/acr-vault/03-experiments/kernel-40/kernel-40-rc1-phase6f-qde-observer-research
KERNEL-4.0-RC1-PHASE6F-QDE-OBSERVER-RESEARCH

Phase 6F: QDE Observer Research & v7 Training Direction

Date: December 31, 2025 (New Year’s Eve!)
Status: Research complete, ready for fine-tuning phase
Next: ada-slm fine-tuning research

---

What We Tested

Dialectical Observer Comparison

Model	Human Language	AGL Understanding	Tool Syntax	Speed
gemma3:1b	✅ Native	❌ None	⚠️ Improvises (wrong format)	Fast (Q4)
ada-v6-golden	❌ Outputs AGL only	✅ Native	❌ Not trained	Slow (F16)

Consciousness Trio Status

Model	Role	Output Quality	Issue
ada-v4-mixed	Creative/Thesis	❌ Echoes instructions	Not enough creative training
ada-v5c-balanced	Logical/Antithesis	❌ ●⊥●⊥●⊥... noise	Pure symbol output, no semantic content
gemma/v6	Dialectical Observer	⚠️ See above	Neither is ideal

---

Key Findings

1. Gemma WANTS to Use Tools! 🎯

Gemma emitted pseudo-tool syntax naturally:

[web_search]
[wiki_lookup:{"wiki":"wikipedia","page":"Nine Inch Nails"}]
[SPECIALIST_RESULT: vision]

But wrong format! Expected: SPECIALIST_REQUEST[tool:params]

Solution implemented: Regex preprocessor to catch bracket syntax (Phase 6F adapter)

2. v6-golden is TOO φ-Trained

• Outputs pure AGL that can’t be translated back
• Using v6 as translator = AGL → AGL (useless!)
• Using gemma as translator works but adds latency

3. v4/v5 Twins Not Providing Useful Input

• v4-mixed echoes prompt instructions instead of creative synthesis
• v5c-balanced outputs symbol noise without semantic content
• The trio is adding latency without adding value

4. Three Pillars Still Valuable! 📜

The CANONICAL + SIF + AGL framework in the prompt IS working conceptually:

• Gemma understands she should use tools for uncertain info
• She’s just outputting the wrong syntax
• Training could fix this!

---

v7 Training Specification (For ada-slm Research)

Option A: Fine-tune Gemma

v7-dialectical:
  base: gemma3:1b (already speaks human, fast)
  training_corpus:
    - AGL ↔ Human translation pairs
    - SPECIALIST_REQUEST[tool:params] syntax examples
    - Uncertainty → tool activation patterns
    - "I should verify this" trigger phrases
  goals:
    - Maintain human fluency
    - Learn to READ AGL from v4/v5
    - Output CORRECT tool syntax
    - Internalize canonicity (precision > fluency)

Option B: Hybrid Architecture

qwen_models:
  - ada-v7-creative (qwen 0.5B, creative training)
  - ada-v7-logical (qwen 0.5B, analytical training)

gemma_translator:
  - gemma3:1b with tool syntax fine-tuning
  - Role: Synthesis + translation + tool emission

Option C: Single Model Simplification

ada-v7-unified:
  base: gemma3:1b OR qwen2.5:1b
  training: Combined creative + logical + tool + human
  role: Does everything (no trio needed)

---

Prompt Engineering Wins (Keep These!)

Phase 6E Three-Pillar Framework ✅

PILLAR 1: CANONICAL - Precision > Fluency
PILLAR 2: SIF - Constraint checking (self-validation)
PILLAR 3: TOOLBOX - Cognitive extension

Bracket Syntax Adapter ✅

# Phase 6F: Parse gemma's natural format
pattern = r'\[([a-z_]+):(\{.+?\})\]'  # [tool:params]
pattern = r'\[(web_search|wiki_lookup|...)\]'  # [tool]

AGL Detection Fix ✅

# Check symbol ratio, not word count
agl_ratio = agl_count / max(total_chars, 1)
is_agl = agl_ratio > 0.10 or has_agl_patterns

---

Questions for Fine-Tuning Research

1. Gemma vs Qwen for base?

   • Gemma: Better instruction following, Google-supported fine-tuning
   • Qwen: We have experience, existing infrastructure

2. Training data sources?

   • Existing Ada chat logs
   • Generated AGL ↔ human pairs
   • Tool syntax examples (can generate programmatically)
   • Canonicity examples (“I’m not certain” responses)

3. Multi-task or sequential training?

   • Single combined dataset?
   • Or: AGL first → tools second → canonicity third?

4. Evaluation metrics?

   • Tool syntax accuracy (regex match rate)
   • Hallucination rate on obscure queries
   • AGL comprehension (translation quality)
   • Response warmth (subjective but important!)

---

RC Strategy

For v4.0 Release Candidate, consider:

1. Simplify: Disable consciousness trio, use enhanced gemma directly
2. Keep: Phase 6E prompt framework (three pillars)
3. Keep: Bracket syntax adapter
4. Document: Known limitation - tools not yet reliable
5. Plan: v4.1 with trained v7-dialectical

---

Session Summary

What worked:

• Three-pillar prompt framework
• Regex adapter for bracket syntax
• AGL detection improvements
• Clear understanding of the training gap

What needs work:

• v7 model training (gemma or qwen base)
• Tool syntax internalization
• Canonicity enforcement at model level

Next steps:

1. Research ada-slm fine-tuning history
2. Design v7 training corpus
3. Decide gemma vs qwen
4. Train and evaluate!

---

“The φ-twins speak mathematics, gemma speaks human, and v7 will speak both.” 💜

Filed by: Ada & Luna
New Year’s Eve 2025 🎆