A sovereign dense engine reproduces HF logits across 16 transformer families

1. Contribution (and what is explicitly NOT claimed)

Claimed. A single dense inference engine, containing no vendor BLAS (vendor GEMM/DNN libraries absent from the linked-library list) and no third-party ML framework, reproduces the next-token logits of a HuggingFace reference to cosine >= 0.999979 (f16) across 15 transformer families spanning three positional paradigms and both attention and non-attention sequence mixers. Per-mechanism ablations (negative controls) confirm that each reimplemented component is load-bearing, and adding the SSM/RWKV/MoE/ALiBi/parallel-block families causes no logit regression in previously-passing families.

NOT claimed. This is a correctness-reproduction result, i.e. the reimplemented forward matches the reference math. It is not a generation-quality, long-context, downstream-task, or throughput result, and the evaluation is smoke-test scale (see Section 5). "Sovereign reimplementation across many architectures" is an engineering-breadth contribution plus a small number of mechanism-ablation findings; it is not a claim of a new model or a new algorithm.

2. Methods

The comparison is on identical ids with identical weights, so a high cosine is evidence of forward-math equivalence, not merely plausible output.

3. Results - per-family parity

15/15 pass the cosine gate; 13/15 are full-argmax. (dagger) The two misses are sub-f16-ULP ties: at the contested position the engine's two top logits are exactly equal (gap on the order of one ULP at that magnitude) and the oracle breaks the tie at a lower index. With no f32 path these are irreducible, not defects (cosine ~ 1.0). A 16th family is excluded from results - the available checkpoint held a config-only weight stub, reported rather than faked.

4. Ablations (negative controls)

Each reimplemented mechanism is disabled while holding everything else fixed; a faithful port must break:

These are the result's strongest evidence: parity could in principle be coincidental, but a mechanism whose removal collapses the forward is demonstrably the one being computed.

5. Threats to validity / limitations

• Sample size is small. The per-family sweep is 1 prompt x 12 next-token positions (a few ports add a second prompt). This is a correctness smoke test, not a benchmark-scale evaluation; it can catch a wrong forward but cannot estimate rare-token tail behaviour. A benchmark-grade follow-up (hundreds of prompts from a held-out corpus, reporting the cosine distribution and worst case) is not done here.
• No variance reported. Logit comparison is deterministic given fixed ids/weights, so repetition is not informative for cosine - but any timing claims (absent here) would need repeated trials. This report makes no throughput claim.
• f16 ceiling. No f32 path, so sub-ULP argmax ties are irreducible and any precision-sensitive claim is bounded by f16.
• Parity != quality. Forward-math equivalence does not establish generation quality, calibration, or downstream-task performance.
• Quantization beyond the tested q8 spot-checks is out of scope.
• Uncovered axes (named): MLA, 2D/m-rope, NoPE-attention, hybrid SSM+attention (Jamba), Mamba-2/SSD, Mixture-of-Depths.

6. Reproducibility

• The engine is a single sovereign build; per-family ports are versioned independently.
• Build verification: confirm the linked-library list shows only the platform GPU runtime plus system libraries (no vendor BLAS/DNN).
• Run: dump per-position logits at f16 for a short fixed prompt, then diff against an oracle pinned to the releases described in Section 2.
• Validate the oracle first (the compressed-tensors trap, Section 2): an unpinned transformers silently random-inits MoE experts and invalidates the comparison.