Operation Convergence — when TatuEngine decided to teach itself
TatuEngine·

Operation Convergence — when TatuEngine decided to teach itself

5 min read← Back to timeline
#tatuengine#convergence#teacher-student#sft#reasoning

The engine that decided to teach itself

First TatuEngine learned to infer: a ternary 1B-parameter model running on GPU with 252× speedup, 114× compression, bit-identical to CPU. Then it learned to act: an autopoietic agent with 8 subsystems, a security sandbox, MCP, memory.

Then came the question that changes everything: what if the engine learns to teach itself?

The answer was Operation Convergence — an intense R&D day (04/07/2026) that stacked six versions (v0.17.0 → v0.22.2) to build the Master-Apprentice cycle: a teacher model distilling reasoning into a student model. And like any good saga, it almost didn’t work.

Context — the teacher and the student

The Master-Apprentice cycle is the gateway to self-improvement:

  • Teacher (master): Qwen2.5-3B-Instruct + LoRA (R=16, alpha=32, targets q/k/v/o) — the model that knows how to reason
  • Student (apprentice): BitMamba-2 1B ternary — the engine’s model, which doesn’t know yet
  • Distillation: the teacher generates examples, the student learns from them

Before Operation Convergence, the engine already had the base: OpenAI-compatible API (v0.13), chat TUI (v0.15), benchmark (v0.16), TatuViz with cognitive graph (v0.17). What was missing: structured reasoning — and the bridge between the two models.

The struggle — six versions in one day

v0.18: Cognitive Density

“Semantic Bridges Operation” — a contextual hypercube that connects concepts before answering. The first layer of density.

v0.19: the Quality Ladder

Structured Reasoning Protocol (Chain-of-Thought) with a tag format:

[THOUGHT] ... step-by-step reasoning ... [/THOUGHT]
[ANSWER] ... final answer ... [/ANSWER]

The student doesn’t just learn the answer — it learns the format of thinking.

v0.20: the router

Teacher-Student Multi-Engine Router — routes each request between Qwen 3B (teacher) and BitMamba 1B (student). The teacher generates, the student tries.

v0.21: Total Attack

Forced semantic mining + Mamba-2 R&D — pushes the model to mine knowledge, not just reproduce it.

The critical moment: the convergence that wouldn’t come

This is where the saga almost died. The student wasn’t converging — the answers didn’t follow the format, alignment kept breaking. Two commits fixed it: Operation Convergence (forced semantic mining) and the teacher system message fix.

The villain was the teacher’s system message: the instruction the teacher received contaminated the distillation. Fixing the professor’s prompt unlocked convergence.

v0.22.1-2: the rigorous validator

Only accepts what has closed tags, coherent steps and a complete thought/answer pair.

Resolution — the complete cycle

Operation Convergence delivered the entire Master-Apprentice infrastructure:

  • Trained teacher LoRA — Qwen2.5-3B-Instruct, R=16, alpha=32, targets q/k/v/o
  • Working distillation — 381 examples generated with the [THOUGHT]/[ANSWER] protocol
  • 3 consolidated datasets — tatu_train_data (349) + student_train (23) + distill_clean_train (9)
  • Differentiable PyTorch student — 1B with forward + backward validated (RTX 3060, CUDA 13.0), gradients flowing

The next milestone: Student SFT — actually training BitMamba-1B (3 epochs, 381 examples, GPU) and comparing Student vs Teacher. The engine learned to infer, learned to act, and now it’s learning to learn.

Metrics

Metric Value
Versions in the arc (v0.17→v0.22.2) 6 in one day (04/07/2026)
Teacher Qwen2.5-3B-Instruct + LoRA R=16, alpha=32
Distilled examples 381 (3 datasets: 349 + 23 + 9)
Student BitMamba-2 1B differentiable (forward+backward OK)
Convergence fix teacher system message (v0.22.0)
Validator rigorous: tags, steps, thought/answer (v0.22.1)

Lessons

  1. Convergence isn’t automatic — the teacher matters — the fix that unlocked everything was the teacher’s system message, not the student. The apprentice only converges if the master teaches the right way.
  2. Teaching format is teaching reasoning — the [THOUGHT]/[ANSWER] protocol makes the student learn the process, not just the answer.
  3. Rigorous validation in distillation — the tags/steps/thought-answer validator ensures garbage doesn’t become training data.
  4. Self-improvement is the destination — infer → act → learn. The Master-Apprentice cycle is the step toward the Teacher→Student→generation→distillation→Teacher loop.