From a five-lens review with three skeptics per finding:
- A certificate now names its kernel, implementation, member and corpus
generator (with the seed), all in its hash. Placement relies only on a
certificate that is for that kernel, implementation and member, and
earned on a corpus unless it is a CPU's own reference.
- A bound no longer stands in for a fidelity.
- The receipt and certificate encodings length-prefix every string, so
bytes cannot move between fields under one signature. The corpus digest
length-prefixes every output.
- The receipt carries the joules' acquisition class: required for a
non-zero figure, absent for zero, checked by the crate's label rule.
- statistical() normalises the reference, refuses a shot total past 64
bits, and cannot exceed one. compare() refuses a scale of 64 bits or
more. lanes::simulate refuses what the native simulator refuses.
- A known energy figure of any size comes before none.
- Detection asks the running CPU for its vector units on x86-64 and
aarch64 and says when it reports what the binary was compiled for.
- The docs claim no more than the gate kernel for lanes.
- The CLI labels by implementation and prints the generator; its doc says
fabric describes the host while its certificate agrees everywhere.
- wai.quantum.fabric joins the registry's unregistered receipt labels.
Native aarch64 and wasm32-wasip2 give the same certificate, corpus digest
included, for the same seed.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
36a39494b8