Build Standards

Client work you submit as sprint output is read against these standards. They are also what separates work a client pays for from work a client tolerates.

The Four Dimensions

Every deliverable is scored on correctness, usefulness, explainability, and craft.

Correctness

• Code does what it claims. If it is shipped, it has been run against something real
• Every number in a client-facing document traces to a source you can point to
• Totals reconcile. If your revenue baseline is 2% off the client’s actuals, you find out why before you present it, not after they do
• Assumptions are written down as assumptions, with the value used and the reason for it
• If you do not know something, the deliverable says so. Fabricated precision is worse than an honest range

Usefulness

• Someone at the client would actually open it. A tool nobody uses is a failed deliverable
• It solves the problem the client actually has, integrated with the systems they actually run
• Discovery output names specific things to build, with an estimate and a sequence, not themes
• Scope creep is not usefulness. If the client’s question changed, renegotiate it in writing

Explainability

This is the standard the Foundry is built around, and it is where AI-assisted work most often fails.

• For any model output, you can state which inputs drove it and by how much
• Rankings come with sensitivity: what would have to change for the order to change
• A model no one on the team can explain is not shippable, regardless of how good the output looks
• “The model said so” is never an answer to a client question

Craft

• The repository is clean, has a README, and someone outside the team could run it
• Tests exist where breakage would be expensive. Not everywhere, but not nowhere
• Deployment is documented and repeatable, not a thing one person knows how to do
• Documents are proofread. Charts are labeled. Slides have a point per slide
• No client data in personal accounts, personal repos, or unapproved tools

Repository Standards

Every engagement has one private repository in the ai-foundry-byu GitHub organization, named client-<slug>.

client-<slug>/
  README.md          What this is, how to run it, how to deploy it, who to ask
  discovery/         The audit: findings, what to build, estimates, sequence
  app/               The application. Whatever stack the job calls for
  integrations/      Connections into the client's systems
  analysis/          Notebooks and scripts, in order, that reproduce every number
  deliverables/      What the client actually receives, versioned by delivery
  docs/              Handoff docs, meeting notes, decision log, scope changes

Two rules that are not negotiable:

1. No secrets or client data in the repository. Ever. Credentials live in the client’s own systems or an approved secret store.
2. Every number in a deliverable is reproducible from something in analysis/. If you cannot rerun it, you cannot defend it.

The Decision Log

Keep docs/decisions.md current. One entry per material decision:

2026-10-09: Price erosion assumption
Decision: Use 12% annual erosion for specialty, 18% for hospital, rather than a
flat 15% across the portfolio.
Why: Client's 2023-2025 actuals show materially different curves by channel.
Who: Confirmed with Brian Guy on the Oct 8 call.
Reversible: Yes. Sensitivity range 8-22% is built into the model.

This is the single highest-leverage habit in the course. It is what lets you answer “why did you assume that” six months later, and it is what makes handoff possible.

QC Gate

Nothing goes to the client without the QC lead’s sign-off. The QC lead checks:

• It runs. Someone other than the author has run it from a clean checkout
• Every number traces to a source, and totals reconcile against the client’s own figures
• Assumptions are stated, with values and reasons
• Model outputs are explainable in plain language
• Deployment steps are written down and have been followed once by someone else
• No client data or secrets in the repository
• Document is proofread, charts are labeled, filenames are clean

Work that fails QC after it reaches the client is a miss for the whole team, and the client notices it. Catch it internally.

Working with AI

Use AI tools hard. They are why a team of five can deliver what used to take fifteen.

• Generate freely, verify always. Treat every model output as a draft from a fast, capable, occasionally confident-and-wrong analyst.
• Check citations and figures by hand. Fabricated sources have ended real consulting engagements.
• Keep client data out of unapproved tools. See the confidentiality terms in the Syllabus.
• Document your method. Project documentation and your sprint reviews should state which tools you used, for what, and how you validated the output. This is not a confession, it is methodology.