◇FeatureGraph

Roadmap

Connect scientific detection to explicit object representation.

The completed CLaP integration establishes a tested boundary between an external state detector and FeatureGraph object materialization. The released BIDMC beta remains the citable record of the earlier native construction.

Where it stands

What is built, what is in development, and what it does not claim.

Built

  • The compiler layers P, D, O and M, deterministic and outside any language model.
  • The featuregraph-core package.
  • Runs on breathing recordings (BIDMC, CapnoBase), a wearable study from PhysioNet and a Tennessee Eastman reactor signal.
  • A public research record.
  • Two of the five agent roles, the Tester and the Recorder.

In development

  • The scientific interpretation layer and its agent loop.
  • Natural-language access beyond the mock-table demo.
  • Support for time series that do not oscillate.

Not claimed

  • FeatureGraph provides structure and an audit trail. It does not provide clinical validation or final scientific interpretation.

Completed interoperability pass

CLaP state-occurrence study.

Complete

External detection

CLaP supplies the scientific state labels and change points without FeatureGraph tuning.

Complete

Public materializer

from_state_sequence constructs FeatureObjects, boundary status, measurements, provenance, and relations.

Complete

Exact reconstruction

Nine objects reconstruct all 20,700 source labels exactly; eleven declared study checks pass.

Frozen beta research pass

BIDMC respiration representation study.

Complete

Independent reference analysis

A separate analysis of raw BIDMC respiration provided candidate breaths for an object-level comparison.

Complete

Object-level comparison

Individual candidate breaths were matched, exposing boundary agreement and unmatched objects instead of relying on aggregate means.

Complete

Self-implemented detector

The beta records explicit lag, threshold, state-gap, peak-boundary, and incomplete-object behavior without native SciPy peak finding.

Complete

Multi-subject transfer

Fixed and MAD-scaled rules were run across BIDMC subjects. The negative transfer result is retained rather than hidden by further tuning.

Complete

Accumulation demonstration

Wave-relative area was represented as a dependent object with parent identity, completeness, timing, rates, and centroid.

Complete

Research manuscript

A working paper documents the successful preservation result, definition mismatches, and detector failure.

Current work

Prepare the beta record for public preprint.

Complete

Release and archive the beta record

FeatureGraph v0.1.0b1, its generated evidence, GitHub checks, and Zenodo archive are public and immutable.

Priority 02

Prepare the manuscript for public preprint

Harmonize measurement contracts, finalize figures and citations, and release the study without claiming biomedical validation.

Priority 03

Maintain the compatible beta line

Use beta/v0.1.x for compatible corrections and evidence updates while preserving the released v0.1.0b1 tag.

Next paper

Transition-only FeatureGraph.

Next

No oscillation assumption

Construct explicit rising, falling, and inactive transitions without deciding that the source signal contains cycles.

Next

Performance evaluation

Measure correctness, scaling, memory, and computation costs for the simpler representation.

Later

Higher-order composition

Let researchers or downstream software decide when transitions should compose into oscillations or other behavioral objects.