Simplify Stage3 chunking to interleave-only and add eval diagnostics

Add eval audit tools, caption-evident set, and logging

Record non-fatal pipeline issues in eval JSONL outputs

Clarify stage3_debug rewrite-first behavior in quickstart docs

Remove outdated branch list from project summary

Add mandatory workspace bootstrap steps for Git alignment

Add comprehensive project documentation for session continuity

Fix eval_categorized.py to work with eval_pipeline.py output

Remove binary PNG files (use Hugging Face XET storage instead)

Add eval results for debugging

Add ranking metrics infrastructure to eval pipeline

Add per-category evaluation metrics script

Fix constraints: body_type, species, gender are multi-select

Add tag categorization pipeline for e621 checklist

Improve topless/bottomless definitions to prevent confusion

Fix Windows encoding issues in diagnostic script

Add diagnostic script for structural clothing inference

Redesign structural inference as group-based system with wiki data

extracted tag groups and wiki definitions

Add per-tag evidence tracking and wiki extraction script

Add compact eval analysis script for new output format

eval with structural inference + strong_implied + implications 50 images

eval with structural inference + strong_implied + implications

Rewrite structural inference prompt for better Llama 3.1 8B performance

Add structural tag inference (Stage 3s) and compact eval output

Default min_why to strong_implied; add retrieval gap analysis script

eval with expanded GT and leaf metrics

Normalize GT annotations: expand implications, exclude non-evaluable tags

eval with min_why=explicit + implications

Add tag implication expansion (fox→canine→canid→mammal)

eval results with min_why=explicit bug fix

Remove data/eval_results/ from .gitignore so eval results are tracked

Fix min_why not passed to workers in parallel eval mode

Add latest eval results

Add --min-why threshold to filter Stage 3 selections by confidence level

adding tag implications file

Add diagnostic eval metrics, why-distribution tracking, and generic character filter

Add n=10 eval results for analysis

Add parallel processing to eval pipeline with ThreadPoolExecutor

Add independent character tag metrics to eval pipeline

Improve eval harness: shuffle samples, always write results

Add end-to-end evaluation harness for pipeline metrics

Expand alias filter tests with real CSV data and pipeline tests

Add alias-based character tag filtering for Stage 3

initial commit