Apparatus-driven companion to the Bun Zig-to-Rust port. Operationalizes the predictions and falsifier-grade discriminator articulated in Doc 702 of the RESOLVE corpus into concrete tools for AI-assisted cross-language code translation generally.
Doc 702 reads AI-assisted cross-language code translation as a Pin-Art bilateral under SIPE-T threshold conditions, with the Bun runtime's experimental claude/phase-a-port branch (Anthropic, May 2026) as the live exemplar. The apparatus yields four predictions and one falsifier-grade discriminator. This repository builds the tools that test them.
Genuine semantic-preserving translation between source and target languages should exhibit three signatures simultaneously and sharply at the convergence point:
- T1 — Substrate-internal semantic preservation. Differential testing against the original binary on production-distribution inputs. Independent of the translation pipeline.
- T2 — Compositional invariance. Fuzz-driven input-output equivalence under perturbations within the operational distribution.
- T3 — Pipeline repeatability. Translation output stability across model variants, seeds, and prompt orderings.
Decoupling — most diagnostically T2 partial via translated tests with T1 failing on differential testing — is the vibe-port failure mode the community has flagged.
differential/— T1 harness. FFI-bridged or reference-implementation-based comparison between source-language binary and target-language binary on production-distribution inputs.fuzz/— T2 harness. AFL/libFuzzer-style mutation fuzzers applied symmetrically to source and target builds with output comparison.repeat/— T3 harness. Translation-pipeline repeatability runner. Executes the same translation step multiple times under varied conditions and diffs outputs.welch/— Welch-bound packing diagnostic. Scans translated Rust for unsafe-block density, lifetime-annotation density, and other idiomatic markers; compares against a baseline of mature Rust crates.l2m/— Per-file translation context-budget calculator. Given substrate effective context, Porting.md size, and per-file LOC, predicts the L2M-bounded knee point above which translation quality degrades non-linearly.
- P1. Per-file translation quality degrades non-linearly above an L2M-bounded LOC threshold.
- P2. Phase A → Phase B compilation success follows a SIPE-T threshold curve, not smooth growth.
- P3. Unsafe-block density per region predicts latent semantic drift under stress fuzzing.
- P4. Translated test suites systematically miss semantic drift the translation introduced.
Repository scaffold. Direction is pending. The keeper sets the work; this README is the orientation pointer.
- Doc 702 — AI-Assisted Cross-Language Code Translation — the apparatus-level reading
- Doc 700 — L2M Resolved — the capacity-bound apparatus
- Doc 696 — Discrete Geometry — the Welch-bound packing apparatus
- Doc 699 — Three-Signature Simultaneity Test — the T1·T2·T3 origin
- Doc 541 — Systems-Induced Property Emergence — including §3.6 rung-1/rung-2 distinction and §7 Fal-T5
To be set by the keeper.