Advanced Proof Pack

Use this route after the lightweight hosted demo or the local first proof. It surfaces stronger AGILAB capabilities that are already packaged but are too deep for the default Hugging Face first-run path.

The goal is not to make the first demo longer. The goal is to keep the newcomer path clean while giving evaluators a second pack that proves richer behavior: mission decisions, execution-model benchmarks, network simulations, service health, app-to-app contracts, connector contracts, and release evidence.

What belongs here

Proof route

What it proves

Where to start

mission_decision_project

Deterministic mission-data decision loop: ingest evidence, score routes, inject an event, re-plan, and export a decision bundle.

Select the built-in app, run ORCHESTRATE, then open view_data_io_decision.

execution_pandas_project

Cython worker speedup demo: a Pandas worker keeps dataframe I/O and reducer evidence in Python while the hot scoring loop runs as a typed contiguous float64 kernel.

Use Execution Playground.

execution_polars_project

Execution-model benchmarking for the comparable threaded dataframe worker path.

Use Execution Playground.

uav_queue_project / uav_relay_queue_project

Network-style experiment analysis: queue buildup, packet drops, routing policy changes, topology, trajectories, and generic network map views.

Select the built-in app, run it, then open view_scenario_cockpit, view_queue_resilience, view_relay_resilience, or view_maps_network.

inter_project_dag packaged preview

App-to-app artifact handoff without private infrastructure: one app produces an explicit artifact contract that another app can consume.

Run src/agilab/examples/inter_project_dag/preview_inter_project_dag.py.

mlflow_auto_tracking packaged preview

Optional experiment memory: AGILAB writes local evidence first, then logs params, metrics, and artifacts through MLflow when the backend is present.

Run src/agilab/examples/mlflow_auto_tracking/preview_mlflow_auto_tracking.py.

sqlite_connector_proof packaged preview

Local database evidence: create a deterministic SQLite database, run a parameterized read-only SQL query, export CSV, and record schema, query, result, and artifact hashes.

Run src/agilab/examples/sqlite_connector_proof/preview_sqlite_connector_proof.py and see Data Connectors.

resilience_failure_injection packaged preview

Resilience comparison: inject one relay degradation event, then compare fixed, replanned, search-based, and active-policy responses on the same scenario contract.

Run src/agilab/examples/resilience_failure_injection/preview_resilience_failure_injection.py.

train_then_serve packaged preview

Service handoff: freeze the trained-policy artifact path, IO contract, prediction sample, and health gate before a serving stack is started.

Run src/agilab/examples/train_then_serve/preview_train_then_serve.py.

service_mode packaged preview

Persistent worker lifecycle and health gates: start, status, health, and stop are presented as explicit operator actions.

Run src/agilab/examples/service_mode/preview_service_mode.py and see Service Mode.

Data connector reports

Connector catalogs, local/cloud/search endpoint shapes, health planning, and credential-safe resolution reports without requiring live secrets.

See Data Connectors.

Release proof

Public trust evidence: PyPI package, GitHub release, CI guardrails, docs-source integrity, and hosted-demo evidence in one page.

See Release Proof.

Optional SB3 industrial optimization examples

Advanced app-repository routes for active mesh optimization, MLflow auto-tracking, multi-app DAGs, resilience/failure injection, and train-then-serve policy contracts.

See Industrial Optimization Examples when sb3_trainer_project is installed.

How to demo it

Keep the story bounded. Do not switch apps randomly. Use one of these lanes:

Decision lane

mission_decision_project -> ORCHESTRATE -> ANALYSIS -> view_data_io_decision. Stop when the selected strategy, re-plan event, decision deltas, and exported artifacts are visible.

Performance lane

Execution Playground. Stop when the viewer understands that pool is AGILAB’s external local fan-out, Pandas can benefit from process fan-out, and Polars may not because it already owns native internal parallelism.

Network lane

uav_relay_queue_project -> ORCHESTRATE -> ANALYSIS -> view_scenario_cockpit -> view_relay_resilience plus view_maps_network. Stop when the baseline/candidate evidence bundle, queue buildup, relay choice, drops, topology, and trajectories are visible from the same run artifacts.

The source-checkout proof can be regenerated without the UI by running uav_queue_project twice, once with the shortest-path policy and once with the queue-aware policy, then packaging the result through the same Scenario Cockpit evidence helpers:

uv --preview-features extra-build-dependencies run python tools/scenario_cockpit_evidence.py --output-dir build/scenario-cockpit-proof --clean

Maintainers refresh the checked-in public sample only when intentionally updating proof evidence:

uv --preview-features extra-build-dependencies run python tools/scenario_cockpit_evidence.py --output-dir build/scenario-cockpit-proof --clean --write-doc-sample docs/source/data/scenario_cockpit_uav_queue_sample.json

The checked-in sample records a promotable queue-aware candidate against the shortest-path baseline, 24 hashed artifacts, and zero missing peer artifacts. It is source-checkout evidence for the proof route; it is not a broader simulator certification.

Operator lane

service_mode preview and Service Health JSON Schema. Stop when the lifecycle and health thresholds are explicit. Do not claim production service certification from the preview.

Tracking lane

mlflow_auto_tracking preview. Stop when the local evidence bundle and the tracking status show the same params, metrics, and artifact path. If MLflow is not installed, the expected status is skipped with an installation hint.

Resilience lane

resilience_failure_injection preview. Stop when the fixed route degrades, the post-failure route ranking is explicit, and the adaptive response wins without implying a certified MARL benchmark.

Train-then-serve lane

train_then_serve preview. Stop when the service contract, prediction sample, and health payload are written without implying that AGILAB is a production serving platform.

Industrial optimization lane

Industrial Optimization Examples. Stop when the reviewer can see the active-mesh or queue-routing artifact contract, optional MLflow tracking status, resilience comparison, or service contract without confusing those examples with the default hosted first proof.

What not to claim

  • Do not present the Advanced Proof Pack as the default hosted demo.

  • Do not claim live cloud validation unless the compatibility matrix says that route is validated.

  • Do not claim Cython end-to-end speedup from the kernel-only benchmark; the full run still includes IO, grouping, artifacts, startup, and orchestration.

  • Do not claim Polars should always benefit from AGILAB pool mode. Polars already manages native internal parallelism.

  • Do not claim the UAV examples are full research simulators. They are compact public scenarios shaped to prove AGILAB workflow value.

  • Do not claim the Active Mesh Optimization example is full decentralized MARL. It is a compact centralized-policy route that makes the agent/action contract visible for later hardening.

  • Do not present mlflow_auto_tracking as an AGILAB model registry. It is an adapter proof that keeps MLflow as the tracking and registry system when used.