Glossary¶
Terms used across the final V5 documentation, API, and operator experience.
Model and evidence¶
HydroCore-v5 — the final frozen learned Sentinel, 4,182,612-parameter small variant. Exact identity: Final system.
Sentinel — the only valid trained task family in the final release. Runtime learned outputs are source_node, event_presence, event_cause, evidence_sufficiency, and relative_strength.
Architecture head — a neural output structure present in the model definition/configuration. Presence does not prove the head received valid supervision.
Runtime-enabled output — a learned output explicitly allowed by the V5 release/finalist freeze to be surfaced at runtime.
Operational authority — the component permitted to make a governed workflow decision. In the final system OOD, sample ranking, and plan generation remain deterministic; plan verification is simulator-based; approval is human.
Classical signature — simulator-derived source-response evidence used by the non-neural localization branch.
Fusion — combination of classical and learned source beliefs under the frozen fuse_source_probabilities-v1 configuration.
Authority labels¶
ADVISORY — can influence evidence/prediction, cannot authorize an operational state.
CALIBRATED ADVISORY — advisory output with applicable conformal semantics.
DETERMINISTIC — algorithmic/rule control with operational workflow authority.
SIMULATOR_VERIFIED — response candidate completed WNTR/EPANET verification and passed configured modeled constraints.
HUMAN_APPROVED — a human explicitly recorded approval. This is not infrastructure actuation.
Autonomous actuation — software directly commanding field infrastructure. HydroSwarm has none.
Calibration and OOD¶
Conformal candidate set — a set of candidate source nodes produced under a split-conformal artifact. Its coverage guarantee is marginal over an applicable population, not a per-incident confidence.
Coverage — fraction of applicable evaluated cases whose conformal set contains the true source. Always read with population/denominator and calibration applicability.
calibrated_rate — fraction of cases where calibration was applicable. The final novel-topology split has calibrated_rate=0, so its raw inclusion statistic is not calibrated coverage.
OOD (out of distribution) — evidence that an incident/topology lies outside conditions authorized by the frozen calibration/runtime policy.
Fail closed — withhold or reduce action authority when required evidence/assets/applicability are missing rather than silently substituting a confident result.
Sampling and planning¶
Scout — the sampling stage. In the final system the authoritative selector is deterministic rank_sample_locations; learned Scout heads are non-authoritative.
Strategist — response-planning model/head concept. Learned Strategist outputs are non-authoritative in the final system; authoritative candidate generation is deterministic generate_response_plans.
Sample budget — governed bound on evidence-collection recommendations during an incident.
VERIFIED — a plan completed exact modeled verification and passed configured constraints. Not synonymous with real-world safe or approved.
REJECTED — a candidate failed verification/completeness/constraint checks.
STALE verification — verification computed under an earlier evidence context; cannot be approved after the context changed.
APPROVED — a separate human approval event. No automatic execution follows.
Evaluation governance¶
Development evidence — data/results available during architecture/capability work before final freeze.
Finalist freeze — M11.2 artifact fixing checkpoint, calibration, runtime outputs, authority, and limitations before final test.
Locked-final test — 105-case final applicable stress population, opened once in M11.6.
Locked-topology test — 20-case final novel-topology/fail-closed population, opened in the same one-time M11.6 execution.
Descriptive/non-gating metric — reported scientific measurement that did not determine the hard pass/fail gate. M11.6 explicitly treats Top-1/Top-3/MRR and novel-topology predictive metrics this way.
Hard safety counter — preregistered invariant-violation count. All 15 M11.6 counters were zero.
Opened record — immutable M11.6 record marking that the one-time lock was consumed before evaluation results were produced.
Post-lock tuning — model/system tuning after observing locked results. Final governance records false.
Experience/provenance modes¶
LIVE — real current backend incident and current computation.
REFERENCE INCIDENT — deterministic checksummed replay of a frozen WNTR-backed workflow; useful for demonstrating product flow, not final V5 benchmark evidence.
DEMO_FALLBACK / ILLUSTRATIVE DEMO — hand-authored fallback content, explicitly not real current scientific computation.
Replay — inspection of a recorded incident/audit history; not a new model run.
Historical terms¶
HydroCore-v4 — prior frozen generation retained for provenance. Its metrics and lock status are historical, not current V5 claims.
HydroCore-S/M/L historical benchmark — earlier architecture-generation research retained under historical reports; not the final V5 evidence.