Skip to content

MODEL_REGISTRY.md

Latest commit

 

History

History
68 lines (48 loc) · 2.1 KB

MODEL_REGISTRY.md

File metadata and controls

68 lines (48 loc) · 2.1 KB
Raw

Model Registry

Stream now exposes a public model and method registry at:

  • GET /api/system/model-registry
  • data/system/model-registry.json

The registry separates trained models, statistical methods, deterministic probes, and compiled kernels. That distinction is important. A useful repo does not need to call every calculation a model.

Trained models

The music lane trains and evaluates a view-predictability ensemble over the committed public corpus.

  • Implementation: music_pipeline.predictability_analysis
  • Estimators: RandomForestRegressor, GradientBoostingRegressor
  • Target: log1p(view_count)
  • Validation: 5-fold cross-validation with out-of-fold ensemble prediction
  • Leakage guard: virality_coefficient is excluded because it is derived from views

The music lane also runs KMeans clustering to describe release archetypes. Those clusters are descriptive. They are not identity labels and should not be used as artist judgments.

Statistical methods

The registry also names the non-trained analytical methods:

  • power-law tail fit over view counts
  • attention inequality metrics
  • tag co-occurrence network analysis
  • controlled decision-lab comparisons

These methods are there to make the shape of attention easier to inspect. They do not turn metadata into cultural certainty.

Deterministic probes

The runtime drift and local event-window surfaces are deterministic checks. They are useful for local inspection, API testing, and event-contract design. They do not claim hidden access to model internals.

Compiled kernels

loopchii-wasm-core contains Rust metric primitives for entropy, concentration, weighted means, top-share calculations, and chi-square goodness-of-fit.

Run:

cargo test --manifest-path loopchii-wasm-core/Cargo.toml

Verification path

python3 run_analysis.py --output-dir exports
python3 -m stream_backend.cli.doctor
python3 -m pytest -q tests
cargo test --manifest-path loopchii-wasm-core/Cargo.toml

If a future model is added, it should appear in the registry with its implementation path, validation method, leakage guard, and claim boundary.