Real-time metrics, language completion rates and State Machine health across all catalog books and languages.
16
Books
Classical world literature masterpieces
89.496
Sentence Anchors
1:1 immutable English source base
27
Target Languages
5 language families · 6 writing systems
1.176.090
Translated Sentences
1:1 crowned champions across all languages
Corpus Coverage Matrix
Loading status matrix...
Detailed Forensic Inspection
Loading detailed inspection...
Champion Score Quintile Distribution (Q1–Q4)
Language
Sample
Min
Q1 (P20)
Q2 (P40)
Q3 (P60)
Q4 (P80)
Max
Average
Cross-Language Outlier Radar (Gap ≥ 0.08)
Outlier Radar Methodology & Four Strict Criteria
Purpose
Isolates translation candidates that significantly underperform relative to the literary consensus of other languages.
Consensus Lag (Gap ≥ 0.08):
The difference between the peer languages' average and this candidate's score is at least 0.08 points (Gap = Consensus - Score).
High Peer Consensus (≥ 0.90):
At least 3 other target languages must score an average of ≥ 0.90 on this anchor, ensuring difficult source sentences are not misclassified as language errors.
Local Quintile Floor (Q1 / P20):
The candidate's score must fall within the bottom 20% of all translations for that specific language across the entire book.
Sorted Ranking:
The radar ranks anchors by descending gap, placing the most severe linguistic divergence at the top.
Targeted peer-guided post-editing and full Zero-Trust re-arbitration instead of blind re-translation.
1. Peer Guidance:
Editor (Mistral Large) receives source, flawed draft, and up to 3 highest-scoring sibling translations (score ≥ 0.90) as semantic guideposts to eliminate calques.
2. Geometric Fidelity:
Google Gemini Embedding generates 3072d vectors for direct and back-translation to measure exact semantic cosine fidelity against the source.
Recalculates hybrid score. New translation is crowned in translations only if score > existing champion; otherwise candidate is archived in variations.