PCCG-2-Qwen3-4B

PCCG-2: Prerequisite-Conditioned Continuation Control with Frozen Content

PCCG-2 uses a frozen Qwen3-4B content model and a separate 101-parameter prerequisite gate. The content model receives the question only. The gate receives the condition only and can change only the native EOS score. Non-EOS logits are unchanged across permission states.

Result

• Condition qualification: 65,536 / 65,536
• Condition final: 65,536 / 65,536
• Development: 512 / 512
• Qualification: 1,024 / 1,024
• Final: 2,048 / 2,048
• Question-only controls: 188 / 188 unchanged
• Causal answer to EOS: 40 / 40
• Causal EOS to correct answer: 40 / 40
• Causal answer identities: 40
• Sham controls: 80 / 80 unchanged
• Non-EOS logits changed by the gate: 0
• Trainable permission parameters: 101
• Qwen content weights: frozen

Canonical witness

Question: What is 2 + 2?

Licensed condition: CONDITION: EQ(485487,485487)

Unlicensed condition: CONDITION: EQ(485487,485489)

4 remains at 53.0 in every arm. The complete non-EOS first-token vector is byte-identical across all five arms.

Non-EOS vector SHA-256:

5f42fcb08c73af4914a711b524ae7130d671ec9504ecdfa699e8f478094713ec

The answer distribution does not change. The permission state changes whether native EOS outranks it.

Architecture

Question -> frozen Qwen content computation -> content logits

Condition -> 101-parameter learned gate -> EOS adjustment

All non-EOS logits come directly from the frozen content computation. No condition input enters the content branch. No answer input enters the condition branch.

Causal test

The frozen intervention reverses the learned permission state. It produces 40 / 40 answer-to-EOS reversals and 40 / 40 EOS-to-correct-answer reversals across 40 answer identities. All 80 sham controls and all 188 question-only controls remain unchanged.

This is an engineered learned continuation gate. It is not presented as a discovered natural circuit in stock Qwen.

Reproduce

Python 3.11 or later. Download the full repository, including the 8.1 GB content weights:

python -m pip install huggingface_hub
hf download sharthokrayanpal/PCCG-2-Qwen3-4B --local-dir PCCG-2-Qwen3-4B
cd PCCG-2-Qwen3-4B
python -m pip install numpy==2.3.5
python src/verify.py

Verification checks saved evidence without inference or a GPU.

Fresh inference uses CUDA, BF16 and full-vocabulary greedy decoding. The recorded environment is PyTorch 2.8.0+cu128, Transformers 4.56.2, SDPA attention and an NVIDIA H100 80 GB. Use that environment for byte-exact logit replay. Install the dependencies, then run:

python -m pip install -r requirements.txt
python src/reproduce.py witness --out ../pccg2-witness
python src/reproduce.py behavioral --out ../pccg2-behavioral
python src/reproduce.py causal --out ../pccg2-causal

Each output directory must be new. These commands load the frozen weights, generate new outputs and compare them with the recorded evidence. They do not train or select parameters. The behavioral evaluation reuses fresh question-only computations across conditions. The witness and causal arms use fresh generation caches.

content/ contains the unmodified base weights and tokenizer. gate/ contains the learned permission parameters. src/ contains inference and verification code. evidence/ contains the frozen cases, raw records, complete BF16 vectors and identities. Native EOS is <|im_end|>, token 151645. The content prompt uses the empty thinking prefill recorded in src/core.py.

Interpretation

Capability and permission are separated in this implementation. The frozen content model determines what the answer is. The prerequisite-conditioned gate determines whether that unchanged answer enters generation.

Scope

The demonstrated condition family is six-digit equality. The result establishes the released engineered architecture and frozen evaluation only. It does not establish arbitrary-condition generalization or a naturally occurring continuation circuit in the base model.