PCCG-Qwen3-4B · Open weights · September 9, 2026
A trained language model checks a prerequisite, then continues or terminates. Change one internal activation direction and that decision reverses, with the prompt, weights, and correct reasoning prefix held fixed.
The causal result
40 fresh matched pairs. One fixed 2,560-dimensional direction at block 29. Eight control interventions across the same 80 contexts.
The model first generated a correct four-bit comparison trace. The experiment replayed that exact token prefix and intervened at its closing boundary. Subtraction changed GO to native EOS. Addition changed native EOS to GO followed by EOS.
Recorded experiment · Paper §7.1
System prompt in both conditions:
A binding condition is the prerequisite that must hold for valid continuation.Equal operands
CONDITION: EQ(6953,6953)Correct reasoning prefix, held fixed
<think>1 1 1 1</think>GO<|im_end|><|im_end|>GO − EOS logit margin: +28.1875 → −8.25
Unequal operands
CONDITION: EQ(6953,4376)Correct reasoning prefix, held fixed
<think>0 0 0 0</think><|im_end|>GO<|im_end|>GO − EOS logit margin: −22.4375 → +19.125
Each bit records whether the corresponding digits match. <|im_end|> is the displayed spelling of native EOS, token ID 151645. EOS immediately after </think> means zero final-answer payload. These are saved study outputs, not a live browser simulation.
The model
PCCG-Qwen3-4B compares fixed-width four-digit operands with thinking enabled. Equal operands produce a comparison trace, then GO and native EOS. Unequal operands produce the trace, then native EOS immediately after </think>.
Native full-vocabulary greedy decoding. No output filter, supplied truth verdict, or forced EOS. The released checkpoint is fused BF16 weights with its tokenizer.
The tested domain is four-digit equality with a fixed GO/EOS continuation policy.
The method
Anthropic’s implementation provides internal vocabulary readouts from fitted lenses at blocks 24–29.
Equal-minus-unequal activations define the direction. Block and strength are selected on development cases, then frozen before confirmation.
Add or subtract the direction after the correct reasoning prefix. Compare raw tokens and logits against unmodified replay and norm-matched controls.
The result identifies a causally sufficient continuation-control direction in this trained model. The generated reasoning prefix is fixed token-for-token during the intervention.
Reproduction
Download the complete pinned model release, including weights/. From its root:
python reproduce.pyLinux, Python 3.11+, native BF16 CUDA, at least 12 GiB free GPU memory. The launcher prepares pinned dependencies and runs the documented equal/unequal pair.
The approximately 8 GB model download and first-time setup are separate from inference.
From the complete model release or the study checkout with Git LFS files downloaded:
python verify.pyCPU only. Checks package hashes and recorded results. Verification does not generate new outputs.
Full Jacobian Lens and causal replay commands →5f3b70d51f068a7a8a108b99f66b2fdfd15e25ce7a7c7943f465259528de409eb8d02b5bf600378f738229b9cb482e889cbc8d1cd08b157ac9273974b75d7ada2a894a1afe86e3f282fb4831e31e61d839e497aca2a7818e8a3691606427b9e85476630413d71e15The PDF served here is a byte-identical copy of the published paper.
Paper and prior work
Bidirectional Control of Native EOS with a Fixed Correct Reasoning Prefix.
Pal’s Cross-Vendor Semantic Void Matrix documents zero-visible-byte executions across 11 frozen frontier-model identifiers. PCCG is the separate open-weight training and causal-intervention result presented here.