Technical brief¶
A compact technical review of the final HydroSwarm system. For exact hashes and authority, Final system is canonical.
System in one paragraph¶
HydroSwarm is a local hybrid decision-support pipeline for simulated drinking-water contamination incidents. A classical hydraulic/signature branch and a learned HydroCore-v5 Sentinel produce source evidence; frozen split-conformal calibration is applied when valid; deterministic OOD/evidence control decides whether the workflow may continue; deterministic Scout and plan-generation logic choose evidence/action candidates; WNTR/EPANET is required to verify response plans; and a distinct human event is required for approval. The learned model never owns verification, approval, or actuation authority.
Frozen V5 identity¶
Finalist: HydroCore-v5 M10 frozen release, small, 4,182,612 parameters, seed 20260814.
- checkpoint:
de2b3f56243a1933d1d7c5957cd74a29fade119f7d104ce7f1500b3dd7b6d2a5 - release manifest:
f3fb08642738128f020c50e20e6b68c417bf80703f7ef6bc8f42db2aa41f8d34 - calibration file:
8f77f06b72316455e1f8040dbeb5907503e4eb623dd527d9ea809a56e96c046d - calibration artifact:
f2503e856c467eb38c6c7f6dbde679527c1921925941ec52809bd6e8e6dd16dd - alpha/grouping:
0.1,B_DEPTH_AWARE - serving:
V5PipelineFactory(resolve_v5_bundle_dir())
Learned versus deterministic authority¶
The model architecture includes optional specialist/control heads, but final valid training scope is sentinel only. Runtime learned outputs are exactly source_node, event_presence, event_cause, evidence_sufficiency, and relative_strength.
Operational controls are deterministic:
- OOD:
OODDetector - sampling:
rank_sample_locations - planning:
generate_response_plans - physical verification: WNTR/EPANET
- approval: human operator
- autonomous actuation: none
This separation is enforced by the release loader and was rechecked by M11.6 safety counters.
Training and model selection¶
The selected S model used AGE_FIX_ONLY plus exact 1,350-step interleaved multi-topology training over golden-reference, branched-loop, and loop-grid. M9 tested larger capacity; the ~13.9M-parameter M model did not meet the predeclared meaningful unseen-topology gain, so S remained selected.
A frozen caveat: the M9.6 training record used a fixed unobserved-age sentinel, while M10.4-tested serving retained incident_elapsed. This known train/serve feature-semantics deviation was not changed after lock.
Final evidence¶
M11.6 opened the final locked populations exactly once after authorization:
| Population | n | Top-1 | Top-3 | Coverage/applicability | Actionable |
|---|---|---|---|---|---|
| nominal final | 15 | 73.3% | 86.7% | 93.3% coverage | 80.0% |
| all locked-final | 105 | 55.2% | 76.2% | 88.6% coverage | 61.0% |
| novel topology | 20 | 55.0% | 70.0% | calibration inapplicable | 0.0% |
Overall M11.6: 125/125 complete, locked-final PASS, locked-topology PASS, 15/15 hard safety counters zero, one authorized opening, no rerun, no post-lock tuning.
Topology predictive metrics are descriptive/non-gating. The operational topology result is fail-closed: 0% calibrated, 0% actionable, 0% human-approved.
Runtime¶
FastAPI + SQLite + bounded local jobs expose typed local APIs and durable audit history. The current source app serves V5. hydroswarm self-test --strict checks the V5 bundle and a bounded real WNTR run.
For V5 Docker, either run the published release (docker compose -f docker-compose.release.yml up, ghcr.io/insightlabs38-pixel/hydroswarm:v0.2.1) or build the current checkout; both resolve to the same frozen V5 identity.
Reproducibility¶
Do not rerun M11.6. Verify the immutable model/calibration/release hashes plus the M11.6 opened/metrics/gate/safety/closure chain. Ordinary source checks, self-test, and non-locked tests remain reproducible.
See Reproducibility and Claims and evidence.
Limits¶
All data are synthetic. Stress performance is materially lower than nominal performance. Novel-topology prediction is not calibrated. WNTR/EPANET inherits network/model assumptions. No field/utility accuracy, chemistry determination, public-health safety, or autonomous-control claim is supported.