Lore Intelligence
A small Legendry — maybe twenty entries, all linked, all filled in — is easy to keep in your head. You know what you’ve written, you know what connects to what, and you know which characters need more work. A large Legendry is different. By the time you have three hundred entries across ten categories, with hundreds of relationships and half-filled sections scattered across entries you wrote six months ago and half-forgot, you can’t hold the shape of your world in your head anymore. You literally cannot see what’s missing, what contradicts itself, or where the hidden structural problems are. Not because you’re lazy — because the world is too big to eyeball. Lore Intelligence is Ishvana’s machine learning layer that finds the patterns and problems you can’t see yourself. Anomalies, contradictions, thematic clusters, underdeveloped entries, structural gaps, entity centrality, community detection — every kind of analysis that only becomes possible when you run actual math over your world data. It’s the tool you reach for when you need structural insight at scale.
Health overview
Section titled “Health overview”The top of the Lore Intelligence panel shows an aggregate health score from 0 to 100, computed from six dimensions:
| Dimension | What it measures |
|---|---|
| Consistency | How free the database is from contradictions and validation issues |
| Depth | How thoroughly entries are filled out |
| Breadth | How well coverage spans across categories and entry types |
| Interconnectedness | How well entries are linked to each other |
| Maturity | Mean maturity score across all entries |
| Overall | Weighted combination of all dimensions |
The current health score is color-coded in the top bar: green (70+), amber (40-69), red (below 40). Health history is tracked over time — the last 20 training snapshots are preserved — so you can see how your world’s health is trending. A slowly rising curve is the healthy trajectory. A flat curve or a declining one is a signal that your recent work isn’t improving the Legendry’s structural health.
Dashboard sections
Section titled “Dashboard sections”The dashboard is organized into four groups with ten sections total. Each section answers a different structural question about your world.
Overview
Section titled “Overview”- Health — the aggregate health score and trend history. The landing page for the module.
Three sections for finding what you didn’t know was in your world:
- Clusters. Groups of similar lore entries found using text analysis. Each cluster gets a label derived from its key terms (the algorithm picks the most distinctive words across the cluster), a depth level, an entry count, and a dominant entry type. Drill into any cluster to see its members. Clustering is what reveals thematic groupings you didn’t plan — “these eight entries are all about political succession even though they live in different categories.”
- Similarity. Pairs of entries that are semantically close, split into three types:
- Suggested cross-links — high similarity but no existing link. The system is suggesting you probably should link these.
- Near-duplicates — very high similarity that might indicate redundant content. Worth reviewing to see if they should be merged.
- Missing entries — concepts mentioned across multiple entries that don’t have their own page yet. A clear signal of “you’ve been writing about X without having an X entry.”
- Tags. Tag usage analysis with suggestions for each entry based on its content, plus groups of tags that frequently co-occur. Useful for cleaning up a messy tagging system.
Three sections for finding what’s broken:
- Anomalies. Unusual entries flagged by six scoring signals — isolation (no relationships), completeness (missing expected sections), tag isolation (unique tags that don’t appear elsewhere), size outlier (unusually short or long), and staleness (not updated in a long time). Each anomaly gets a composite score and a list of human-readable reasons explaining why it was flagged.
- Contradictions. Conflicting descriptions of the same entity across different lore entries. Found by comparing context snippets from each mention of the same entity and flagging where the descriptions disagree. Each contradiction records the entries involved, the entity, the severity, and a confidence score.
- Validation. Structural integrity checks — missing cross-references between entries that mention the same entities, orphaned entries with no relationships, and entities referenced in prose but missing from the entity table. Each issue has a severity, description, and suggested fix.
Three sections for understanding shape:
- Coverage. Which parts of your world are well-documented and which are under-served. Broken down by entry type, by category, by entity type, and by mention distribution. A project with 300 character entries and 10 location entries has a coverage problem; Coverage makes it visible.
- Maturity. Each lore entry scored on completeness using eight signals — word count, section fill, section depth, entity density, relationship count, summary presence, tag presence, section status. Entries receive a tier: stub, draft, developing, mature, or comprehensive. Filter and sort by tier, score, title, or entry type. The view that answers “which entries need more work?”
- Entity Graph. Analysis of your entity relationship network — degree centrality (how connected), betweenness centrality (how central), PageRank (importance), community detection (groups of closely connected entities), isolate detection (entities with no relationships), bridge detection (entities whose removal would disconnect parts of the graph), and missing reciprocals (one-way relationships where the reverse link is missing).
How to actually use it
Section titled “How to actually use it”- Open the Lore Intelligence panel from the Analysis workspace.
- The top bar shows the current health score (if trained) and the last training timestamp.
- Click Train to run the full pipeline. Training loads all your lore data, runs all the analysis modules, and persists the results. A results banner shows counts — clusters found, anomalies detected, contradictions found, validation issues, training duration.
- Use the left sidebar to navigate between the ten analysis sections. Start with Health for an overview, then drill into whichever section has the most interesting signal.
- Click Refresh to reload the current section’s data without retraining. Useful after you’ve made changes and want to see the updated view for a specific section without waiting for a full retrain.
Smart training
Section titled “Smart training”Training is smart in a specific way: if your lore corpus hasn’t grown significantly since the last training run, the pipeline skips retraining unless you force it. This prevents wasted time on unchanged data. The threshold for “grown significantly” is configurable but defaults to a reasonable value.
The minimum corpus size for training is 5 lore entries. Below that, there’s not enough data for the statistical analyses to produce useful results. Above that, results get richer as the corpus grows.
Where Lore Intelligence fits
Section titled “Where Lore Intelligence fits”The ML pipeline is the analytical counterpart to the rest of the agent system. Where Lorekeeper checks prose against the Legendry for runtime consistency, Lore Intelligence analyzes the Legendry itself for structural patterns. The two complement each other — Lorekeeper is “is this chapter consistent with my canon,” Lore Intelligence is “is my canon itself structurally sound.”
Lore Intelligence lives in the Analysis workspace alongside Editorial Analysis, Agent Overview, Etherforce Observability, Legendry Bench, and Error Tracking. All of them are diagnostic views over different parts of your project’s health.