How runtime-weekly/runtime-tutorials works
RUNTIME tutorials is a collection of four independent, runnable guides—not a shared application. A reader chooses a guide, supplies or inspects synthetic examples, and gets classifications, proposed tool calls, or recomputed research summaries. Laya adds an optional local drafting step; the research guides compare saved model decisions and resource measurements, with Tev1 also documenting optional local inference. Sampled code supports the Laya and Needle runtime links; research-guide wiring is based on their documented workflows and sampled verification scripts, so uncertain internal call paths are omitted.
Main entry points: analyze.py Recompute comparison · demo.py Propose tool calls · verify.py Verify findings · examples.py Classify and gate
Components 22
Laya classification
- Classify and gate
laya/examples.py - Custom categories
laya/custom_notes.py - Fetch checkpoint
laya/download_model.py - Laya checkpoint
Needle tool proposals
- Propose tool calls
needle/demo.py - Tools and queries
needle/cases.json - Saved demo results
needle/example-results.json
Tev1 research
- Verify findings
tev1/verify.py - Frozen cases
tev1/data/cases.json - Recorded decisions
tev1/data/results.json - Environment evidence
tev1/data/tested-environment.json - Fetch checkpoints
tev1/download.py - Optional inference
tev1/infer.py
Kev Laya research
- Recompute comparison
kev-laya/analyze.py - Frozen cases
kev-laya/cases.json - Saved decisions
kev-laya/results.json - Resource measurements
kev-laya/measurements.json - Expected summary
kev-laya/expected-summary.json
Other components
- Guide reader
- Local LLM server
- Needle runtime
- Hugging Face
Connections 25
- Guide reader → Classify and gate: runs demo
- Guide reader → Custom categories: runs example
- Classify and gate → Laya checkpoint: loads
- Custom categories → Laya checkpoint: loads
- Fetch checkpoint → Hugging Face: downloads
- Fetch checkpoint → Laya checkpoint: writes checkpoint
- Classify and gate → Local LLM server: requests draft
- Guide reader → Propose tool calls: runs or inspects
- Propose tool calls → Tools and queries: reads
- Propose tool calls → Needle runtime: calls
- Propose tool calls → Saved demo results: reads saved output
- Guide reader → Verify findings: verifies findings
- Verify findings → Frozen cases: reads
- Verify findings → Recorded decisions: reads
- Verify findings → Environment evidence: reads
- Guide reader → Fetch checkpoints: requests checkpoints
- Fetch checkpoints → Hugging Face: downloads
- Guide reader → Optional inference: runs optional rerun
- Optional inference → Frozen cases: reads
- Optional inference → Hugging Face: uses downloaded assets
- Guide reader → Recompute comparison: recomputes findings
- Recompute comparison → Frozen cases: reads
- Recompute comparison → Saved decisions: reads
- Recompute comparison → Resource measurements: reads
- Recompute comparison → Expected summary: checks optional