Classification Pipeline

Classification Pipeline

Every source classification decision in MarieLib is inspectable. This page explains how the pipeline works and how to debug classification results.

---

🔍 Pipeline Overview

MarieLib classifies edible items through a multi-stage pipeline. Each stage records a trace step so you can see exactly why an item resolved the way it did.

flowchart TD
    discovery[ITEM_DISCOVERY]
    tagLookup[VALUE_TAG_LOOKUP]
    cache[RESOLVER_CACHE]
    community[COMMUNITY_TAG_SIGNAL]
    keyword[KEYWORD_SUFFIX_SCORING]
    recipe[RECIPE_LOOKUP]
    blend[TAG_RUNTIME_BLEND]
    winner[WINNER_SELECTION]
    confidence[CONFIDENCE]
    discovery --> tagLookup --> cache
    cache --> community --> keyword --> recipe
    recipe --> blend --> winner --> confidence

Loading

Trace Step IDs

Step	What it checks
ITEM_DISCOVERY	Item is edible / has food properties
VALUE_TAG_LOOKUP	Registered value tags on the item
EXTERNAL_CLASSIFICATION	Compat or API-registered mappings
RESOLVER_CACHE	Cached prior resolution
COMMUNITY_TAG_SIGNAL	c:foods/* tag matches
KEYWORD_SUFFIX_SCORING	Item name keyword analysis
RECIPE_LOOKUP	Recipe ingredient inheritance
INGREDIENT_RESOLUTION	Per-ingredient classification
NAMESPACE_PEER	Same-namespace item clustering
PRIMARY_RECIPE_MERGE	Multi-ingredient recipe merge
TAG_RUNTIME_BLEND	Tag data blended with runtime inference
SIGNAL_AGGREGATION	Multiple signals combined
WINNER_SELECTION	Highest-scoring value key chosen
CONFIDENCE	Spread-based confidence check
HARD_FALLBACK	Last-resort classification
APPLY_GATE	Final apply decision

---

⚖️ Signal Priority

When multiple signals compete, MarieLib uses weighted precedence:

Signal	Priority
Value tags (modid:values/* or consumer-specific tags)	Highest
Community tags (c:foods/*)	5×
Namespace grouping	4×
Name suffix match	3×
Keyword match	2×
Recipe inheritance	1×

Tag vs runtime blend: When both tag data and runtime inference produce results, tag data typically wins for the dominant value. Runtime results may supplement secondary values.

---

🧪 Inspecting Classifications

Hold a classified item and run:

/<modId> debug held

For Nourished: /nourished debug held

This writes a full classification trace to:

config/<modId>/debug/trace_dump.txt

The trace shows every pipeline stage, scores considered, blend precedence decisions, and the final value assignment.

---

📊 Scanner Output

The scanner analyzes all edible items and can write:

• Tag recommendations for data/<modId>/tags/item/values/<key>.json
• Multi-value reports for items with secondary classifications
• Overlap matrices showing co-occurrence between value pairs
• Ambiguous item lists for manual review

Run multi-value analysis with:

/<modId> scan_analysis

Output lands in config/<modId>/scanner_analysis/.

---

💡 Tips

• Items with c:foods/* tags classify reliably — add them if you're a mod author.
• Recipe inheritance means complex foods inherit from ingredients (e.g. stew → multiple values).
• Low-confidence items appear in config/<modId>/unassigned_sources.txt via /get_unassigned.

---

📚 Related Pages

• Diagnostics & Tracing
• Community Tags
• Compatibility Guide
• Commands Reference