# Dispositive Questions ## Fourteen questions that move the decision, with the green and red answers Each instruction gives one diligence question. Each instruction has an answer pattern that raises or lowers conviction. ## Very important 1. **Bill of materials and licenses for the software.** Get the bill of materials from the company. Find if the company uses open-source low-code tooling. Find if the company rebrands the frontend of the tooling. Find if a tier runs multi-tenant on the tooling. Find which code is proprietary. - *Green*: The company gives a full bill of materials. The company uses the tooling in the backend only, or with a commercial license. The company shows a bounded proprietary codebase. - *Red*: The company refuses, or says "all in-house" and does not show the repository. The company rebranded the console and has no license. 2. **Ownership of data and weights for each vertical.** Find who owns the training data, the knowledge bases, and the resulting model weights. Find if the company can use them again. - *Green*: The company has permanent rights in writing to use them again. Rights that are exclusive for a period are better. The contract has background-IP carve-outs. - *Red*: The institutes own everything. The institutes can license it to others. 3. **Source of the outcome metrics.** Find which customer gave the outcome metrics. Get the baseline and the period. Find if that customer agrees to speak with the company. - *Green*: The company names the customer. The company has a documented baseline and a production period. The customer confirms the outcomes. - *Red*: The source is simulations, targets, or composites. 4. **Fine-tuned model against an off-the-shelf model.** Find which model the company fine-tuned. Get the method of the fine-tuning. Get the performance against an open-weight model with the same retrieval. - *Green*: A held-out test shows a clear margin. The license permits resale. - *Red*: The company has no test. Only the retrieval causes the gains. ## Important 5. **Last three deployments.** Get the time to go-live of each of the last three deployments. Get the engineer-days of each deployment. Find how much of the work was new code and how much was configuration. - *Green*: A deployment takes about two months or less. The work is mostly configuration. The work for each deployment decreases. - *Red*: Each deployment has a bespoke repository. 6. **Deployment map.** Get a map of each live deployment. Find if the deployment is on-premise or in the cloud. Find which GPUs the deployment has and who bought the GPUs. Find the tier of the serving license. Find the data flows to external APIs. 7. **Engineering roster.** Get a roster of each person who writes production code or trains models. Get the tenure of each person. - *Green*: The company has four or more experienced engineers. The company gives the name of the person who leads AI. - *Red*: The company has two or fewer engineers. Contractors or advisors make the core parts. 8. **Inspections of protection.** Get the questionnaires on protection that the customers approved. Get the penetration test. Get a certification plan with dates. 9. **Alpha-stage components.** Find if alpha software operates in a customer environment. - *Red*: Alpha software operates in a customer environment. ## Moderate 10. **Patent list.** Get the patent list. Get the number, the title, the jurisdiction, the inventors, and the assignee of each patent. 11. **Academic or government claims.** Find if an academic body has a claim on the models, the methods, or the code. Find the same for a government body. 12. **Test and monitoring stack.** Get the test and monitoring stack. Get the regression tests, the hallucination tests, and the dashboards that show a change in the performance of the model. 13. **Roadmap arithmetic.** Make sure that the budget for agents agrees with the stated cost of each agent. Refer to [[Roadmap Budget Realism]]. 14. **Vision pipeline.** Get the models of the vision pipeline for defect detection. Get the volume of labeled data. Find how accurate the pipeline is. Get the hardware. ## Why it matters Questions 1 to 4 show if the thesis is correct. The other questions measure the execution risk. Refer to [[What Must Be True]] for the conditions that the questions test. ## Related - [[Foreman MOC]] - [[Claims Versus Evidence]] - [[Technical DD Framework]]