M12: Capstones

[tamnd]

Twelve weeks. All three capstone tracks, each with a skeleton, a specification, a test suite, a grader, a scorecard generator, and a reference solution that CI keeps green.

The tracks are a verified pass suite that runs at E0, a backend for the RV32-Toy target that needs a local machine, and a mini LLVM written in Rust from the blueprints alone.

Exit criterion. Each track has a reference solution that produces a published scorecard.

Tasks

• Track A, the verified pass suite: twelve required passes, six optional, four gates
• Track B, the RV32-Toy backend: ISA specification, simulator, twelve week plan
• Track C, mini LLVM in Rust: scope, skeleton, grading
• A reference solution per track, kept green by CI
• The capstone-smoke weekly job

Gates

• Every track has a reference solution producing a published scorecard
• Track C's reference implementation has found and filed at least five blueprint defects
• At least two external readers have completed a track and published a scorecard
• The capstone-smoke weekly job is green

The Track C defect count is the interesting one. If somebody builds a working mini LLVM from the blueprints alone and finds nothing wrong with them, either the blueprints are excellent or the implementer read LLVM's source when they got stuck. The retrospective works out which, and a zero here is a warning sign rather than a triumph.