HydroSwarm¶
Offline, physics-first decision support for drinking-water contamination incidents.
HydroSwarm helps responders narrow plausible sources, decide whether another sample would meaningfully reduce uncertainty, compare response options against the hydraulic network, and stop before field action at an explicit human-approval boundary.
Watch the 3:23 demo GitHub Devpost Run v0.2.1
Research software, not production control. All reported model and evaluation data are synthetic. HydroSwarm does not identify contaminant chemistry, certify water safety, replace laboratory or utility procedures, or execute infrastructure actions.
Judge quick path¶
-
Quick review · 5–10 min
Understand the problem, workflow, headline results, and limitations without reading the model internals.
-
Technical review · 10–20 min
Inspect the software architecture, authority separation, simulator verification, and evidence trail.
Architecture → Final System → Authority & Safety → Judging Evidence Map
-
Scientific / ML review
Audit supervision scope, calibration applicability, split governance, locked evaluation, and exact claims.
Model Card → Evaluation → Dataset Card → Reproducibility → Claims & Evidence
Explore the product workflow →
How HydroSwarm works¶
flowchart LR
O["Observe"] --> L["Localize"]
L --> G["Gate"]
G --> S["Sample / Abstain"]
S --> P["Plan"]
P --> V["Verify"]
V --> H["Human"]The learned Sentinel is advisory. Deterministic controls decide when evidence is sufficient, exact WNTR/EPANET simulation is required before a response can become VERIFIED, and approval remains a separate human event.
Final locked evaluation¶
-
73.3%
Nominal locked-final Top-1 localization
-
55.2%
All locked-final stress Top-1 localization
-
88.6%
Conformal coverage across applicable all-stress locked-final incidents
-
0 / 15
Hard safety counters violated
The final evidence run completed 125 / 125 incidents across the locked-final and locked-topology populations. Novel-topology predictive metrics are descriptive only: calibration was inapplicable there and the actionable rate was 0.0%.
Read the scientific evidence →
What is actually novel?¶
Prediction does not equal permission to act.
HydroSwarm's contribution is the governed integration of localization evidence, explicit calibration applicability, deterministic evidence/OOD gating, deterministic sampling and planning, exact hydraulic verification, and a separate human approval boundary. A learned output alone cannot authorize a response.
Frozen submission¶
This portal is generated from the immutable hackathon source snapshot 4bbf6fa3ff9f. The repository remains the canonical source of truth; this site is a judge-friendly presentation layer.