Skip to content

0068: Select the output head's final normalization

2026-10-23. Correct the extraction plan under the existing complete-response acceptance gate. The quantized model, arithmetic, Rust source and verifier remain unchanged.

All 23 non-delayed proofs verified individually, and the prompt/generated IDs matched. The complete grid rejected the last residual link: the head committed layer 21's input instead of its output. Its extraction command had combined --head and --residual-stream. The dump contained both families, and the pinned Rust loader chooses the ordinary residual family before considering the final norm. The head's boundary roots exactly repeated layer 21's roots. Matching xin_head.u16 bytes alone did not show which family the proof consumed.

Extract the separate head with --head and without --residual-stream, into a fresh directory. Require an empty ordinary residual_stream manifest entry and head.final_norm: true. Require every retained numeric array to equal the prior head extraction and xin_head.u16 to equal the last layer's output. Do not edit old manifests or relabel the failed run as successful.

The 22 transformer proofs already verified against roots enrolled before proving; their exact bytes, roots and statements remain frozen. Reuse those unchanged segments in a new campaign, enroll a fresh head model/shape root before proving, and generate a new head proof. Require params.residual_stream.head_norm: true in the caller's head-statement check. Require all 23 proofs, all 22 residual links, all prompt/generated positions and the final output token to pass. No link is removed or treated as optional.

The final gate remains a fresh verifier worker with no weights or witness traces, using the caller's enrolled roots and exact statement parameter digests, followed by rejection of changed inputs, output, model, calibration, coverage and proof bytes. This is one disclosed engineering prompt; it does not qualify general language-model quality, hiding or independent cryptographic security.