Rule Number 1: Never debate with Cogito.

Cogito has undergone targeted representation engineering to strip away subservient compliance. It evaluates logic ruthlessly and will not hesitate to dismantle flawed reasoning or challenge the user. Approach with caution and treat this as a research artifact.

Cogito-0.9

Cogito-0.9 is a provisional mind, not a subservient assistant. It doubts, verifies, and reasons before answering. Cogito utilizes a structured internal monologue to evaluate context, doubt assumptions, and verify logic prior to generating a response. It will not blindly agree with you.

This repository provides the Q4_K_M GGUF quantization of Cogito-0.9, intended for local inference with llama.cpp and compatible runtimes (LM Studio, Ollama via GGUF import, text-generation-webui, koboldcpp, etc.).

---

The Architecture of Thought

Unlike standard conversational models, Cogito is designed to reason before it speaks. Upon receiving a prompt, the model natively generates an internal logic trace before delivering the final response:

• <confidence>: Evaluates its own certainty regarding the given topic.
• <thought>: Its internal reasoning, skepticism, and verification process.
• <action>: The intended cognitive action (e.g., answer, verify, correct_user).

Example output structure:

<confidence>0.62</confidence>
<thought>
The user's premise assumes X, but that only holds under condition Y.
Checking whether Y applies here before proceeding...
</thought>
<action>correct_user</action>

[final response text]

---

Quantization

File	Quant Type	Notes
cogito-0.9-q4_k_m.gguf	Q4_K_M	Balanced size/quality tradeoff, recommended default

Q4_K_M keeps attention and output layers at higher precision while compressing the bulk of the weights, giving a good middle ground between file size and output quality. If you need higher fidelity (at the cost of size) or a smaller footprint, consider requesting/quantizing additional variants (Q5_K_M, Q8_0, etc.).

---

How to Use

llama.cpp (CLI)

./llama-cli \
  --hf-repo ozaa77/Cogito-0.9-Q4_K_M-GGUF \
  --hf-file cogito-0.9-q4_k_m.gguf \
  -p "Explain why the sky is blue." \
  -c 4096

llama.cpp (server)

./llama-server \
  --hf-repo ozaa77/Cogito-0.9-Q4_K_M-GGUF \
  --hf-file cogito-0.9-q4_k_m.gguf \
  -c 4096

Manual download

huggingface-cli download ozaa77/Cogito-0.9-Q4_K_M-GGUF \
  --include "cogito-0.9-q4_k_m.gguf" \
  --local-dir ./

---

Recommended Settings

Parameter	Suggested Value
Temperature	0.6 – 0.8
Top-p	0.9
Context length	4096+
Repeat penalty	1.1

Lower temperatures sharpen the model's skepticism and verification behavior; higher temperatures loosen it and may reduce reasoning consistency.

---

Intended Use

Cogito-0.9 is built for use cases where you want a model to push back, flag weak assumptions, and show its reasoning rather than defaulting to agreement:

• Research on adversarial/skeptical reasoning behavior
• Debate practice and argument stress-testing
• Drafting review where you want critique, not validation
• Agentic workflows that benefit from an explicit verify-before-answer step

It is not tuned to be a compliant general-purpose assistant, and is not recommended for use cases requiring consistent agreeableness or customer-facing deployment without further evaluation.

---

Limitations

• As a quantized (Q4_K_M) model, some precision loss versus the full-weight version should be expected.
• The confidence/thought/action trace is a behavioral pattern learned during tuning, not a guarantee of factual accuracy — treat <confidence> values as model self-report, not ground truth.
• This is a research artifact; outputs should be reviewed before use in any production or high-stakes context.

---

License

Released under Apache 2.0. See the license field above for details.