03 — The owned disciplines, in depth

The five topic areas this site owns outright, each with its position, its best verbatim material, and its honest counter-evidence. These become the site's core pages; 01__ gave each a summary, this gives the substance.

1. Testing

The four non-negotiables, verbatim:

1. No mocks. No patches. Use register_playwright_service__in_memory() and in_memory_stack-style composition.
2. Assert on contracts — schemas, status codes, persisted artefacts — not implementation details.
3. Real Chromium for integration tests. Gate on SG_PLAYWRIGHT__CHROMIUM_EXECUTABLE; skip cleanly when absent.
4. Deploy-via-pytest. Deploy tests are numbered (test_1__create_lambda, test_2__invoke__health_info, …) and run top-down.

Why no-mocks is affordable here and not elsewhere: Type_Safe objects are cheap to construct, and every service has an in-memory composition path — so the real thing is as easy to instantiate as a mock would be. The philosophy and the type system are one decision. The proof: 4,785 tests in 81 seconds.

The structural-guard pattern — CI tests that enforce architecture rules rather than behaviour (object = None banned, UI-in-the-wheel, component snapshots) — with the rule that makes the set healthy: every guard encodes a rule that was violated at least once. Grown from incidents, not checklists.

Counter-evidence to publish: CI runs 67.4% of collectible tests; six import-level breakages hide in the un-run third; one guard has never worked; three tests fail on date arithmetic. All measured, all in the sg-compute pack.

2. CI pipelines

The pattern, distilled from the 36,815-byte pipeline: build once, verify before naming. Native per-arch builds → push by digest only → integration-test the pre-tag image → only then combine digests into a manifest and tag. Nothing gets a name until it has passed as an anonymous digest.

Same shape one level up: the AMI bake relaunches from the baked image and re-verifies before tagging healthy — the artefact must prove itself twice, once as built and once as booted.

Versioning: one repo-root version file, auto-incremented by branch policy (dev bumps minor, main bumps major), read at runtime, used as the image tag. One source, many consumers, no drift — and the counter-evidence is sg_compute/version, a second version file read by nothing.

3. Documentation and the reality system

The governing rule, which is the estate's single best NFR idea:

"If the reality document doesn't list it, it does not exist. … Briefs are aspirations, not facts."

With its supporting disciplines: PROPOSED labels on unbuilt features · the markdown twin · day-indexes closing "all documents are em-dash-free and released under CC BY 4.0" · debriefs classifying failures as good-failure/bad-failure · the session-handover guide ("Don't improvise").

The full honest story: the system exists, 72,339 words across 12 domains — and the index was 41 versions stale while the README described a repository that does not exist. The rule is right and enforcement is manual, which is the same finding as the coding pack's: discipline without tooling is real but fragile. The page should propose the fix the estate would recognise: a reality-doc freshness check in CI, exactly like the licence audit.

4. Budgets and finance

Three pillars, all publishable as discipline (never the estate's own figures):

Profitability-first — "until we know the traction, the product lines, the services, the cost lines, and what users actually buy, financial projections and future predictions are made-up… The path is to discover the market and ship quasi-daily; investment only accelerates the path to profitability." The Explorer-phase corollary: do not build Town-Planner financial artefacts on hypothesis.

Pre-approve the ladder — the overrun positions approved at approval time, the kill point "named while it is still cheap to name", value milestones and sponsor probabilities recorded as a calibration record, grounded in the escalation literature ("the person who approved the first million is, on the evidence, the worst available decider on the second" — with the preregistered finding that a public conditional pledge makes stopping raise trust). Plus the original move: opportunity cost as a first-class acceptance, with an owner.

Budget-on-the-step — budgets attached to workflow steps rather than projects, so containment is structural.

5. Project management

Two layers. The position: "the project manager is where the register becomes work" — a register per project, net score, erosion, the missing reference class; risk management and project management as one discipline with two vocabularies.

The demonstrated system: the estate's own way of working is its PM methodology — briefs with acceptance criteria (the naming brief's nine, scored in the sg-compute pack), cross-team reviews the same day, numbered asks and tasks, version-prefixed filenames sorting chronologically, handover guides, and the good-failure/bad-failure debrief convention. Roughly 4,000 documents of it. No page anywhere describes it as a system. Writing that page — "how a brief becomes work here" — is this site's most original PM contribution, and it doubles as onboarding for every new agent session.

This document is released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).

== briefs/v0.33.62__nfrs-brief-pack__04__site-architecture.md