Union Alpha

Independent YFarmX measurements indicate that Union Alpha is consistent with the Llama3 family at moderate confidence. These measurements establish lineage evidence only: a shared tokenizer or serving contract shows shared ancestry, and does not independently establish the precise underlying release.

Listed anonymously on 16 September 2026 and measured the same day. This endpoint returns TWO different token counts for a byte-identical request, and they describe different tokenizers, which is why early readings of it disagreed. A differential test settles which is real: tripling an input must triple its content cost, and any fixed chat template cancels in the subtraction. On the higher path the content scales exactly, 30 digits costing 10 and 60 costing 20, a Chinese phrase 13 and its triple 39, one family emoji 15 and three 45, once a constant 17 is removed. On the lower path the same sums come out one or two tokens adrift, so it reports a counter rather than a vocabulary. The surviving counts are Llama 3s: 47 of 50 strings match, against a median of 15 across 281 measured models, and the three template-free differential values are predicted by 14 models out of 309, every one of them a Llama-3 build. This places a VOCABULARY and not a company: Perplexity Sonar and the Hermes and sao10k lines sit in that same group. An earlier reading of this record followed the lower path and reported a Qwen family; it is superseded. Raw measurements, the differential and the working are in docs/research/union-alpha/.

Union Alpha on OpenRouter (stealth listing, 16 September 2026)Meta Llama 3 tokenizer (Llama-3.1-8B-Instruct repository)