FIPS Protocol Wind Tunnel

FIPS Wind Tunnel is a deterministic, multi-fidelity protocol experimentation platform for constructing FIPS networks, injecting stochastic and adversarial events, observing exact causal propagation, finding protocol limits, comparing variants, and reducing large failures into replayable cases.

The primary product surface is a native macOS workbench. It renders the same versioned event stream that is persisted in run evidence; the UI does not own or reimplement protocol semantics. The existing static browser analysis remains a secondary forensic surface for completed artifacts.

The authoritative product specification defines the required engine, campaign, search, shrinking, visualization, validation, and natural-language authoring capabilities. The executable capability audit records implemented, modeled, partial, and missing behavior. A backend milestone or static export must not be described as the complete product.

Run the native workbench

scripts/prepare-macos-engine.sh
open FIPSD.xcworkspace

Select the FIPSD scheme and run it. The workbench supports direct experiment controls, local Claude/Codex campaign authoring, deterministic event playback, forward-only natural-language changes to the currently rendered run, saved-artifact event-aware replay with exact dense-window summaries, pause/step/scrub/speed controls, node inspection, scheduled node failures and recoveries, manual lower-root injection, partitions and merges, editable link conditions, parent/ancestry swaps, cost-driven parent oscillation with hysteresis and hold-down, authenticated Sybil waves, direct-link topology capture/editing, weighted per-node Wi‑Fi/BLE/Tor/Ethernet profiles, authored network zones with real shared-medium queue contention, configurable CPU/memory/cache/table/queue limits and deterministic slow-node heterogeneity, exact control-message propagation, graph-native split-horizon Bloom waves, routed synthetic payload projection including persistent segmented streams and synchronized burst processes, explicit source/destination object transfers with automatic MTU packetization and delivered-byte progress, graph-native lookup/cache/session recovery, and honest one-billion-node analytical cohort rendering and evidence export. Its analysis inspector ranks recorded link bottlenecks, reconciles causal work by stage, shows root-arrival amplification, delivery and Bloom-FPR percentiles, congestion groups, causal flame graphs, and heavy-link anomaly samples while keeping fidelity labels beside every result. A side-by-side runner holds seed, topology, traffic, and interventions constant while comparing protocol timer variants. The workbench also launches bounded pairwise adversarial searches and automatically shrinks the highest-amplification reproduction while preserving 90% of its score. For tiny networks, it can exhaust every supported authored action order and retain independently replayable counterexamples. The Bloom, payload, explicit object transfers, natural-language amendments, lookup/session recovery, multi-scale visualization, and mixed-connectivity, shared-medium zones, and interactive interventions, session rekey waves, lookup storms, parent instability, authenticated Sybil admission, and transport-class failover fidelity boundaries are documented explicitly.

Control the app from agents

The local MCP control server lets Claude, Codex, and other MCP clients launch and manage the same visible workbench. Agents can author or amend experiments, inject exact events, change configuration, control playback, inspect current state, wait for completion, analyze artifacts, and explain the active causal state without UI automation.

The macOS app also embeds Claude over ACP behind the left sidebar's Experiment / Agent switch. That session receives the same MCP server directly, runs in bypass-permissions mode, and renders streamed agent replies as Markdown.

Run from the CLI

cargo run -p fips-cli --bin fips-wind-tunnel -- \
  stream examples/m1/root-ratchet-12.yaml --output runs/root-ratchet-12

The stream is JSON Lines using event-stream-v1alpha1. Its event records are identical to the ordered trace written into artifact.json. The native renderer separately records its source-mapped projection as render-frame-v1alpha1 JSON Lines. Its canvas consumes only those frame primitives; an independent raw-artifact oracle regression-checks every visualization mode while keeping synthetic layout and compressed-event boundaries explicit.

Other campaign, scale, oracle, analysis, atlas, and release commands remain available through fips-wind-tunnel --help. Their verified scope and fidelity are documented under docs/.

Grounding

The simulation is pinned to FIPS commit 80c956a6fdb85dde1450969a21891c1158e43267. Executable codecs and checked-in tests outrank prose. Every result declares fidelity and provenance; aggregate, sampled, calibrated, and unsupported behavior stays labeled.

Development gates

cargo test --workspace
cargo clippy --workspace --all-targets -- -D warnings
scripts/check-loc.sh

Hand-written source files have a 300-line soft limit and a 600-line hard limit. See AGENTS.md for the complete engineering contracts.

License

FIPS Wind Tunnel is available under the MIT License.