 2030+ · confidence low · mapping tier nontrivial.
Stochastic optimization
Definition: Optimization under uncertainty (sampling-hard). Basis: Montanaro 2015 (as above) applied to stochastic objectives. Signals from your metadata file: structural markers stochastic_objective; tool priors smps, stochastic-programming; size input n_scenarios_log2 in 10–40. Horizon 2028 · confidence medium · mapping tier nontrivial.
How a workload becomes a classification.
Rule-based. Auditable. No ML black box.
Ingest scheduler logs
The CLI parses Slurm sacct output, Kubernetes pod JSON (including Argo workflows), or PBS Pro accounting records for the window you export. Raw logs never leave the head node.
Build workload signatures
Jobs are grouped by name into workloads: run count, mean and p95 runtime, cores, GPUs, queue wait, and window length. If you supply a metadata file, it is joined by job name (command, modules, linked libraries, problem dimensions, structural markers); the name is then replaced by its hash.
Match against the archetype library
For every archetype, the classifier checks three signals (STRUCT, TOOL, DIM), computes s, and assigns strong / partial / weak / none. The best match wins; ties go to the nearer horizon, then higher confidence, then archetype id.
No match is a result
A workload with no signals, for example one with no metadata entry, is classified no_candidate and still scored, so it lands in the "Not a candidate" band. Confidence (high / medium / low / speculative) belongs to the archetype, as listed above, not to the individual match.
No quota, no upcharge
When a workload portfolio contains zero quantum-suitable jobs, we report that and the audit price does not change. The Latency Index value comes from the full inventory, including the no-match class.
The Quantum Suitability Score, formally.
Five inputs. One score per workload. Reproducible from your scheduler logs, your compute rates, and your metadata file.
Archetype Match
Three signals per archetype: STRUCT (0.50) a registered structural marker, TOOL (0.30) a tool prior among the job command, modules, or linked libraries, DIM (0.20) the size input inside the archetype range. Their sum is s. Strong requires STRUCT and s ≥ 0.80; partial s ≥ 0.45; weak s ≥ 0.20. Markers are declarations in the metadata file you supply; the CLI does not inspect job inputs.
Scale: strong 30·s · partial 18·s · weak 7·s · none 0. A tool prior alone caps at 2 points.
Problem Size
Estimated logical qubits q̂ for the matched archetype (e.g. one per spin-orbital for VQE), compared with Q-Intercept’s own conservative planning budget for the archetype’s horizon: 12 (2026), 100 (2028), 1000 (2030+). Pending advisor review.
Scale: Below 4 qubits: 4. Within budget: 16–20. Up to 2× budget: 10–15. Beyond: 3. Speculative horizon: 3. Unknown size or no match: 4.
Classical Cost
Annualized cost C = runs per year × mean runtime × (core rate × cores + GPU rate × GPUs), using only the rates you supply. Queue-wait cost is computed separately and never folded in. If a needed rate is missing, the score is reported as not available.
Scale: Log-linear: $1k/yr or less → 1, $10k → 9, $100k → 17, $1M or more → 25. Anchors are provisional until calibrated on the first five paid audits.
Hardware Horizon
Base from the archetype’s horizon band, adjusted +1/−1 for fault-tolerance dependence (NISQ-shallow vs. deep fault-tolerant circuits) and +1/−1 for whether a published end-to-end hardware demonstration exists. Algorithm-class facts only; never vendor roadmaps or vendor performance.
Scale: Base 2026 → 13 · 2028 → 9 · 2030+ → 5 · speculative → 1, then −2 / 0 / +2, clamped to 0–15. No match: 0.
Translation Difficulty
Ease of mapping the problem to its quantum formulation, set once per archetype: a published, routine mapping; a mapping that needs non-trivial modeling; or a research-level mapping. It does not analyze your code.
Scale: published 9 · nontrivial 5 · research 2. No match: 0.
Banding
Why we report horizons in bands, not quarters.
Horizon bands are coarse on purpose: 2026 / 2028 / 2030+ / speculative. They describe algorithm classes, not vendor delivery dates.
What we score against
- The archetype's horizon band, set in the methodology with its advantage citation
- Fault-tolerance dependence: shallow NISQ-era circuits vs. deep fault-tolerant circuits
- Whether a published end-to-end hardware demonstration exists, with citation
- Logical-qubit planning budgets per horizon (our own, conservative, pending review)
What we will not score against
- Vendor marketing claims of imminent advantage
- Unpublished error-correction claims
- Vendor roadmaps, device generations, or vendor performance comparisons
- Pre-print results that have not been replicated
The honesty guardrail, enumerated.
These rules govern the methodology and every report. Each report prints the claims it does not make.
No specific speedup multipliers on today's NISQ-era hardware
Reports may cite published asymptotic complexity bounds. They will not extrapolate to wall-clock speedups on hardware that does not yet exist.
No promised quantum-advantage delivery dates beyond banded horizons
We commit to bands (2026 / 2028 / 2030+ / speculative). We do not commit to quarters, vendors, or device generations within those bands.
No vendor performance comparisons
We classify workloads against published archetypes. We do not predict relative performance of IBM Heron vs. IonQ Forte vs. neutral-atom platforms.
No ROI projections beyond classical bottleneck cost
Cost enters the score only as the classical cost you already pay, from your own rates. Reports never project savings that depend on quantum hardware delivering claims we cannot verify. (Since CLI v0.0.4 the report prints the annualized classical cost per workload and a top-10 by cost: your own spend, never a projection.)
No forced classifications
When no archetype credibly fits, the workload is labeled "no quantum candidate" and reported. No quota for quantum-applicable findings.
Printed in every report
Each report ends with the list of claims it does not make, so the refusal travels with the document. These rules are applied by how the report is written; there is no automated phrase linter.
What leaves your network. What does not.
Reports stay local. The CLI uploads nothing. The only artifact designed to leave is the sanitized JSON manifest, and only if you write it with the consent token and move it yourself.
| Field | Treatment | Leaves network |
|---|---|---|
| Workload ids | HMAC-SHA256 of the job name under a key you hold (--manifest-key), 16 hex characters; bare SHA-256 with a CLI warning if no key is given | Sanitized |
| Archetype, band, cost band per workload | Categorical labels only (cost band: low / mid / high / unknown) | Sanitized |
| Aggregate counters | Workload count; counts per archetype, band, and cost band | Sanitized |
| Scheduler type | One of {slurm, k8s, pbs} | Sanitized |
| Methodology version + hash | Pins the scoring constants | Sanitized |
| Raw job names, container images, scripts, file paths | Not in the manifest schema | Never |
| User IDs, group IDs, IP addresses, hostnames | Not in the manifest schema | Never |
| Dollar values, rates, runtimes, sub-scores, problem dimensions | Not in the manifest schema | Never |
--manifest-preview. The CLI refuses to write it to disk unless you add --consent EXPORT-SANITIZED-MANIFEST, and it never transmits it. Any transfer is a manual step you perform.The technical advisor.
Methodology, every revision, every audit report: built for advisor sign-off.
Methodology 1.0.0-draft
To be signed by [Technical Advisor — recruitment underway], a credentialed HPC operations veteran with peer-reviewed publications in the operator's vertical. No advisor is engaged yet; a false internal record stating otherwise was corrected in public on 2026-08-10 (roadmap register).
Once engaged, the advisor's name is disclosed to enterprise customers during scoping, under mutual NDA. The standard engagement requires a credentialed advisor for the customer's industry. One disclosed exception (ADR-0018): a first design-partner cohort proceeds without an advisor signature, stated in writing before contracting, with reports re-signed at no charge once the methodology is published.
Advisor role, once seated
- Will review the methodology before publication and on every revision
- Will review and sign standard-engagement reports; first-cohort reports are re-signed on publication
- Will be available on standard-engagement debrief calls as a credentialed reference
- Will be able to refuse to sign; a refusal is reported to the customer with rationale
Hashed onto every page of every report.
Methodology revisions are tracked. Every report pins to the version that produced it.
Each customer audit report carries the version + hash of the methodology that produced it on every PDF page footer.