These are dataset-backed workloads for Plexus. 02_incident_commander is the actor-model reference application. The other five currently serve as real-data integration/acceptance workloads; they exercise Plexus and real TypeSafe calls, but they should not be treated as evidence that an actor architecture is superior to an ordinary centralized async pipeline.
The reference application has a stricter bar: entities persist across many evidence epochs, raw evidence arrives while the computation is running, actors message and invalidate one another directly, semantic results change future topology/lifecycle, branch credit affects local investigation, pruning can cancel live subtrees, and termination emerges from telemetry EOF plus quiescence rather than a coordinator waiting for a fixed count.
All six use recognized external datasets or APIs. Downloaded data is not committed to Plexus; fetchers write beneath .plexus-data/ by default and .gitignore protects that directory. Pass --data-dir to keep datasets elsewhere.
Prerequisites
- Elixir/OTP supported by Plexus and
mix deps.getcompleted. TYPESAFE_API_KEYfor the run scripts.TYPESAFE_BASE_URLis optional.TYPESAFE_MODELis accepted as an example-specific override; the SDK's standardTYPESAFE_DEFAULT_MODELenvironment variable is also honored.- Internet access for fetch scripts.
curlis preferred for large downloads; the helper falls back to Erlang HTTP for small files.git-lfsand7z/7zzare required only for the GAIA example.
The examples use real TypeSafe/Jev calls. There is no fixture-mode fallback in examples/; deterministic fixtures live only in test/fixtures/examples/ for offline parser/release tests.
Catalog
| Example | Standard data | Actor computation | TypeSafe/Jev density |
|---|---|---|---|
00_issue_swarm | SWE-bench Verified | issue actors route into repository/fix-shape populations; gold patches score a fixed prediction vocabulary post hoc | one request per issue |
01_city_signal_tracker | NYC 311 | report actors feed spatiotemporal incident clusters | one request per selected dense cluster |
02_incident_commander | GAIA / MicroSS | living system twin: raw chronological telemetry drives persistent services, repeated hypothesis revisions, peer conflict, invalidation, branch credit and live pruning | event-driven; a hypothesis may issue many measurements over its lifetime |
03_dependency_upgrade_search | deps.dev v3 | resolved dependency graph diff + semantic risk measurements + pruned migration-order beam search | only changed dependency nodes |
04_research_evidence_graph | SciFact | claim/evidence graph with typed support/contradiction edges and gold labels | one request per claim/document pair |
05_alert_swarm | NOAA Storm Events | large dormant event population awakened by historical day events | only highest-impact state/day groups |
Typical workflow
Each directory has a fetch.exs and run.exs:
mix run examples/00_issue_swarm/fetch.exs
TYPESAFE_API_KEY=... mix run examples/00_issue_swarm/run.exs --limit 50
Fetch once, then rerun from the local cache. The run scripts expose workload controls rather than substituting toy datasets for the real source.
Workload scale and provider calls
Dataset size, live actor population, and semantic-call count are separate quantities. The GAIA reference intentionally allows a long-lived actor to perform many semantic measurements as new evidence arrives; scale is driven by the real event stream rather than by a loop whose only purpose is to manufacture requests.
Before a large run, inspect the example README and choose its call-driving controls. Every run exposes an explicit semantic-work ceiling, and TypeSafe telemetry reports the calls/tokens actually consumed. In TypeSafeSDK 0.4, distinct uncached inputs remain distinct provider requests; coalescing deduplicates identical work and controls concurrency.
Runtime-only experiments stay in experiments/
Plexus already has controlled characterization for actor birth, measurement/coalescing, calibration, and scheduling replay. In particular, experiments/node_birth.exs has retained measurements through 100,000 actors. The application examples do not duplicate that experiment merely to advertise an actor count.
UI boundary
These examples exercise Plexus directly. They intentionally contain no Phoenix, LiveView, web UI, or visualization layer. Interactive visualization belongs in the separate Plexus LiveView poncho repository.
See DATASETS.md for source/provenance details.
Live TypeSafe transport evidence
Every dataset-backed run creates its TypeSafe client through examples/support/runtime.exs.
The shared runtime installs a privacy-safe TypeSafe telemetry collector before semantic work begins.
At process exit each run prints a TYPESAFE LIVE TRANSPORT SUMMARY containing the selected endpoint,
requested model and transport, evaluate counts, confirmed HTTP response counts, HTTP status distribution,
returned model distribution, input/output token totals, retry counts, and provider request IDs.
A response counts as a confirmed HTTP response only when TypeSafeSDK telemetry contains both an HTTP status and a non-empty provider request ID. The collector never records API keys, semantic state, prompts, request bodies, response bodies, authorization headers, or other customer content.
Application-level semantic measurements and confirmed TypeSafe HTTP responses are deliberately reported
separately: the former demonstrates Plexus scheduling/accounting; the latter proves that the remote TypeSafe
service actually answered the semantic request.