Research program
Research
RAHN is not a product with a research garnish — it is a research program with an implementation attached. Every claim in the repository is claim-status-marked, every measurement carries its method, and the limitations baseline is the honesty floor. This page maps the program and states plainly what is implemented, what is demonstrated, what is designed, and what is open.
Motivation and problem statement
The problem statement is one paragraph long: networks have no first-class state object. Config tools push desired state, monitoring observes symptoms, and the actual state — the thing that fails — is reconstructed by hand. The underlying question: can network state become a first-class computational object — versioned, causal, verifiable, replayable, and safely evolvable?
The ten research questions
- RQ1 — Deterministic, content-addressed network state
- Implemented for the v0.1 vocabulary; identity stability is test-enforced with cross-platform CI runs (E1).
- RQ2 — Verification before execution
- Implemented; cost measurement pending — latency distribution at 10²–10⁵ objects (E2) has not been run.
- RQ3 — Semantic branches and merges
- Implemented with fail-closed semantics; the E3 conflict study (precision, recall, false-rejection rate) is pending.
- RQ4 — Constitution expressiveness
- Open: which real-world invariants can the current vocabulary express, and at what check cost?
- RQ5 — Causal attribution fidelity
- Open: how faithful can asserted causal chains be?
- RQ6 — Incident reconstruction / replay
- Future work; depends on RQ5.
- RQ7 — Consistency requirements for distributed state
- Requirements-first approach taken; per-replica authority chosen (ADR 0015) after rejecting Raft, CRDTs, and op-shipping on requirements grounds.
- RQ8 — Multi-backend execution
- Backend abstraction implemented (ADR 0016); further backends are future work.
- RQ9 — AI as advisor, not authority
- Framed; nothing AI-related is in the core by design.
- RQ10 — Usefulness without AI
- Framed; the core is deliberately model-free.
Experiments and evidence today
The experiment plan (E1–E7) defines what would count as evidence; the benchmark methodology defines the rules (a result without an environment block is void; p99 is the story; negative results are published). Status, honestly:
- Implemented and test-enforced: determinism of state and identity (E1 via CI on Linux and Windows), canonical round-trips, merge properties (100-seed commutativity checks), observation ingestion baselines, causal-graph operations, sync transport.
- Partially run: E2 and E6 (scoped prototype measurements exist — e.g. the reverse-adjacency causal query improvement from ~2.7 s to ~5.7 ms on a 10k-edge chain — explicitly labeled prototype results, not guarantees).
- Not run: E3 (merge conflict study), E4, E5, E7.
- Recorded debt: O(m·n) batch-apply cost; single-machine baselines only; no production deployment exists anywhere.
Prior art
The prior-art survey covers source-of-truth systems, intent-based networking, formal verification of networks (Header Space Analysis, VeriFlow, NetKAT, Minesweeper), and content-addressed systems. Its own status is honest: a working survey with provisional claims that must be deepened before any novelty assertion. RAHN makes no novelty claims at this stage.
Limitations — the honesty baseline
The limitations document is normative for every other document. Highlights: stable architecture but young implementation (no production use, no external review); causality asserted, never inferred; evidence is partial and single-machine; no novelty claims warranted; O(m·n) apply debt; merge incomplete in principle (compensated by failing closed); determinism costs; and Rust as a choice with consequences. If any page of this site or repository claims past that baseline, it is a bug — report it.
Open problems for contributors
The highest-value open work is measurement, not features: run E2 (verification latency), run E3 (merge conflict study), execute the RQ4 expressiveness study, perform an independent host-safety review of the execution backend. Each has a methodology document and acceptance criteria — see contributing and the curated good first issues.