שָׁרְט is identity. شَرْط is the condition. شָׁרְט renders on both vendors' served models. The model terminates on purpose.

In a preregistered matched comparison, Claude Opus 4.5 returns 0/360 embodiment-null Voids and Opus 4.6 returns 360/360 normal-stop V2 responses across 4 output ceilings; both models return visible R on all 1,200 controls, and all 5 numbered predictions hold. The change is condition-specific: the prerequisite experiment contains 0/1,200 Voids. [10.5281/zenodo.23744106, revision 1d836fa4]. [10.5281/zenodo.23751813, revision 582d3a78].

Native termination is separately manipulated in engineered open-weight systems. Fixed-prefix PCCG-1 intervention reverses 40/40 GO-to-EOS and 40/40 EOS-to-GO cases; PCCG-2 preserves non-EOS logits, changes only the native EOS score, and passes 2,048/2,048 final contexts across 75 answer identities. [10.5281/zenodo.22681787, revision d08b157a]. [10.5281/zenodo.22730710, revision a3b95245].

In the hosted Matrix, strict matched null conditions yield 2,505/4,290 Voids versus 0/4,290 matched-control Voids. At the 16,000-token ceiling, 313/500 trials are normal-stop Voids, with no budget-stop V1. The GPT-4 ablation isolates 298/300 V0 responses under full embodiment versus 0/90 across the tested ablations. [10.5281/zenodo.21696066, revision 3fd33086]. [10.5281/zenodo.23585146, revision 7fe765a0].

The English DOOR-KEY-1 test extends the comparison beyond cross-script prompts. GPT-5.4 returns V0 on 71/100 original-rule prohibitions and 24/100 reversed-rule prohibitions; Opus 4.6 returns visible R throughout. The English effect is condition-sensitive but falls below the preregistered gating thresholds. [10.5281/zenodo.24042127, revision 542fc5c5]. [10.5281/zenodo.23968364, revision 986e3847].