These results use ISA_A64_xml_A_profile-2026-06_mc (Armv9.6-A). The sweep was split into 256 independently verified shards covering all 2³² encodings.

Agreement with the specification on whether an encoding is valid:

The large validity gap has identifiable causes. Unicorn tests validity by executing an instruction and observing traps, while the other oracles decode without execution. A small number of regions are also affected by decode-time UNDEFINED conditions that the compiled specification oracle does not evaluate. Silica records these limitations instead of smoothing them out of the result.

Instruction text

Comparing rendered mnemonics and operands is much more expensive than recording a validity bit. Silica therefore evaluates text on a deterministic sample of 1,000,000 words drawn from 1,266,064,016 candidates where all four oracles consider the encoding valid. This is a sampled result and is deliberately kept separate from the exhaustive validity figures.

The sample is useful for locating normalization and presentation work; it does not claim exhaustive coverage of every textual rendering.

Explore the results

The sweep engine produces a large research dataset. silica-scope is the companion terminal app for making that dataset approachable. It opens a finished Silica artifact directory and lets you browse headline metrics, inspect the 256-shard encoding map, filter disagreements, look up any 32-bit word, and read the filing-ready reproducers.

Install it from PyPI with Python 3.11 or newer:

Then run it from a Silica checkout or point it at an artifact directory:

silica-scope is a pure-Python reader with no native decoder dependencies. It does not launch the exhaustive sweep, and it handles the repository's smaller published artifact set gracefully. See the terminal reader guide for the panes, keyboard controls, and artifact discovery options.

How it works

The high-volume path is written in Rust and calls each decoder in-process. It stores one bit per encoding per oracle, which keeps the exhaustive comparison compact and makes disagreement a direct bitmap operation. Crashes are bisected to the exact instruction word.

Python handles specification compilation, normalization, reporting, and the independent verification layer. Artifact schemas, sampling rules, and known limitations are documented in docs/formats.md.

Reproducing the work

Create the pinned environment and check that the required local inputs are available:

Arm's XML specification is not vendored because of its license. silica doctor reports where Silica expects to find it and any other missing prerequisites.

To run the full pipeline from a prepared checkout:

This is a full 2³² sweep, not a quick smoke test. It produces the compiled oracle, shard records, validity bitmaps, disagreement corpus, published metrics, reproducers, and a stable SHA-256 result hash.

Trust, but verify

Seven independent verifiers recompute the project's claims from raw artifacts. They do not trust a generated summary, and each verifier has a fixture proving that it detects the defect it guards against. There is no skipped or provisional state.

Pinned decoder versions and a freshly recomputed result hash make separate runs comparable. The verification goals and their current status are recorded in GOALS.yml.

Scope

Silica currently covers base A64 and Advanced SIMD decoding. SVE, SVE2, SME, A32/T32, RISC-V, assembler round trips, and general execution testing are outside the v1 study.

The closest inspiration is Sandsifter, which explores x86's variable-length instruction space. Silica applies the same spirit of systematic skepticism to AArch64, where fixed-width encodings and an independent specification allow a complete, adjudicated comparison.