Untrusted Authors, Trusted Answers: A Calculus of Fidelity-Graded Translations

2026-07-03 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

A new calculus of fidelity-graded translations, implemented in the hurdy-gurdy platform, addresses the challenge of ensuring correctness in multi-language program translations. This system models translations as a graph, where each edge (language pair) has a declared fidelity grade (e.g., "reproducible", "checked", "proved"). The calculus defines how these grades compose, how per-run checks re-establish faithfulness, and how independent translation routes corroborate each other. A key finding is the asymmetry between witness-carrying answers, which are self-certifying at the source, and universal answers, which rely on grades and branch agreement. The hurdy-gurdy platform, comprising 13 languages and 13 pairs around BTOR2 and SMT-LIB reasoning hubs, was largely built by independent LLM agents. Its evaluation, current as of July 2026, demonstrates per-pair coverage, branch agreement, end-to-end witness replay, and certified unreachability, highlighting the architecture's role in catching defects.

Key takeaway

For Machine Learning Engineers building or verifying complex multi-language systems, you should adopt a fault-tolerant architectural approach to translation correctness. Explicitly define fidelity grades for each translation step and implement independent, cross-checked routes. Prioritize witness carry-back for existential proofs, as these are self-certifying. This strategy allows you to scale language support and integrate LLM-generated components reliably, shifting trust from author to architecture.

Key insights

Trust in multi-language translations stems from architectural checks and graded fidelity, not author provenance.

Principles

• Translation trust is architectural, not authorial.
• Fidelity grades compose by weakest link.
• Independent translation routes corroborate each other.

Method

Define language "pairs" as commuting squares with declared projections; check faithfulness per-program; compose squares by pasting with projection-compatibility.

In practice

• Implement translation "pairs" with decidable square checks.
• Utilize LLM agents for translator and interpreter development.
• Employ source-level witness replay for existential proofs.

Topics

• Fidelity-Graded Translations
• LLM Code Generation
• Program Verification
• Multi-language Systems
• Formal Verification
• Translation Validation

Code references

• cksystemsgroup/hurdy-gurdy
• YosysHQ/riscv-formal

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Faithful Autoformalization of Natural Language Assertions — Monty reliably converts natural language assertions into executable formal specifications by combining LLM generation with testing-based validation and semantic conformance checks.
• Diversifying to Verify: When Task-Equivalent Programs Differ in Verifiability — Task-equivalent program implementations vary significantly in automated verifiability, and diversity aids verification success.
• Cycle-Consistent Neural Explanation of Formal Verification Certificates — Trained neural architectures can explain formal verification certificates more effectively and efficiently than general LLMs.
• The Model Called a Tool I Never Gave It — When LLM tools run on uncontrolled client devices, the architecture shifts from prompt engineering to trust and reliability.
• Do LLMs Game Formalization? Evaluating Faithfulness in Logical Reasoning — LLMs can produce unfaithful logical proofs despite high compilation rates, highlighting a gap between validity and faithfulness.
• Training Large Language Models for Self-Explanation Faithfulness — Reinforcement Learning can directly optimize LLM parameters to generate faithful self-explanations.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.