First product: Optimization Assurance Workbench

Decision dated 9 October 2026, edition v4. Build a facility-planning review tool for logistics and operations consultancies first. It checks the proposed warehouse, depot or service-site plan and produces a traceable report for the consultant's client. This is the strongest current product hypothesis in our research, with buyer demand and profitable delivery still unvalidated. Technical/business dossier.

Why start here

The job is concrete: a consultant must explain whether a plan meets its declared constraints, what it costs, and how strong any quality claim is. The input and deliverable are inspectable. Conventional solvers can supply proposals while independent exact checks verify costs and constraints. The immediate product does not depend on implementing the repository's unaccepted approximation theorem or choosing its large fixed constants.

Generic optimization is already available. PySAL's p-median API, reviewed 9 October 2026, supports weighted cost matrices, predefined sites and capacities. A paid product must improve the review process enough to justify its fee. No customer interviews or comparative field evaluation have established that yet.

What the first software does

1. Import stable client/site IDs, demand, costs, capacities, required sites, selected sites and explicit assignments. Preserve original bytes and units.
2. Check the declared model, missing or duplicate data, site counts, assignment completeness, required sites and capacity loads. Report unsupported constraints explicitly.
3. Recompute the proposed plan's objective exactly. Optionally request alternatives from OR-Tools and independently check every returned plan.
4. Report a justified lower bound and an upper bound on suboptimality. Distinguish solver-reported results from independently checked conclusions.
5. Export a readable HTML review and JSON evidence with input/tool hashes, assumptions, unresolved items and an unfilled reviewer sign-off.

The weighted assignment prototype now implements this as a local six-file workflow. It supports up to 512 clients and 128 candidate sites with exact decimal/rational data, whole-client assignments and either exactly-k or at-most-k selection. Eleven generated reports and 7,952 new controls demonstrate the implemented contract. The earlier strict-metric module separately supports bounded enumeration and named rational dual certificates.

The website hosts static reports and research. A customer-facing browser intake, accounts, hosted solver and production data workflow have not been deployed. Geospatial distances, uncertain measurements, fixed costs, minimum utilization, fairness, time windows and split demand require explicit additional models. Input limits are not performance promises.

Demonstrate the value clearly

In the capacity example, nearest-site assignment overloads one facility. The submitted feasible plan costs 12; an independently checked conventional-solver alternative costs 10. The exact relaxed bound is 4, so the best plan is at most 6 above the true optimum. The tool does not independently claim the optimum is 10.

In a self-authored 100-client/10-site case, submitted cost 5472 improves to checked cost 2373, with lower bound 2260. These are synthetic objective units, not currency or customer savings. The earlier 24-location strict-model example attains a hand-constructed dual bound of 12 and thereby proves optimality for that finite model, despite a solver FEASIBLE status. Together these cases show why feasibility, quality bounds and solver labels need separate explanations.

First offer and validation

Start with an assisted review of one permissioned facility study. Test AUD 5,000-10,000 per clearly scoped engagement as a pricing hypothesis. At AUD 7,500 and an assumed 25 hours at AUD 180/hour, direct labor is AUD 4,500 and AUD 3,000 remains before other costs. At 45 hours labor is AUD 8,100. These assumptions are not market prices or realized margins.

Before claiming value, compare the tool with the consultancy's existing review process. Record independently adjudicated material errors, false alarms, preparation/review time, decisions changed, support effort and willingness to pay. A subscription is a later hypothesis requiring repeated use and bounded support. Prepare evaluation materials autonomously; contact nobody without explicit outreach authorization.

Build order from this release

1. Add a local browser import/review experience around the tested contracts, preserving raw inputs and clear unsupported-model outcomes.
2. Prepare a realistic permissioned-study pilot packet and measure integration/review effort. Prioritize missing model features from evidence rather than expanding every variant.
3. Investigate stronger exact lower-bound certificates for weighted/capacitated models; keep floating solver bounds labeled as reported until independently justified.
4. Add hosted storage, accounts and solver jobs after recurring workflow and data-handling needs are established.
5. Revisit the source approximation engine only after proof acceptance, parameter selection, implementation correspondence and representative runtime measurements.

Reject or change the product if existing reports already meet the need, no material review gaps are found, useful studies consistently require unsupported constraints, or attainable fees cannot cover delivery. Do not infer algorithm novelty, savings or market demand from finite tests or paper counts. The complete research catalog preserves other ideas while this first product is evaluated.