Roadmap

Roadmap

The project advances in phases; each closes only when its success criterion is automated as a test. Milestone notes are in the ADR's Hitos section (docs/decisiones.md).

Phase	Scope	Status
0 — Feasibility spike	Compile FBW's systems + a320_systems natively, instantiate the A320, tick it, read electrical vars, inject a failure	Done — success criterion met; zero patches to vendored code; 102 upstream electrical tests pass natively; stack settled (D-004)
1 — Core + API + CLI	Persistent runtime over Simulation<A320>, variable registry, set/get/step API, human REPL; electrical vertical slice on ground	Done (2026-07-15) — automated as core-rs/tests/electrical_slice.rs; curated control catalog, PyO3 bindings (a320_sim) and REPL with live watch shipped
2 — Failures + detection	inject_failure / list_failures, read_ecam	Done (2026-07-17) — automated as core-rs/tests/generator_caution.rs: the APU generator fails and its caution appears. Stable failure ids (D-013); the ECAM catalog is a homegrown rule engine — no FWC in the Rust (D-014)
3 — MCP server	Expose the API as MCP tools; end-to-end demo	Done (2026-07-17) — nine tools over stdio; an LLM resolved the APU GEN failure from the ECAM alone. The active-failure list is deliberately not a tool (D-016)
4 — More systems	Hydraulics, APU, fuel, engine start; richer world boundary (N2 input)	Done (2026-07-22) — the full cold & dark → engines running sequence on catalog names alone, automated as core-rs/tests/cold_dark_to_engines_running.rs and available via a320-mcp --start engines-running. Own engine spool (D-019), bleed-gated starts (D-020), seeded panel states (D-021)
5 — Benchmark (research)	Scenario suite with QRH ground truth, trajectory-level compliance scoring, baselines + ablations	In progress — scenario loader + runner + recorder + scorer and 4 scenarios shipped (D-022 – D-026); remaining: broader coverage, multi-model baselines + ablations

What "done" meant per phase

Each closure is backed by an automated success criterion and left the vendor pin intact with zero patches to FBW's code:

• Phase 1 — cold & dark → batteries → external power, the network coming alive.
• Phase 2 — drop a generator, see the caution appear (the loop that justifies the project).
• Phase 3 — an LLM closed the observe→reason→act→advance loop over the real MCP protocol.
• Phase 4 — cold & dark → both engines at idle on their generators, all three hydraulic circuits pressurized, APU shut down, clean ECAM.

What's next (Phase 5)

The evaluable environment and the scoring metric exist. The remaining work is the paper: broader scenario coverage (more systems), the blind multi-model baselines, and the ablations (e.g. the instructions-variant axis). See Benchmark-Overview and Design-Decisions.