Skip to content

0067: Require complete activation-scale binding

2026-10-23. A new quantized candidate under the owner's standing approval and decision 0064. The model, authored prompt, calibration texts and complete-response acceptance gate stay fixed.

The first real layer's detailed verdict reports rq_scale_open: ["qkv/v"]. At the pinned SASH revision, a delayed-value fold component publishes its divisor instead of its input-scale exponent. The verifier explicitly omits the equality between that scale and the producing requantizer's scale. Its broader binding table still labels the seam bound. That summary is insufficient to establish complete inference integrity, even if the proof eventually verifies.

Do not qualify y-delay-vw12 or spend on its complete-response proof. Retain its failures, proof bytes and constructor investigation without treating them as evidence for a fully bound LLM. No claim of an exploited forged proof is made; the missing equality is explicit in the pinned source and execution verdict.

The next candidate is the existing y-attnout-w12 profile: the same quantization settings with value projection rounding restored before its next quantization. It uses the chain gadget, whose input-scale equality is implemented. Keep the exact-integer Python witness fix from decision 0065. Freeze a new experiment and regenerate inference and every witness; do not reuse the delayed profile's proof or roots. The generated answer may change and must be disclosed.

Acceptance requires every detailed open-binding list to be empty, no unsound skip, exact caller-enrolled roots, full layer/head coverage, prompt/output and token continuity, and rejection of changed claims. A successful native run or empty broad binding.unbound list alone does not pass. This requirement follows the original complete-inference target; it does not relax it or change prior results. Existing classifier results remain unchanged.