Repository skeleton

Repository skeleton

PLATFORM REPOS (built once, reused by every client)
  pe-ontology/                      Knowledge layer: versioned releases, read-only from above
    releases/vX.Y/                  concepts, relationships, document taxonomy, declared rules
    overlays/                       strategy overlays, selected on the plane
    benchmarks/                     promoted, anonymized aggregates only
    docs/GLOSSARY.md                one word, one meaning

  pe-solution-layer/                client-blind code, reused unchanged
    CLAUDE.md                       root constitution: standards, commands, review rules
    .claude/
      settings.json                 shared configuration, checked in
      settings.local.json           personal overrides, never committed
      commands/                     run-use-case.md  validate.md  regenerate-views.md  promote.md  compare.md
      hooks/                        blindness-linter  schema-check  ref-integrity  tenant-isolation  eval-regression  formatter
      agents/                       orchestrator.yml + one .yml per business agent (subagents defined inside their parents)
    skills/
      <use-case>/
        SKILL.md                    the one file a new use case adds
        scripts/                    executable steps the skill calls
        references/                 documents loaded on demand, never pasted into context
        assets/                     output templates and static files
    domain-lib/                     deterministic rule implementations, exposed as thin tools
    .mcp.json                       south server registrations (which servers exist; who may call them is plane content)
    tests/                          unit/  integration/  e2e/ for the deterministic code
    docs/                           architecture.md  api-reference.md  onboarding.md

  pe-platform-infra/                the harness and platform half
    terraform/                      cloud modules, one set per provider
    gateway/north/                  per-tenant endpoint template, stamped at provisioning
    servers/south/<source>/         one tool server per source, definitions to the tool-design standard
    evals/runner/                   suite execution, score storage, judge machinery
    skeleton/                       the SSRA templates this whole tree is stamped from
    ci/                             the five gates
    scripts/                        setup  deploy  seed
    Dockerfile · .env.example       runtime image; variable names only, values live in the vault

  pe-platform-apps/                 P1Z-built client surfaces: thin shells over one shared core
    packages/ui-core/               render tree, streaming message views, conversation state machine
    apps/web/                       browser shell
    apps/desktop/                   native shell + webview
    apps/mobile/                    iOS and Android shells
    apps/voice/                     future surface; same core, new shell

PER-ENGAGEMENT REPO (stamped from the skeleton; values only)
  client-<name>/
    CLAUDE.md                       GENERATED from the binding; regenerates on plane changes
    binding.yaml                    every plane slot valued
    ontology.lock                   pinned Knowledge release ID
    .claude/agents|skills|commands  pinned from the solution layer, unchanged; activation list from the binding
    mcp.config                      north endpoint (tenant auth, host allow-list) + south credentials as vault references
    connectors/                     source selections and field mappings from discovery
    evals/baselines/                the client's golden exemplars, per the data-rights clause
    README.md                       generated engagement summary

The runtime walk of these files (which file governs which step of a live request) is the Interactive Judgment pattern, P1, animated in the Architecture section, on the Pattern walkthroughs page.