RFC-001: I'm not a developer, but my model got 5K downloads on HF this month. It's shaky, and I honestly need your advice.

[ittirit-delentia]

Hello engineering community,

I want to be completely honest right from the start: I have no formal training in Computer Science, and 12 months ago, I didn't even know how to write code. I graduated with a degree in Architecture, majoring in Facilities Management. I am not a software engineer. I just wanted to learn, challenge myself, and build a computer project of my own—combining my background in resource allocation with AI to manage hardware constraints. That's how this project came to be.

Recently, I checked Hugging Face and saw that my base model card (delentia-slm-jitna-v0.4) has 5,247 downloads last month (over 7,400+ combined downloads across the entire v0.4 cognitive framework collection).

To be frank: I don't know if these downloads are bots or real users testing it. I am honestly confused, and a bit scared.

I built this SDK (delentia-os) and got it to pass a lot of unit tests, but in my eyes, the system is not stable yet. I have reached the absolute boundary of what a solo self-learner can figure out alone. I am not opening this discussion to show off a finished product. I just want to find people to talk to, ask for advice, and learn from.

If you have experience with local LLMs, compilers, or matrix math, I would be deeply grateful for your perspective on these issues I'm currently struggling with:

---

1. The Swapping Lag (Guardian -> Scribe Transition)

Dynamic switching between the Guardian adapter (safety check) and Scribe adapter (context compression) under high-intent variance (Δ) scenarios causes latency spikes and matrix initialization overhead on low-VRAM edge devices.

• If you've done dynamic adapter swapping on consumer hardware, how do you handle memory transitions without crashing the inference context?

2. Hallucination Chaos (Three-Body Synthesis)

When I load multiple adapters (Guardian, Scribe, and Executor) simultaneously on the frozen 8B base model, the output sometimes turns into absolute neural chaos and gibberish.

• I read about Rank-Stabilized LoRA (RSLoRA) and consensus thresholds. In v0.4.3, I tried forcing J-Space parameters directly during training via <cognitive_state> tags to stabilize outputs. But does this training-level constraint actually prevent multi-adapter weight interference at runtime, or do I still need active weight routing / SVD merging?

3. The Re-training Nightmare

Every time I update the base model, the old adapters suffer from Distribution Shift and stop working. Running SFT training scripts from scratch overnight for every update feels incredibly inefficient.

• How do open-source developers maintain backward compatibility for PEFT weights when the base model shifts? Is there a way to avoid retraining from scratch?

4. Breaking the Safety Equation

I wrote a basic mathematical gate to control model output safety: F = D^I x A (where A = 0 represents a constitutional kill switch). If you want to try and break my safety checks, please clone the repo, run the adversarial suite, and tell me where it fails:

git clone https://github.com/delentia-labs/Delentia-AI-SLM.git
cd Delentia-AI-SLM
pytest training/test_adapters_hypothesis.py -v

5. Mapping Southeast Asian Languages

I built the JITNA coordinates with Thai linguistic structure in mind. If you work with regional NLP (Thai, Vietnamese, Indonesian, etc.) - does mapping intent to a 6-coordinate system (I, D, Δ, A, R, M) make sense, or is it a fundamentally flawed approach?

6. Am I solving the right problem? (The Endgame)

I don't know what the final form of this project should be.

• Should it continue towards the Sovereign AI Protocol path—acting as an open, architecture-agnostic standard that allows different regions to plug in their own national models (like Typhoon or Mistral) while inheriting mathematically enforced safety rules? Or should it just remain a local privacy framework for PDPA/GDPR compliance?

---

I am just a regular person trying to build one thing of my own that actually works. English is not my native language, so some of my documentation might sound a bit messy.

I would be so grateful if you could share your advice or criticism. Let's talk!

• Hugging Face Article (RFC-001): https://huggingface.co/blog/Ittirit-delentia/rfc-001-constitutional-ai-os
• Zenodo Paper: DOI: 10.5281/zenodo.20920052
• PyPI Package: pip install delentia-os