RAGfish is the architecture hub for the Noema Architecture — a governed AI design in which human judgment governs and AI accompanies. This repository holds the architecture decision records, the governance contract, the capability matrix, and the execution roadmap. Runtime code lives in the sibling repositories.
AI is not the subject. AI is a capability exercised under human-controlled policy, at defined boundaries, with a durable audit trail.
The current baseline is the epistemic lifecycle introduced by ADR-0004:
Evidence
│ user-owned corpus, plus — only after a Human Gate approval — externally acquired sources
▼
NOESIS investigation and calibrated epistemic assessment
│ (may produce a proposal to acquire more external evidence)
▼
Human Gate a human explicitly authorizes or declines any crossing of an
│ external trust boundary — required only when acquisition is needed
▼
External acquisition executes only within the approved scope
│
▼
NOESIS re-analysis recompute over the enlarged evidence set — not a patch
│
▼
investigation_result the sole NOESIS → NOEMA handoff object
│
▼
NOEMA presentation / narrative transformation only
The lifecycle may iterate (NOESIS → Human Gate → acquisition → re-analysis) while a
human keeps approving. It always terminates in NOESIS producing an
investigation_result; NOEMA is always downstream of it.
This is an added dimension over the existing governance model, not a replacement: the decision / execution / knowledge / audit planes remain separated.
| Principle | Meaning |
|---|---|
| Human sovereignty | Human noesis governs the lifecycle and owns every governance decision. AI accompanies. |
| NOESIS assesses; human noesis governs | The NOESIS layer produces advisory assessments, confidence, hypotheses, and recommendations. It never owns sovereign judgment, routing authority, or approval. |
| "Can't rather than won't" | Governance-critical restrictions are enforced by absent capability — no credential, no endpoint, no code path, no field, schema/type constraint, separate execution context — not by prompt instruction alone. |
investigation_result is the sole handoff |
NOEMA receives investigation_result and nothing else from the investigation side. |
| NOEMA has no evidence-acquisition capability | No retrieval, no evidence store, no external API, no corpus-boundary crossing. NOEMA also cannot introduce new claims or raise confidence above the NOESIS result. |
| Separation of planes | Decision, execution, knowledge, and audit remain distinct. No component both decides and executes. The audit record is append-only. |
Full definitions and the enforcement mechanism for each restriction are in the Governance Contract and the Capability Matrix.
Primary entry points — read these first:
| Document | Purpose |
|---|---|
| ADR-0004 — Noesis / Noema Epistemic Separation | The current architectural baseline: the epistemic lifecycle, the Human Gate, and "can't rather than won't". |
| Governance Contract | What each architectural actor may, must, must not, and cannot do — with the enforcement mechanism for every governance-critical restriction. |
| Capability Matrix | Capability ownership as an explicit grid (18 capabilities × 6 actors) plus the enforcement-mechanism mapping. |
| Execution Roadmap v3 | The current execution baseline. Phase 0–8 plan, task catalog, and dependency graph. |
| Project Charter | Vision, mission, principles, and success criteria. |
| Human-Governed Development Loop | Task lifecycle, task fields, branch and PR discipline. |
Foundational ADRs — still authoritative:
- ADR-0000 — Human Sovereignty Principle
- ADR-0001 — Noema Architecture — four-repo structure and responsibility boundaries
- ADR-0002 — Noema Governance Pipeline — the nine-stage pipeline
- ADR-0003 — Governance Platform — four-plane model and machine-checkable rules G1–G6
Historical / superseded — kept for traceability, not current guidance:
- Execution Roadmap v2 — superseded for execution planning by Roadmap v3; retained as historical evidence.
| Repository | Role |
|---|---|
rag-fish/RAGfish |
Architecture hub — ADRs, governance contract, capability matrix, roadmap, public narrative. |
rag-fish/NoesisNoema |
Client / runtime implementation — local retrieval, on-device inference, user interaction. |
rag-fish/noema-agent |
Constrained execution service — exists to execute, not to decide. Stateless; policy enforcement, route-contract formation, verification, decision logging. Distinct from the epistemic NOEMA layer. |
rag-fish/noesisnoema-pipeline |
Evidence / package pipeline — source extraction, chunking, embedding, corpus quality gates, provenance and trust metadata. |
Repository boundaries are enforced by ADRs and contracts, not by code coupling.
All work follows the Human-Governed Development Loop: 1 task = 1 issue = 1 branch = 1 PR, with a human as the final reviewer and merge owner.
Ready task
→ read the authoritative references (ADRs, governance contract, roadmap)
→ inspect the target repository's code
→ produce an implementation plan
→ review
→ implement + validate
→ open a PR
→ human review / merge
Every task states the standard fields (target repo, task type, owner, branch, Definition of Done, validation) — see the loop document — plus:
README Impact: YES / NO— with a one-line rationale. IfYES, the task must update this README (or spawn a follow-up task that does), so the public entry point never drifts from the architecture baseline.
TBD