research papers

GPU results are not reproducible across vendors — the WGSL spec leaves rounding, reassociation and fusion unspecified. REPROGPU pins byte-identical SHA-256 outputs for a declared integer/fixed-point kernel subset via a 320-bit superaccumulator, Philox4x32-10 and integer SHA-256, with a cross-adapter conformance harness as the witness.

Derived content — hints, explanations, difficulty labels — is consumed with positive evidence only: a passing check says nothing about what it would take to overturn it. This paper attaches auditable warrants to derived claims, so contestation has a ledger to answer to instead of a boolean.

A build cache key describes the inputs a project declared, not the inputs the build actually read. KeyFuse detects undeclared reads — env vars, working-directory drift — with minimal collision witnesses and applies conservative key repair, turning cache trust into a falsifiable claim.

A learner's repeated Run on an auto-graded exercise usually leaves one bit — pass or fail — and no memory of the previous program. We define a per-edit behavioral signature: deterministic detection of which edits actually changed behavior, on a hidden basis derived from the exercise's own tests.

Auto-graded exercises ship a reference program and a handful of tests — a program can pass every shipped test and still be wrong. We define the alibi distance as the minimum number of hidden tests a wrong-but-passing program can be distinguished by, then close that aperture corpus-wide.

Spaced-repetition schedulers for coding practice see more than a pass or a fail: DeepForge observes how many test runs failed before the pass, which hint tier was revealed, and when the solve happened. Ladder-graded spacing turns those execution signals into a scheduling ladder that respects real difficulty.