Demonstrated here
- Frozen, reproducible historical replay
- Versioned model and data lineage
- Station health and inspectable artifacts
Historical replay · bounded capability evidence
A governed path from research algorithm to 24/7 seismic operations
NEIC's scientists and researchers own the science. The engineering problem is the ecosystem around it: streaming ingest, detector execution, monitoring, archiving, and distribution built so AI-driven and classical algorithms run dependably on continuous data, with verification built into every delivery increment. This historical replay demonstrates the engineering pattern — reproducible execution, data provenance, model lineage, station-health monitoring, inspectable evidence. Continuous streaming, NEIC integration, and operational readiness are not demonstrated here.
Proposed delivery
- Streaming ingest and detector execution on NEIC-defined interfaces
- Monitoring, archiving, and distribution around Government-selected algorithms
- Operator review and Government-controlled rollout
Not yet demonstrated
- Continuous streaming
- NEIC integration or downstream effects
- Operational readiness or mission outcomes
Loading scored results…
Loading scored artifacts…
Illustrative operating and qualification path
Government controlledIllustrative; subject to Government architecture and operator approval.
Potential operating pattern
- Government-controlled continuous inputs
- quality and interface checks
- Government-selected classical and ML detectors in parallel
- existing association and analysis path
- analyst-reviewed products
Delivery path
- Interface discovery NEIC-defined
- historical replay harness Demonstrated on this site
- streaming ingest and detector execution Verification built in
- operator-reviewed rollout Humans in the loop
- Government-controlled promotion or fallback USGS decides
Stage 2 runs live on this site. Verification happens inside each delivery increment — not as a separate evaluation phase — and the Government holds every gate.
Evidence mapped to the notice's five tasks
| Notice task | Demonstrated here | Proposed next artifact | Current gap |
|---|---|---|---|
| Develop and Maintain Real-Time Seismic Data Streaming and Detection Systems | Historical waveform ingest and detector execution | Mirrored continuous-input adapter and recovery test | No continuous stream or event association demonstrated |
| Modernize and Optimize Detection Algorithms | Classical and ML methods run through one frozen evaluation harness | Government-selected incumbent and candidate acceptance tests | No incumbent replacement or mission improvement demonstrated |
| Integrate and Advance AI/ML Capabilities | Versioned worked examples with weights, thresholds, and lineage | Model execution path on NEIC interfaces with approval gates and controlled promotion evidence | No Government-system integration demonstrated |
| Build and Integrate Monitoring, Archiving, and Distribution Systems | Run health, checksummed artifacts, and read-only API delivery | Government-interface integration and operational health measures | No downstream product distribution demonstrated |
| Support Scientific Research and Innovation | Reproducible, inspectable bounded experiment | Repeatable research-to-operations qualification process | No representative global or sustained evaluation demonstrated |
Acceptance measures for delivered work
Targets established with NEIC before scoring- TimelinessIngest-to-result and result-to-product timing
- AccuracyPick-, event-, location-, magnitude-, and product-level accuracy
- ReliabilityAvailability, restart recovery, backlog, and degraded-mode behavior
- Operator impactAnalyst workload, overrides, and model disagreement
- InterfacesInterface compatibility and downstream effects
- AuditabilityData, model, configuration, code, and approval lineage
- Global domainRepresentative global source distances, phases, magnitudes, station density, network conditions, and noise regimes
Measures and target values would be established with NEIC before scoring; this replay establishes no operational threshold.
Bounded replay evidence
Frozen protocol · ±1.0 s vs cataloged-arrival comparison setDeterministically selected cases
30-second close-ups · successes and failure modesCases are selected by committed rules from the frozen match ledger and raw waveform cache—not hand-picked after viewing results. They illustrate detector behavior and data quality, not associated events or operational decisions.
Loading case artifacts…
Model registry
Versions, weights, thresholdsFederated data sources
Provenance & coverageTechnical waveform drill-down
Full-day overview · not an interpreted case studyOperations
This run, measuredModel lifecycle
Method, reproduce & boundaries
Every number on this console comes from one scored run under a version-frozen protocol: matching is maximum-cardinality, minimum-total-residual one-to-one assignment within a disclosed ±1.0 s tolerance; the headline benchmark is phase-agnostic for all detectors, with per-phase stratification reported for the ML pickers (phase labels joined from ComCat origin arrivals); STA/LTA parameters are fixed conventional values chosen independently of the scored window; picker weights are SeisBench stead rather than the in-domain scedc set — STEAD's compilation window ends in August 2018, so the July 2019 evaluation day cannot appear in training, while the SCEDC corpus (2000–2020) contains the Ridgecrest sequence itself. The scored window contains the M7.1 mainshock origin and roughly 20 hours of its aftershock sequence — a stress regime, not a typical operating day — and the reference set is restricted by protocol to arrivals from M≥3.5 events. Headline recall counts P and S arrivals separately, which structurally favors phase-resolving pickers over the single-onset vertical-component baseline; the comparison is disclosed rather than equalized. Unmatched picks remain unclassified because ComCat is an aggregated comparison set, not a complete label set — the sequence produced thousands of cataloged earthquakes below the protocol floor the same day. This demonstration makes no event-level, completeness, or operational-timing claim. Replay compute figures describe offline batch execution on one laptop CPU. Data are public (EarthScope FDSN waveforms, USGS ComCat arrivals); PhaseNet, EQTransformer, and SeisBench are open-source work credited to their authors. Morton Analytics built the ingest, frozen-protocol evaluation harness, artifact lineage, serving, and observability around them.
Protocol · Reproducibility manifest · Coverage report · Metrics JSON
Application logs record timestamp, path, and coarse user-agent class only — no cookies, no IP addresses; platform-level logs are retained per DigitalOcean defaults.
Morton Analytics, LLC — SBA-certified 8(a) small business · UEI U9QHJDUKGUM5 · CAGE 9VN85
[email protected] · 406-546-0902 · morton-analytics.com · Market-research demonstration — not an offer or production-deployment commitment.