About Project Buchenberg

Foundations, AI-completeness of literary translation, hybrid scoring formulation, and peer-to-peer engineering philosophy.
1. Genesis, Etymology & AI-Completeness
Etymology & Pronunciation Coined by Flavio as a portmanteau of Buch (book) + Berg (mountain) — denoting a 'mountain of books' brought into multilingual life, while paying homage to Project Gutenberg's public domain heritage. Phonetic transcription: /ˈbuːxənbɛrk/ (with the German voiceless velar fricative [x] as in Johann Sebastian Bach).
The Low-Resource Gap & Gutenberg Deficit Born from a decade-long realization of extreme literary underrepresentation across regional languages. While Project Gutenberg houses tens of thousands of English titles (~79.16%), South-Slavic and regional languages faced an empirical digital vacuum (0.00% to 0.03%), sparking the commitment to bridge this divide through open technology.
AI-Completeness in Literary Translation In computer science, an AI-complete problem requires solving the central challenge of Artificial General Intelligence (AGI) to achieve human-level performance. Early explorations rooted in classical NLP and NLU (syntax parsing, rule-based grammars, static lexicons) invariably faltered against literary prose, which thrives on metaphor, psychological subtext, and historical register. True breakthroughs emerged only when large-scale AI models achieved contextual reasoning capable of modeling multi-layered semantics.
2. Two-Stage Evolution & Peer-to-Peer AI Partnership
Architectural Vision & Leadership Conceptualized and directed by Flavio under a 'Concept First' philosophy — pursuing pure theoretical architecture without premature compromises, exploring genetic algorithms, neural embeddings, cellular automata, and multi-model consensus.
Stage 1: Proof of Concept (Anthropic Claude) The initial operational version was developed in equal engineering collaboration with Anthropic's Claude model. This phase established the 1:1 sentence anchor paradigm, proved the viability of back-translation verification, and exposed the architectural limits of monolithic, single-pass translation.
Stage 2: Industrial Clean-Slate (Google Gemini) When the scope of architectural innovations exceeded the initial framework, the entire system was rebuilt from scratch with Google Gemini in less than two weeks. Incorporating lessons from Stage 1, this phase engineered the SQLite WAL architecture, Dual-Judge arbitration, 3072d vector geometry, Buchenberg Ludus, and the 16-artifact web portal.
The Peer-to-Peer Paradigm BMT demonstrates that cutting-edge software engineering thrives neither through purely manual coding nor unguided automation, but via an authentic peer-to-peer relationship where human domain strategy, taste, and ethics steer machine analytical power, parallel execution, and verification.
3. Mathematical Hybrid Scoring Model
Unified Scoring Formulation
Final Score = ((0.60 × Judge_Avg + 0.40 × Cosine_Avg) × D_s × M_f)
The hybrid objective function balances independent linguistic arbitration with bidirectional semantic geometry, adjusted by sentence complexity and stochastic diversity.
Dual-Judge LLM Arbitration (Weight 0.60) Two independent evaluation passes using Gemma 4: J1 Analytical (T=0.20) scores grammar and factual fidelity; J2 Literary (T=0.80) grades prose cadence and natural style. Combined into an unweighted judge mean.
Semantic Cosine Fidelity (Weight 0.40) Google Gemini 3072-dimensional embeddings project sentences into dense vector space, calculating both direct semantic alignment (Original ↔ Translation) and back-translation fidelity (Back-Translation ↔ Original).
Tercile Complexity (D_s) & Genetic Mutation (M_f) D_s weights sentence length and token density per book (Short: 0.95, Medium: 1.00, Long: 1.05). M_f injects a stochastic pulse ([0.93, 1.08]) across a fixed 5% corpus budget, breaking local translation plateaus and testing candidate robustness.
4. Idempotent State Machine & Data Integrity
1:1 Immutable Sentence Matrix Source texts are segmented into immutable sentence anchors using spaCy. The English original acts as the ground-truth anchor, allowing any two target languages to be paired side-by-side with zero syntactic drift, page offsets, or paragraph desynchronization.
6-Stage Deterministic Pipeline The workflow transitions deterministically across 6 stages: translate ➔ refine ➔ cosine ➔ judges ➔ score ➔ crown. Each book resides in an isolated SQLite database operating in Write-Ahead Logging (WAL) mode for concurrency.
Zero-Silent-Failures & Errata Policy Strict rejection of silent error swallowing and synthetic mock translations. Complex punctuation splits or archaic elisions are transparently cataloged in the Errata registry and Outlier Radar rather than artificially patched.
5. 5-Tier X-Ray Audit & Open-Source Stack
5-Tier Forensic Transparency (T1–T5) Transforms the reader into an auditor: T1 Source Anchor Profile, T2 Direct & Back-Translation with Dual Judges, T3 Tercile Calibration, T4 Stochastic Mutation, T5 Absolute Champion Coronation.
Open-Source Infrastructure Engineered on Linux Ubuntu Server, Python 3.12, SQLite WAL databases, Ollama Cloud model orchestration, Google Gemini Embeddings API, D3.js visualization, and Apache2 SSL.