# Step 3 — Implementation Detail Built on 2026-10-11. **Not run with real models.** ## What a stated choice now has to reach Case 08 showed a stated topology named in a design and not built, with every Check passed. A stated choice was an instruction; nothing held a design to it. 1. **The catalogue marks the fields that bind a design** (`binds: true`, 26 of them): regions and residency, availability and recovery targets, compliance frameworks and classification, encryption, audit logging, every networking field, identity, environments, and required and forbidden products. 2. **The agreed scope lists the binding fields the customer stated,** as `binding` in `scope.json`. A field assumed for them is not among them: an assumption is not the customer's decision. 3. **A requirement says what it comes from,** in `from`: the customer's own requirement ids and the scope fields it carries. 4. **A submission of requirements is refused** (`_carries_the_scope` in `app/workflows/chain.py`) when: - one of the customer's own requirements is named by no requirement; - a binding field is named by no requirement; - a requirement that carries a binding field is not non-functional, or has no entry in `conditions`. 5. **From there the existing chain does the rest.** A non-functional requirement's conditions each need a Check on the Target architecture, and validation gives each Check a verdict. So a stated choice now reaches a Check, in code, at every link. What this does not prove: that the condition the analyst writes says what the customer meant, or that the Check the architect writes verifies it. It guarantees a stated choice cannot be silently dropped; whether it is honoured well is still judged by the validator. ## Readiness in the guide The packager is told to open the executive summary with what the design rests on: how much the customer stated, how much was assumed, how much is unknown. An instruction, not a check. ## Also closed, from the same run - **Only the pinned providers.** `check_providers` in `app/tools/local.py` refuses, before Terraform runs, a resource or data source of a provider the bundle does not pin, and a `required_providers` entry for one. The target architect is told that something outside the Cloud provider's own provider is not an Element. A module's own providers are untouched: they are declared inside the module. - **The scope statement names only the customer's exclusions.** Not done: `intent.exclusions` are free text, so nothing in code checks that an excluded thing was not built. The Specialists are told; in case 08 they complied. Tests: `tests/execution/test_requirements_chain.py` (6 more), `tests/tools/test_local.py` (5 more). ## First run with real models, 2026-10-11 `case-08-landing-zone` again, results `20261011T185421Z`: completed, 91%, 60 minutes, about $11.94. - **The stated topology was built this time.** `networking.topology` is carried by NFR-2, whose first condition is that a Shared VPC host project is connected to service projects across the environment tiers. The code has a host project and the connections. The run before had neither and passed every Check. - **Only the pinned provider is in the code.** The run before brought in the GitHub provider. - **It cost more:** $11.94 against $8.76, with three Revisions of the Target architecture and three of the code, against two and one. Two of them were two Checks that contradicted each other on one key's role, sending the code back and forth before the design was corrected. That is the design's fault, not this step's, but a stricter chain gives a large design more to get wrong. - **One evaluation check failed, and it is a real finding:** the requirements name "Cloud KMS". The customer asked for keys they control; the product was the analyst's addition.