nx_llm_capcheck.nx

nx_llm_capcheck.nx -- SOVEREIGN DEEP CAPABILITY CHECKER for the no-float Qwen2.5-0.5B (our own "Unsloth-style" harness: measure SPEED + ACCURACY together, reproducibly, on the REAL model -- no TSV, no strawman). For a suite of known-answer prompts it runs the full sovereign pipeline (BPE encode -> 24-layer integer forward -> final norm -> 151936-vocab LM head -> argmax -> BPE decode) and measures, via the sys_now_ms clock, isolated: * forward_ms -- the 24-layer integer stack (incl. on-the-fly weight dequant; no weight cache yet) * lmhead_ms -- the one full-vocab output projection * load_ms -- one-time model parse (amortized in real use) and CAPABILITY: does the decoded prediction contain the expected answer? -> accuracy X/N. HONEST: a measurement harness PASSES when it MEASURES correctly; the accuracy number is the finding, not gated -- it surfaces the known faithfulness gap (component-verified matmul/RoPE, full-stack H2H still open) so we can fix it. No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL