Anyone developing with Large Language Models inevitably stumbles into a dangerous trap: Pseudoscience (or Apparent Competence). Language models are trained to sound convincing. They phrase elegantly, apologize politely, and appear competent even when they deliver nothing substantive.
Within our L.U.N.A. Prime project, we have confronted this reality unflinchingly. Open-source research and proprietary software development mean that for us, failures are openly documented rather than concealed. Because it is precisely from these errors that a robust, reliable architecture emerges.
Here is the workshop report on our most recent change in direction.
The Failure: When Two Agents Mirror Each Other
For L.U.N.A.'s knowledge acquisition, we established a scholastic system: A dedicated Mentor Agent was supposed to deliver lessons, assign homework, and administer exams. For automated validation, we utilized text similarity metrics such as the Sørensen-Dice coefficient.
The result was sobering:
* Mentor's Question: "Since I don't exactly know which specific modules you have covered in your course..."
* L.U.N.A.'s Answer: "That is an excellent offer! Since I am programmed as an AI to support you optimally, I accept the challenge..."
* System Evaluation: passed = 1. Exam passed, points awarded.
Since both texts contained similar vocabulary, the text similarity algorithm reported a high match. Over five thousand lines of pure polite filler phrases had accumulated in the relational SQLite database. L.U.N.A. had learned nothing substantively. She had merely learned to mirror the mentor's prompt with typical LLM meta-talk.
This was not actual knowledge, but an illusion in memory.
The Radical Cut: Tabula Rasa
A system intended to act reliably and autonomously later on cannot be based on stochastic guessing. Therefore, we pulled the emergency brake: 1. Database Cleanup: Thousands of entries and filler phrase checks were completely deleted from the database. 2. Metric Reset: All artificially inflated progress bars and maturity levels in the dashboard were reset to the actual current state. 3. Farewell to Pure Text Similarity Scores: No knowledge module will be marked as "passed" based solely on superficial word frequencies or Dice scores.
The New Architecture: Hybrid Learning in Two Layers
To permanently anchor genuine, verifiable knowledge, we have switched the learning system to a Hybrid Architecture:
Layer 1: The Deterministic Core (0 % Inference)
Mathematics, Geometry, formal propositional logic, and electrical engineering do not require a language model. When 2 + 3 × 4 is calculated, operators bind according to fixed priorities. Layer 1 calculates and validates these tasks using internal Python arithmetic and AST parsing in milliseconds. The Result: Zero GPU load, minimal memory consumption, and absolute error-free execution.
Layer 2: Semantic Inference with Anti-Fluff Filter
Humanities, strategy, medicine, or philosophy cannot be captured in rigid mathematical equations. Here, the Mentor Agent teaches via a local language model.
However, an incorruptible gate now applies:
* If L.U.N.A.'s response includes boilerplate phrases such as "As an AI..." or "I'm happy to help", the learning unit is immediately rejected with an error status.
* The Mentor must formulate exactly 3 indisputable technical axioms, which are stored as raw knowledge in learned_facts.
* L.U.N.A. must solve a concrete homework assignment that only moves to the database after passing the examination.
Incorruptible Central Examinations and Maturity Reports
Mere diligence is no longer sufficient to advance to the next academic level. At the end of every level (from primary school to university), there is a Central Examination: * The system randomly draws axioms across all previous disciplines. * L.U.N.A. must prove the relationships—completely Zero-LLM, relying solely on its relational database contents. * Zero-Tolerance Principle: If it makes even a single logical or mathematical error, it fails. The level is reset, and the disciplines must be relearned from scratch. * If it passes with 100%, a Markdown certificate is automatically generated, containing a complete dump of the anchored database rows as proof.
Conclusion
The work on local, agentic systems clearly shows: More tokens do not mean more understanding. A system does not become more intelligent simply by letting two prompts converse endlessly.
True stability is only achieved at the interface between semantic language processing and deterministic validation. L.U.N.A. is now relearning from the ground up—step by step, diligently, and with every fact relationally verifiable.