Stop raw sewage from overflowing into city rivers during a cloudburst—using the motorized floodgates already in the ground instead of a $300 billion concrete rebuild.
Built for the Meta Global AI Developer Hackathon · Indie Track · Theme: Push the Frontier · Open the live twin · Watch the 3-Minute Demo Video (YouTube)
Not a U.S.-only system. Newark and Chicago are the test storms. Any city with a
SWMM5.inppipe file and standarddBZweather radar—London's Thames Water, the EU's 650,000 combined-sewer outfalls, Tokyo, or Nairobi—can run this same controller on its own gates.
Real NEXRAD N0Q radar mosaics (not painted by the app): Left three: Hurricane Ida over Newark, 23:30–00:45 UTC 1–2 Sep 2021 (traces/nexrad/ida-2021-newark/) — municipal float switch spills 12,096,528 gal vs StormGate-Zero 0 gal (J1 3.338 psi). Right three: Chicago west side, 21:00–23:30 UTC 2 Jul 2023 (traces/nexrad/chicago-2023-0702/) — float switch spills 13,673,729 gal vs StormGate-Zero 0 gal (J4 1.846 psi). sam-3.1 segmented both events in the <0f> box grammar (traces/live_models.json → sam_events), meta/muse-glimmer-30b read frame_02.png directly (8ad068e1ef1cadad → SW / SU2, traces/glimmer_radar.json), and basin rainfall is integrated directly from the radar pixels via Marshall-Palmer.
In London, Paris, Tokyo, across 650,000 combined-sewer outfalls in the EU, and in 746 U.S. communities (New York, Chicago, Boston, Newark, St. Louis), household toilets and street storm drains share the same underground pipe—and in separate-sewer cities like Nairobi, storm floodwater pours through surcharged manholes and inflow-and-infiltration (I&I) to overwhelm trunk interceptors. In a cloudburst, those tunnels fill in minutes. To stop raw sewage from erupting out of basement toilets and flooding streets, utilities worldwide open relief weirs and dump untreated waste and stormwater straight into local rivers, harbors, and drinking-water intakes. In the U.S. benchmark cities we test here, that means 850 billion gallons a year of untreated overflow (EPA 833-R-04-001, Aug 2004, pp. ES-2 to ES-5) and civil penalties up to $68,445 per day per violation when a discharge breaches a permit condition, dry-weather ban, or federal consent-decree milestone (33 U.S.C. 1319(d); 40 CFR 19.4 Table 1, effective 8 Jan 2025).
Digging new storage tunnels takes 25–50 years and billions per city (Chicago TARP: since 1975, $4B+; St. Louis Project Clear: $4.7–6B, 2011–2039; NE Ohio Project Clean Lake: $3B, 25 yrs). Yet the motorized floodgates already in the ground still spill because every previous approach has a blind spot:
- Legacy float switches & single-point PLCs are blind: 1950s–1980s mechanical float/vortex gates and motorized gates wired to local single-point PLCs (
IF local_level > threshold) only read water depth at one manhole. With zero radar lookahead, they stay shut until a downstream basin hits 96% while upstream tunnels sit empty. - In-pipe "smart sewer" sensors react too late: Pre-draining a 10M-gallon basin takes 20–45 minutes. By the time an ultrasonic pipe sensor or ground rain gauge rises, the rainwater is already in the concrete.
- Classical MPC & black-box Deep RL failed in utilities: Classical nonlinear MPC solvers go infeasible the moment a gate jams on debris or a crew radios in a fault, and water utilities reject black-box Deep RL because operators are legally liable for basement backups and require glass-box
SWMM5physics verification.
StormGate-Zero replaces that blind float switch with a physics-guarded control loop built on six Meta models (sam-3.1, muse-spark-1.3, muse-wonder-1.0, muse-voice-transcribe-1.0, muse-image-1.0, and open-weight meta/muse-glimmer-30b) and the U.S. EPA SWMM5 hydraulic solver:
- Sees the storm 45 minutes early:
sam-3.1tracks convective rain cores on Doppler radar (IEM N0QNEXRAD mosaics) before the water hits the pavement, while multimodalmeta/muse-glimmer-30bcross-reads the radar crop at the edge. - Pre-drains the sewer before the cloudburst arrives:
muse-spark-1.3sets continuous gate openings[OR1, OR2, OR3, OR4]every 5 minutes, emptying the downstream basin to the treatment plant ahead of the storm crest and holding early runoff in the upstream basins. - Survives broken hardware: when a field operator radios in that Sluice Gate #2 is mechanically jammed at 50%,
muse-voice-transcribe-1.0locks that constraint in real time andmuse-wonder-1.0computes a new gate schedule around the stuck valve. - Never trusts an unsafe AI move: every gate vector proposed by the models is simulated 60 minutes into the future across 12 parallel St. Venant hydraulic rollouts and checked by a local freeboard interlock. Any move that spills a single gallon more than the physics winner is rejected before it touches the plant.
| Without StormGate-Zero (Municipal Float Switch) | With StormGate-Zero (Radar + Meta MPC + SWMM5) |
|---|---|
| Waits until Basin 4 hits 96% depth to open the gate | Sees Core B on radar and pre-drains Basin 4 to 14% |
Spills 14,820,000 gallons of raw CSO on design storm DS-1482 |
Spills 0 gallons (and 0 gallons when Gate #2 jams at 50%) |
Spills 26,498,221.5 gallons on the double-lock storm (A-predrain spills 8.53M, B-balanced spills 4.36M) |
Spills 0 gallons (traces/wonder_unaided_recovery.json & traces/double_lock_closed_loop.json) |
| Spills ~12.1M gal on the Newark Ida radar mosaic & ~13.7M gal on Chicago | Spills 0 gallons on both radar mosaics |
| A violating discharge hits the $68,445 / day Clean Water Act ceiling | $0 exposure on a dry weir · kept schedule peaks at 3.539 psi on Junction J3 |
Two floodgates jam mid-cloudburst. Static sewer rules dump 26.5 million gallons of raw sewage into the river. StormGate-Zero hears the crew's radio call, re-solves the broken pipe network on the fly, and spills 0.
Pre-trained neural controllers and classical MPC solvers are locked to a fixed pipe and gate layout: the moment a physical floodgate jams on debris mid-storm—or a field crew reports a broken valve over VHF radio—fixed PLC rules overflow and classical optimizers fail. StormGate-Zero uses six Meta models to turn live Doppler radar pixels and spoken radio calls into updated hydraulic constraints on the fly, finding safe gate openings when hardware breaks:
- Static rules fail (
4.36M – 26.50M galspilled): On the compound double-lock storm (DS-1482 × 2.2,OR1locked at5%,OR2locked at90%from minute 20), the municipal float switch spills 26,498,221.5 gal, the hand-writtenA-predrainlaw spills 8,534,246.3 gal river + 897,741.2 gal street, and the best fixed law (B-balanced) still spills 4,358,862.8 gal. - Blind one-shot AI fails (
4.33M – 6.97M galspilled, rejected): Askingmuse-spark-1.3(resp_6ae9e184c25d838e5d884bcd) ormuse-wonder-1.0(resp_6ae9e19fa36ac7ab266e474f) to guess all 18 steps blind in one shot spills 4,333,880.8 gal and 6,966,494.7 gal—and the physics guard automatically rejects both (rejected-spilled-more). - Physics-grounded Meta AI succeeds (
0 galriver,0 galstreet): Shown only the 4 failing hand-written laws at the moment the step-4 jam hits—with no grid points in the prompt—muse-wonder-1.0(resp_6ae9f0ee5b7721d9ff6a4da4,11.85 s) synthesized an off-grid continuous recovery vectoru = [0.05, 0.90, 0.12, 0.90]("Hold back SU3 with OR3=0.12 to buffer the stuck-open OR2 inflow without flooding SU3. Drain SU4 aggressively with OR4=0.90 to create downstream capacity and route stored flow to treatment") that holds the entire storm to 0.0 gallons of river CSO and 0.0 gallons of street flooding (traces/wonder_unaided_recovery.json). Closed-loop step-by-stepmuse-spark-1.3(18calls,traces/double_lock_closed_loop.json) also holds 0.0 gallons.
| Forcing on the 4-basin interceptor | What the tape is | Municipal float switch | All gates wide open | StormGate-Zero |
|---|---|---|---|---|
| DS-1482 (design storm) | 90 min, 18 × 5-min steps, played in ~20 s | 14,820,000 gal | ~3,100,000 gal | 0 gal |
| DS-1482 + OR2 jammed at 50% | Sluice #2 mechanically locked from minute 40 | 14,820,000 gal | — | 0 gal |
| DS-1482 × 2.2 + OR1=5%, OR2=90% | Compound double lock from minute 20 (B-balanced spills 4.36M) |
26,498,221.5 gal | 4,531,586.3 gal | 0 gal |
| Hurricane Ida mosaic (menu) | IEM N0Q crops, Newark, 1 Sep 2021, 23:30–00:45 UTC | ~12,096,528 gal | — | 0 gal |
| Chicago 2 Jul 2023 mosaic (menu) | IEM N0Q crops, west side, 21:00–23:30 UTC | ~13,673,729 gal | — | 0 gal |
| Newark 3.24 in/hr gauge hour | Saturation check: 8–9 p.m. local cloudburst hour | ~52.8 million gal | — | ~26.2 million gal |
| Chicago gauge hour | Saturation check: gauge-shaped cloudburst hour | ~24.5 million gal | — | ~9.2 million gal |
- Why the float switch is worse than leaving the gates open: the municipal float holds OR4 at 2% until Basin 4 is at 96% of crown, then throws it to 100%. It stores the first wave and dumps the second over the river weir. Leaving every gate open lets the treatment plant pass a flatter hydrograph (~3.1M gal spilled instead of 14.82M). StormGate-Zero pre-drains Basin 4 before the second convective core arrives and holds the crest off the weir.
- Why the Newark 3.24 in/hr row still spills ~26.2M gal: real tunnels have finite concrete volume. The mosaic menu plays the IEM palette rainfall, which sits below the gauge peak and holds at 0 gallons. The gauge-shaped reconstruction of Newark's published 3.24 in hour exceeds physical storage. A controller that claimed 0 gallons on that hour would be faking the mass balance.
- Statutory exposure: the planning readout treats any overflow above 0 gallons as one violation-day at the 40 CFR 19.4 ceiling of $68,445 (
33 U.S.C. 1319(d), penalties assessed on or after 8 January 2025). That ceiling applies when the discharge violates a permit condition, a dry-weather prohibition, or a consent-decree milestone, not to every permitted wet-weather overflow. The board's E. coli readout dilutes a planning density of1,000,000 CFU / 100 mLinto10,000,000 gallonsof receiving water against the EPA 2012 beach action value (235 CFU / 100 mL). - Dynamic-wave check:
verify_dynamic_waveruns SWMM5DYNWAVEon the committed gate schedule. The keptmuse-spark-1.3refine schedule on DS-1482 peaks at Junction J3, 3.539 psi above the crown. The pure pre-drain law on that same storm peaks at J1, 2.51 psi. The Ida mosaic peaks at J1, 3.338 psi. The Chicago mosaic peaks at J4, 1.846 psi. The double-lock schedule below peaks at J3, 5.306 psi. Level-pool gallons and these heads are two different checks.
Design storm DS-1482 scaled by 2.2×, with OR1 locked at 5% and OR2 locked at 90% from minute 20:
| Controller | River spill | Street flood | Verdict / Trace |
|---|---|---|---|
| Municipal float switch | 26,498,221.5 gal | 0 | Baseline (traces/compound_fault.json) |
| C-retention (hold upstream) | 24,303,980.4 gal | 1,096,811.6 gal | Fails (traces/compound_fault.json) |
| A-predrain (the hand-written StormGate law) | 8,534,246.3 gal | 897,741.2 gal | Fails (traces/compound_fault.json) |
muse-wonder-1.0 blind 18-step one-shot |
6,966,494.7 gal | 0 | rejected-spilled-more (resp_6ae9e19fa36ac7ab266e474f) |
| D-bypass (all free gates 100%) | 4,531,586.3 gal | 0 | Fails (traces/compound_fault.json) |
| B-balanced (best of the four named laws) | 4,358,862.8 gal | 0 | Best static law (traces/compound_fault.json) |
muse-spark-1.3 blind 18-step one-shot |
4,333,880.8 gal | 0 | Beats B-balanced by 24,982 gal; rejected-spilled-more (resp_6ae9e184c25d838e5d884bcd) |
muse-wonder-1.0 unaided step-4 recovery (no grid shown) |
0 gal | 0 | Kept — synthesized off-grid u = [0.05, 0.90, 0.12, 0.90] (resp_6ae9f0ee5b7721d9ff6a4da4, traces/wonder_unaided_recovery.json) |
muse-spark-1.3 closed loop (18 step-by-step calls) |
0 gal | 0 | Kept (resp_6ae9e477f80e73f48f264cc5 .. resp_6ae9e54995a7cf28d83b49cc, traces/double_lock_closed_loop.json) |
Across 100 deterministic convective-shear × valve-lock scenarios, the schedule in traces/stress_matrix_100.json never spilled more than the best named law (14 outright wins, 86 ties within 1 gal, 1,716,996,992.5 gallons saved vs. the float switch). Every row, PySWMM pressure head, and hosted trace is sealed in traces/compliance_ledger.json (4 blocks, tip 37c1c271ebc7b645).
Every model feeds the gate command or the operator verification of that command.
IEM N0Q NEXRAD mosaic / DS-1482 field SCADA war-room audio (WAV)
| |
v v
sam-3.1 (POST /v1/responses) muse-voice-transcribe-1.0
<0f> boxes + RLE masks ∩ basins POST /v1/asr/transcribe
+ meta/muse-glimmer-30b radar vision "Sluicegate 2 ... jammed at 50%"
| |
+------------------+---------------------+
|
v
12 parallel 60-min St. Venant rollouts
A-predrain · B-balanced · C-retention · D-bypass
|
v
muse-spark-1.3 — Turn 1 continuous [OR1..OR4]
muse-spark-1.3 — Turn 2 score-conditioned refine
|
v
muse-wonder-1.0 — jammed-gate MPC recovery (OR2 = 0.50)
|
v
Physics ranker + local freeboard interlock
+ meta/muse-glimmer-30b edge operator ruling
(depth > 92% & inflow > 35 mm/hr -> force >= 60%)
|
+----------------+----------------+
v v
Actuate OR1..OR4 on the plant muse-image-1.0
PySWMM DYNWAVE check on finish Outfall visual twin (once)
kept schedule J3 = 3.539 psi 14.82M gal weir vs. dry weir
Why one model cannot do all six jobs (measured in traces/):
- Vision LLM vs.
sam-3.1spatial geometry: When we fed the Hurricane Ida radar crop (frame_02.png) to multimodalmeta/muse-glimmer-30b(8ad068e1ef1cadad,57.2 s), it could only name the hottest quadrant (SW / SU2) and explicitly noted that "pixelation and the lack of calibrated dBZ values limit precise pixel-level quantification."SWMM5cannot integrate runoff from a quadrant label; it needs exact<0f>bounding boxes and RLE masks (<|box;x1=...;y1=...|><|mask;...|>), which onlysam-3.1produces (14.5 sfor 6 frames). - Text/vision
/v1/responsesvs.muse-voice-transcribe-1.0diarized WAV audio: Field crews report jammed valves over VHF radio./v1/responsesdoes not ingest raw 16 kHz PCMaudio/wavwith speaker diarization; onlymuse-voice-transcribe-1.0(/v1/asr/transcribe) turns spoken radio audio into a speaker-tagged lock (OR2 = 0.50). - Routine step planner (
muse-spark-1.3) vs. fault-recovery synthesizer (muse-wonder-1.0):muse-spark-1.3drives the 5-minute step loop (0|..17|), whilemuse-wonder-1.0is invoked only on the beat a mechanical fault hits to synthesize an off-grid continuous recovery equilibrium (u = [0.05, 0.90, 0.12, 0.90]in11.85 s,resp_6ae9f0ee5b7721d9ff6a4da4) without needing a pre-solved grid. - Cloud planner vs. open-weight edge safety (
meta/muse-glimmer-30b) & visual twin (muse-image-1.0): Municipal SCADA rooms require an open-weight edge model (meta/muse-glimmer-30b) that explains local freeboard interlock overrides without relying on the cloud planner, while onlymuse-image-1.0emits pixel plates (image_generation_call) for the outfall counterfactual.
| Model | Endpoint | Recorded proof in traces/ |
|---|---|---|
sam-3.1 |
POST /v1/responses |
DS-1482: boxes + masks at t=40 ([0.3792, 0.4813, 0.5417, 0.6417]) and t=55 ([0.5229, 0.6083, 0.6771, 0.7604]) in <0f>0<|box;x1=…|> grammar. Hurricane Ida mosaic: storm boxes on 4 of 6 frames. Chicago mosaic: yellow-core boxes on 5 of 6 frames ([12, 10, 10, 0, 2, 10] objects). |
muse-spark-1.3 |
POST /v1/responses |
On the double-lock tape, 18 closed-loop calls each returned the lowest measured opening. Replaying them spills 0 gal against 4,358,862.8 gal for the best named law (traces/double_lock_closed_loop.json, resp_6ae9e477f80e73f48f264cc5 … resp_6ae9e54995a7cf28d83b49cc). A blind 18-row guess from the same model spills 4,333,880.8 gal. DS-1482 turn-1 keys 0|..17| and the refine pass (15 kept, 2 rejected, 1 null) stay in traces/recorded_refine_pass.json. |
muse-wonder-1.0 |
POST /v1/responses |
On the step-4 compound double lock (OR1=0.05, OR2=0.90, no grid shown), synthesized off-grid u = [0.05, 0.90, 0.12, 0.90] (resp_6ae9f0ee5b7721d9ff6a4da4, 11.85 s), cutting spill from 4.36M gal (best static law) to 0.0 gal (traces/wonder_unaided_recovery.json). On the DS-1482 step-8 jam, returned u = [0.04, 0.50, 0.04, 0.6209] (resp_6ae9adeac24e0d3e800f40f3, 6.50 s). |
muse-voice-transcribe-1.0 |
POST /v1/asr/transcribe |
Multipart request + 16 kHz mono WAV audio (audioEncoding: "WAV", mode: "DIARIZATION"). Transcribed: "Operator 4, Sluicegate 2 is mechanically jammed at 50% open." Parser locks OR2 = 0.50. |
muse-image-1.0 |
POST /v1/responses |
Rendered the side-by-side outfall plate once (image_generation_call WebP converted to PNG and locked behind data/outfall/.muse_spent so loops never re-bill). |
meta/muse-glimmer-30b + local freeboard interlock |
glimmer_edge_guard.py + POST https://integrate.api.nvidia.com/v1/chat/completions |
1. Freeboard override (traces/glimmer_nvidia.json, bd76ae3ac432bcce, 848 tokens): When SU1 hits 95% depth at 48 mm/hr with OR1 pinched to 10%, Python forces OR1 -> 60% ([0.60, 0.40, 0.40, 0.62]) and Glimmer issues the operator ruling. 2. Radar vision (traces/glimmer_radar.json, 8ad068e1ef1cadad, 1,610 tokens): Reads ida-2021-newark/frame_02.png and identifies the peak echo in the SW quadrant (SU2, 28.08 mm/hr). |
| Proof | Result | Trace & Test |
|---|---|---|
| Six Meta models | Recorded traces for sam-3.1, muse-spark-1.3 (Turn 1, Turn 2, and 18-step closed loop), muse-wonder-1.0, muse-voice-transcribe-1.0, muse-image-1.0, and meta/muse-glimmer-30b (freeboard ruling + radar crop). |
traces/live_models.json · traces/recorded_refine_pass.json · traces/double_lock_closed_loop.json · traces/glimmer_nvidia.json · traces/glimmer_radar.json |
| Compound-fault recovery | Under DS-1482 × 2.2 with OR1=0.05 and OR2=0.90, all four hand-written laws spill 4.36M–24.30M gal. Unaided muse-wonder-1.0 (resp_6ae9f0ee5b7721d9ff6a4da4, off-grid u = [0.05, 0.90, 0.12, 0.90], no grid shown) and 18-step closed-loop muse-spark-1.3 both hold 0 gal CSO and 0 gal street flood. |
traces/wonder_unaided_recovery.json · traces/double_lock_closed_loop.json · tests/test_closed_loop.py |
| Physics rejection guard | Blind 18-step one-shot schedules from muse-spark-1.3 (4,333,880.8 gal, resp_6ae9e184c25d838e5d884bcd) and muse-wonder-1.0 (6,966,494.7 gal, resp_6ae9e19fa36ac7ab266e474f), plus steps 7 & 11 of Turn-2 refine, are rejected (rejected-spilled-more). |
traces/double_lock_hosted.json · traces/recorded_refine_pass.json · tests/test_double_lock_hosted.py |
| 100-scenario stress matrix | 100 convective shear × valve-lock scenarios: 0 / 100 losses against the best named law (14 outright wins, 86 ties within 1 gal) and 1,716,996,992.5 gallons saved vs. the municipal float switch. | traces/stress_matrix_100.json · backend/physics/frontier.py · tests/test_frontier.py |
| SWMM5 dynamic wave + SHA-256 chain | PySWMM (SWMM5 DYNWAVE) surcharge pressures (psi) across DS-1482, Ida, Chicago, and compound-fault, sealed into a 4-block SHA-256 chain (tip = 37c1c271ebc7b645). |
traces/wave_storms.json · traces/compliance_ledger.json · python scripts/verify_all.py |
The public twin is stormgate-zero.onrender.com. It replays the saved model traces through the live hydraulic solver.
STORMGATE_REFINE defaults to 0, and traces/live_models.json is committed. You can clone the repo without any API key, verify all 27 tests and the 4-block SHA-256 ledger in under 2 seconds, or launch the interactive SCADA twin:
git clone https://github.com/brian-mwirigi/stormgate-zero.git
cd stormgate-zero
python -m venv .venv
.venv\Scripts\activate # Windows
# source .venv/bin/activate # macOS / Linux
pip install -r requirements.txt
copy .env.example .env # leave MODEL_API_KEY and NVIDIA_API_KEY blank for replay mode
python scripts/verify_all.py
python -m pytest tests -qTo launch the live SCADA twin (http://localhost:5173):
python scripts/generate_demo_assets.py
uvicorn backend.main:app --port 8000Second terminal:
cd frontend
npm install
npm run dev27 passed, offline (STORMGATE_OFFLINE=1 in tests/conftest.py). The test suite locks:
14,820,000 galfloat-rule spill on DS-1482 (tests/test_hydraulics.py).0 galStormGate-Zero spill on DS-1482, both clean and withOR2jammed at0.50(tests/test_hydraulics.py).- All-gates-open spilling less than the float switch, and pre-drain ranking ahead of balanced/retention/bypass (
tests/test_hydraulics.py). - The
<0f>object-multiplex grammar and radio fault parser (tests/test_parsers.py). - IEM N0Q palette anchors (
index 150 → 45 dBZ,index 170 → 55 dBZ), live 60-dBZ cell drop,$68,445statutory ceiling,keep_refinementguard, andPySWMMdynamic-wave verification (tests/test_depth.py). - Double-lock compound fault replay (
0 galvs4.36M galbest static law), 100-scenario stress matrix (14wins,86ties,0losses),PySWMMsurcharge heads onDS-1482,Ida,Chicago, andcompound-fault, and tamper detection on the 4-block SHA-256 compliance ledger (tests/test_frontier.py). - Hosted
muse-spark-1.3(4,333,880.8 gal) andmuse-wonder-1.0(6,966,494.7 gal) blind 18-step schedule scoring and physics rejection (tests/test_double_lock_hosted.py). - Unaided
muse-wonder-1.0off-grid recovery (u = [0.05, 0.90, 0.12, 0.90],resp_6ae9f0ee5b7721d9ff6a4da4,0 gal), closed-loopmuse-spark-1.318-call zero-spill replay (resp_6ae9e477f80e73f48f264cc5..resp_6ae9e54995a7cf28d83b49cc), andmeta/muse-glimmer-30bHurricane Ida radar quadrant identification (tests/test_closed_loop.py). meta/muse-glimmer-30bfreeboard interlock phrasing (bd76ae3ac432bcce) and offline invariant isolation (tests/test_glimmer_trace.py).
backend/
config.py Model IDs, GIS subcatchments, NVIDIA & Brev hooks
episode.py 20-second twin loop, storm switching, DYNWAVE trigger
main.py FastAPI + /ws/twin WebSocket + /api/storm & /api/verify/wave
physics/
design_storm.py DS-1482 two-core storm and VOLUME_SCALE calibration
storm_catalog.py Active forcing (ds-1482, Ida, Chicago, click-to-drop 60 dBZ)
nexrad_mosaic.py IEM N0Q archive downloader, 0.005° world-file crop, Z-R
swmm_engine.py 4-basin level-pool St. Venant plant + 12 parallel rollouts
dynamic_wave.py PySWMM DYNWAVE heads, kept schedule J3 = 3.539 psi
frontier.py 100-scenario search against the four named laws
compliance_ledger.py SHA-256 chain over those measured files
exposure.py 33 U.S.C. 1319(d) $68,445/day ceiling + E. coli dilution
agents/
sam_radar_tracker.py sam-3.1 <0f> prompt builder + <|box;...|><|mask;...|> parser
spark_open_worker.py muse-spark-1.3 Turn-1 vector + Turn-2 score-conditioned refine
wonder_mpc_supervisor.py Ranker, keep_refinement physics guard, muse-wonder-1.0 jam call
glimmer_edge_guard.py Local freeboard interlock + meta/muse-glimmer-30b NIM ruling
voice_scada_bridge.py 16 kHz WAV converter, /v1/asr/transcribe, OR2 jam parser
image_counterfactual.py muse-image-1.0 outfall plate with single-spend lock
meta_client.py Windows system trust store TLSv1.3 client for api.meta.ai & NIM
model_traces.py Zero-rebill replay cache loader/writer
frontend/
src/App.tsx Command center, NEXRAD provenance banner, storm & SWMM5 readouts
src/components/RadarSamCanvas.tsx NEXRAD frame viewer, <0f> SAM overlays, click-to-drop cell
src/components/HydraulicTwin.tsx Dual 4-basin HGL sewer cross-section and gate indicators
src/components/SpillRace.tsx Cumulative river overflow chart (float switch vs. StormGate-Zero)
src/components/AgentCommentary.tsx 60-min rollout ranking and freeboard interlock ruling
src/components/VoiceWarRoom.tsx Field radio microphone & mechanical gate jam injector
scripts/
verify_all.py One-command offline audit of all traces, PySWMM heads & SHA-256 chain
build_frontier_proof.py Rebuilds compound_fault, stress_matrix_100, wave_storms & ledger
record_wonder_unaided.py Unaided muse-wonder-1.0 step-4 recovery recorder (no grid shown)
record_closed_loop.py 18-step closed-loop muse-spark-1.3 recorder on the double-lock storm
record_double_lock_hosted.py Blind 18-step Spark & Wonder recorder + physics guard verdict
record_glimmer_radar.py Multimodal meta/muse-glimmer-30b recorder on the Ida NEXRAD crop
traces/
compliance_ledger.json 4-block SHA-256 chain (tip 37c1c271ebc7b645)
compound_fault.json Double-lock schedules and gallon scores
wonder_unaided_recovery.json Unaided muse-wonder-1.0 off-grid [0.05,0.90,0.12,0.90] -> 0 gal
double_lock_closed_loop.json 18 closed-loop muse-spark-1.3 responses replaying to 0 gal
double_lock_hosted.json Blind 18-step Spark & Wonder responses rejected by physics
glimmer_nvidia.json meta/muse-glimmer-30b freeboard ruling (bd76ae3ac432bcce)
glimmer_radar.json meta/muse-glimmer-30b Ida radar crop read (8ad068e1ef1cadad)
stress_matrix_100.json 100-scenario shear x fault matrix (14 wins, 86 ties, 0 losses)
wave_storms.json PySWMM DYNWAVE heads for DS-1482, Ida, Chicago, and compound fault
recorded_refine_pass.json 18-step Turn-1 -> Turn-2 -> solver accept/reject log + Wonder jam
recorded_live_pass.json First scored Spark (17/18), SAM (t=40,55), and Voice pass
live_models.json Full replay cache used by the running server
nexrad/ Ida & Chicago 480x480 N0Q crops, meta.json, and rain.npy
tests/
test_hydraulics.py 14.82M vs 0 gal, jam resilience, rollout ranking
test_parsers.py <0f> grammar round-trip and radio fault extraction
test_depth.py Mosaic Z-R, cell drop, $68,445 ceiling, refine guard, PySWMM wave
test_frontier.py Double-lock replay, 100-scenario matrix, PySWMM heads, SHA-256 tamper
test_double_lock_hosted.py Blind Spark/Wonder replay and physics guard rejection
test_closed_loop.py Unaided Wonder 0-gal + 18-call Spark 0-gal + Glimmer radar read
test_glimmer_trace.py Glimmer NIM freeboard ruling + offline interlock isolation