01 — The topic map
The hub's core artefact: one page an agent reads to know what each NFR means here, where the canonical material lives, and what the estate actually does about it. Each entry: the position in one line, the best source, the owner site.
Testing — OWNED
Position: "No mocks. No patches." Assert on contracts, not implementation; real Chromium gated on an env var, skipping cleanly; deploy via pytest, numbered tests run top-down. Affordable because Type_Safe objects are cheap to compose in memory — the type system and the testing philosophy are one decision.
Evidence: 4,785 tests passing in 81 seconds; 799 test files; the four tests/ci/ structural guards (one broken — the honest footnote). Canonical: library/guides/v3.1.1__testing_guidance.md + the sg-compute measurements. Cross-link: coding.sgit.ai owns the conventions; this site owns the philosophy and the numbers.
CI pipelines — OWNED
Position: the pipeline is a verifier, not a runner: native per-arch builds, push by digest only, integration-test the pre-tag image, only then assemble the manifest; AMIs are baked, relaunched, and re-verified before tagging healthy; versions auto-increment from one repo-root file.
Evidence: the 36,815-byte ci-pipeline.yml, the two-phase bake-ami.yml — and the lapse: that workflow invokes a binary (sg-play) defined nowhere, sixteen times. Owner of the artefacts: sg-compute.sgit.ai; this site owns the pattern write-up.
Documentation — OWNED
Position: documentation is a truth system, not prose. The reality-document rule: "If the reality document doesn't list it, it does not exist. Briefs are aspirations, not facts." Proposed features labelled PROPOSED. Every page has a markdown twin. Day-indexes, debriefs with good-failure/bad-failure classification, CC BY footer on ~1,100+ files.
Evidence: 1.27M words in one repo alone; 72,339 words of reality docs across 12 domains — and the lapse: the reality index itself 41 versions stale, and a README describing a repo that does not exist. The page writes itself: the system, the rule, and the two places it failed.
IFD — Iterative Flow Development — OWNED
Position: the estate's named methodology — "rapid software development using AI assistance while maintaining engineering rigor… centred on preserving developer flow state." Versioned guides (intro, testing, versioning) at v1.2.x, written explicitly "for LLMs assisting with IFD-based development." Minor versions are the Explorer team's output unit; majors are Villager releases.
Canonical: library/guides/development/ifd/. Nobody outside the estate has ever seen this methodology written up. It is this site's most original owned asset.
Budgets and finance — OWNED (the discipline, not the business)
Position: three pillars. Profitability-first: "until we know the traction… financial projections and future predictions are made-up, because we do not yet have the data." Pre-approve the ladder: approve the overrun positions at approval time — 1.5×, 2×, 5× — "with the stopping point named while it is still cheap to name", because the research says "the person who approved the first million is, on the evidence, the worst available decider on the second." Plus the novel acceptance: "the things not done because this project was funded should themselves be accepted, with an owner." Budget-on-the-step: budgets attached to workflow steps, not projects.
Boundary: the discipline is publishable; the estate's own figures are Tier-3 everywhere. Cross-link: token-spend-as-engineering-problem is wardley-maps.'s line.
Project management — OWNED
Position: "the project manager is where the register becomes work" — a register per project, net score, erosion, the missing reference class. Plus the estate's own PM system as the exhibit: briefs with acceptance criteria, cross-team reviews, day-indexes, numbered asks (N1…) and tasks (T1…), version-prefixed filenames, session handover guides. The brief/debrief system is itself the estate's PM methodology, demonstrated across ~4,000 documents.
Resilience — OWNED (patterns); artefacts live in sg-compute.
Position: design for the failure you had, not the one you imagine. The watchdog (os._exit(2) through a deadlock, GIL reasoning in the source), the two-phase health poller, halt-means-terminate, idle reconciliation, build-time guards each citing a production incident. Every resilience mechanism in the estate traces to a named incident — that is the doctrine, and it is nowhere written down.
Security — LINKS
The NFR posture only: audit-before-the-key · deny-by-default allowlists · key-prefix validation · no-credentials-in-git with CI gates · the disclosed-incident norm. Domains belong to pki. nhi. sg-sentinel. standards.; the LLM boundary to llms..
Serverless and scalability — LINKS
The requirement framing only — "no servers running when there is no traffic, with a defined cost to start" — and the honesty note that throughput is argued, never measured. Platform: sg-compute..
Explainability — LINKS
The engineering habit only: computed, not claimed · unanswered-is-an-output · estimates rendered with ~. The grounding ladder belongs to risks./standards.; provenance to llms..
Architecture — LINKS
Responsibility boundaries as a requirement class ("X is the ONLY class that…"), single-source-of-truth by structure. Conventions: coding..
Backups — ⚠️ THE GAP
151 mentions, no doctrine. The one estate with irreversible publishing (frozen vaults), a keys-vault design, and no written backup/recovery discipline beyond the keys-vault RECOVERY.md proposal. The site should say so — it is the eighth item on the corpus's own NFR list and the only one with no material.
This document is released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).
== briefs/v0.33.62__nfrs-brief-pack__02__the-memory-thesis.md