EVALUATION_AND_RESULTS.md

Evaluation and Results

Publication posture

This page uses a conservative pre-filing posture:

• report selected V40 results only
• avoid releasing raw arrays, raw examples, or full private evaluation internals
• show enough quantitative evidence to establish technical credibility

V40 result summary

Stage 1 relevance filtering

Best reported Stage 1 row:

• epoch: 2
• macro F1: 0.990
• relevant F1: 0.985
• irrelevant F1: 0.995

Interpretation:

• Stage 1 performed at a very high level on its binary gating task.
• That supports the broader architectural choice to screen inputs before deeper expression/state classification.

Stage 2 calibrated classification

Reported calibrated V40 snapshot:

• epoch marker: 5.985
• macro F1: 0.902

Selected calibrated examples:

• neutral: 0.978
• surprise: 0.976
• happiness: 0.973
• neutral_speech: 0.951

Classes that remained materially harder:

• contempt: 0.793
• disgust: 0.792
• sadness: 0.853
• speech_action: 0.882

What these numbers mean

The results support several careful public claims:

• the project had a real staged CV/ML core, not just a concept
• the relevance filter was strong enough to justify the two-stage architecture
• downstream performance was strong for several classes while still leaving a meaningful hard tail of difficult cases
• curation and review-routing remained justified because not all classes were equally easy

Scope of the published results

This page focuses on the V40 result surfaces that best explain the system: the strength of Stage 1 gating, the calibrated Stage 2 snapshot, and the class-specific tail that continued to drive curation and review work.

Accordingly, the repository centers the reader-facing summary rather than the full private evaluation package. The published view emphasizes the most informative aggregate result surfaces while leaving out raw arrays, raw confusion-case image references, private audit artifacts, and model weights.

How to read these results

The most important takeaway is not a single headline number. The V40 snapshot is more useful when read as a system result:

• Stage 1 was strong enough to justify the staged architecture.
• Stage 2 showed solid calibrated downstream performance across much of the target state space.
• The remaining hard tail still mattered, which is why curation, hard-negative review, and routing continued to be part of the engineering story.

Evidence basis for this page

The figures and summaries on this page correspond to the V40 result surfaces represented in this repository’s documentation and examples, especially docs/MODEL_AND_TRAINING.md, examples/plot_v40_results.py, and the summary asset assets/v40_results_overview.png.