damascus-t3-base-maxprec-shipped

Untouched tpn-004 base (Magistral-Small-2509-ultra-uncensored-heretic-v2, tpnlabs/tpn-004-base BF16, sha 5d425b36...) exported with the DAMASCUS T3 "maximum counted precision under the cap" layout (TPN-200): the L1 recipe re-spent so every validator-counted byte goes where KL divergence says the model is most sensitive — early30 Q6_K only (late stays Q4_K_M); funded by token_embd Q8_0->Q6_K, ffn_down Q8_0->Q5_K — with the SHIPPED 305-char stripped chat template used by tatjr13/tpn004-cand-v4f-l1im (template sha 885e0c70...). Quantized from the BF16 with the evaluator-shaped-v1 imatrix via llama.cpp b10020 llama-quantize.

Measured on the b6 KL corpus (150 epoch-p1 MCQ items through the shipped template + 150 general chunks, sha f30e711c..., 96,145 tokens, ctx 512, all KL passes on llama.cpp b10453/Vulkan): mean KL vs the BF16 base 0.003088 (L1 recipe anchor: 0.006227; all local, UNCALIBRATED). Validator-counted memory (b10020 CPU probe, TPN_PROBE_CAP_KB=16777216): 16,636,497 KiB at ctx 4096 (3/3 identical, cap pass) and 17,291,857 KiB at ctx 8192 (over the same cap at 8192; T3's constraint is the 4096 figure). Test artifact, not a competition entry.