Governments and financial institutions run critical infrastructure on 800+ billion lines of legacy COBOL, powering $43T in annual commerce and 95% of ATM transactions. The GAO's latest report on critical federal IT (GAO-25-107795) highlights agencies spending the majority of their budgets maintaining aging mainframes while COBOL expertise vanishes. Today, AI coding agents are tasked with rewriting these legacy systems into modern languages. The Critical Challenge: AI Coding Agents Test Their Own Blind Spots When an AI agent writes both the modern port and its unit tests, they share identical blind spots. The tests merely verify that the code compiles and does not crash. Case in point: France's official DGFiP property-tax calculator. A public Java port co-authored with an AI agent passes all unit tests (4 of 4 PASS). Yet when run on real 2018 commune data, the AI port silently fails on 408 of 408 communes in Ain, and 35,268 of 35,389 communes nationwide (99.7% error rate). On Commune 01001, the COBOL computes 137,415 euros while the AI port outputs 144,041 euros (+6,626 euros discrepancy). Two silent bugs passed testing: fee constants swapped (3% and 8%), and the household waste tax (tctom) omitted from the taxable base. The Solution: Tare | The Original Program is the Answer Key To tare a scale is to zero it out. Tare turns the original legacy program into an immutable deterministic oracle. Tare executes the original COBOL under GnuCOBOL, executes the target port on identical inputs, and weighs every record and field. Target Users & Interaction Tare is built for enterprise software engineers, government modernization teams, and systems architects migrating mission-critical legacy code. Developers interact with Tare directly inside the IBM Bob IDE using Bob's custom tare mode. When Bob attempts to run git commit, Tare's PreToolUse lifecycle hook intercepts the action and refuses the commit (Exit 2) if the port has not balanced against the COBOL answer key or if code was modified after weighing. When mismatches occur, Bob calls Tare's explain MCP tool, which traces differing fields directly to the exact COBOL source statements and target code lines. Why It Is Unique & Creative Unlike existing solutions that rely on LLM self-evaluation or synthetic unit tests, Tare provides zero-hallucination mathematical parity. It replaces subjective code reviews with cryptographic Parity Receipts. Demonstrated Impact & Scale - IBM Bob IDE Proof: In a live recorded take, Bob spent 3.74 Bobcoins. Blocked by Tare, Bob deployed two parallel subagents to diagnose fee rates and totals, corrected swapped constants, resolved the missing tctom tax, and achieved 408 of 408 bit-exact parity. - National French Tax: Scaled across all 35,389 French communes, eliminating all 35,268 errors. - US Medicare Hospice Pricer: Tested on 5,000 real CMS FY2021 claims ($23B annual program), resolving 140 discrepancy bugs found even in CMS's official Java release. - 100% reproducible: Run offline in seconds via python -m tare reproduce --offline.
Category tags: