Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Limitations and maturity

The five projects are runnable reference architectures for learning. They are not products, not audited and not certified. They are written to be understood, and each documents its own gaps. This appendix collects the gaps that matter most when you read the book, as of the pinned revisions.

Maturity at a glance

ProjectWhat is implemented and testedNotable gaps (documented upstream)
I · templateRust/Rig agent; gateway with OIDC/PKCE and sessions; segmented networks; MCP capability boundary; deterministic tool policy; input/output controls; four evaluation suites; OpenTelemetry tracing; opt-in signed approvals with point-of-mutation verificationno per-user data authorisation, secrets management, migrations or trace-store access control; the agent runtime holds the MCP key and symmetric approval-signing secret; overall workflow.request_timeout_seconds is parsed but not enforced; pending approvals are in memory; deterministic PII detection is narrower than contextual NER; nonce conflicts and audit rollback still need live checks. → Limitations
II · Sophossingle-agent graph; SQLite checkpoints with sync durability; run records and lazy recovery of interrupted runs; resume; SSE with reconnectionno approvals, guardrails, OpenTelemetry, evaluation or cancellation; no exactly-once side effects; no multi-agent orchestration; several runtime behaviours shown by labs, not unit tests
III · ETF Researchspec-driven, versioned rules engine (extensively unit-tested); advisory-only model enforced in code; mandatory signed approvals with recomputation under lockthe same runtime-held signing secret pattern as Part I; dated data snapshot, no live data, no brokerage or trading; investor profile parsed but not validated; append-only trigger untested; boot-time fund refresh without a history row; past decisions not fully reproducible. → Limitations
IV · Arktosfour public-only tools (pinned by tests); HMAC API-key lookup; HKDF key separation; versioned AES-256-GCM; SQLCipher; standard derivations with test vectorsno signing, broadcasting or balances; plain HTTP (TLS expected from deployment); no master-key rotation or HSM/KMS; admin key operations not logged; the operator can decrypt phrases
V · Taurosactors, ownership, destinations, immutable revisions, state machine, idempotency, exact-payload approval, locks and unique indexes, application-level event record, eight-tool MCP surfaceaudit not database-enforced; no four-eyes rule; no payments, issuing or reconciliation; concurrency tests do not exercise real lock contention; some decisions available only via REST or console

How to read claims in this book

  • Implemented means the pinned code does it.
  • Tested means a named test in the pinned repository checks it. Where a behaviour is only observed in a lab run, the lesson says so.
  • Documented limitation means the project itself states the gap.
  • Challenge, future design or roadmap means it does not exist yet.

The book's editors do not treat a framework feature, a configuration field or a documented intention as a guarantee. A claim must be supported by the pinned implementation and, where stated as tested, by a named test. If you find a statement in the book that the code contradicts, please report it (see Corrections and contributions).

What this book does not claim

  • That any project is secure, production-ready or compliant with any law, regulation or standard.
  • That any project gives financial advice, executes trades or moves money.
  • That checkpointing, approvals or audit trails provide guarantees beyond what their chapters state.
  • That the projects are integrated with one another. See What is integrated, and what is only designed.