IBM Bob changes the code. Tarsius protects the business behavior.
A formal pre-generation governance layer that captures unwritten legal carve-outs, locks decisions into an immutable SHA-256 Merkle chain, and mathematically proves equivalence via the BETH Differential Oracle.
// Written during 2010 CS-4471 Settlement
IF suspendedFlag = 'Y' AND pOrderType = 'DISC';
wsResult = 'A'; // Overrides prior rejection!
ENDIF;
The Conflict: Technical documentation Section 2.2 explicitly says "No exceptions exist for suspended accounts". Naive AI would delete this block as dead code.
Class action carve-out CS-4471
Plan-7 customer discount lock
Demographic Cliff
72%
RPG devs aged >50 (Fortra 2026)
Mainframe Exits
>70%
Modernizations fail (Gartner 2026)
Unconstrained LLM
9% – 19%
Edge equivalence (arXiv:2605.17535)
Tarsius Governed
100%
Behavioral Equivalence Rate (BER)
Standard code-generation LLMs suffer from the Self-Fulfilling Test Fallacy: they read the outdated spec, omit the unwritten carve-out, and generate matching unit tests that pass 100% green—until production crashes with millions in regulatory fines.
Section 2.2: Suspended Account Enforcement
"Accounts with status 'S' (Suspended) are blocked from all transactions. This rule applies universally. No exceptions exist for suspended accounts under any order or transfer type."
Every naive AI assistant (and new junior engineer) strictly follows this document, deleting any exception as "obsolete dead code".
// Reason: Ticket CS-4471 (2010 Class Action Settlement)
// DO NOT REMOVE without legal sign-off.
IF suspendedFlag = 'Y' AND pOrderType = 'DISC';
wsResult = 'A'; // Overrides prior rejection!
ENDIF;
Written 11 years ago during class action settlement CS-4471. Removing it violates compliance (the exact failure mode of TSB Bank's £48.65M penalty).
Test how IBM Bob 2.0 halts autonomous hallucination and initiates a 60-second Socratic inquiry with the human SME before emitting modernized code.
// Ready. Select a legacy workload and click 'Run Bob Analyzer' to begin AST discovery.
See the exact contradiction lines side-by-side. See why unconstrained AI deletes the rule and how Tarsius creates the cryptographic bridge.
IF-BLOCK (Lines 77–81)
Tacit Carve-Out
Section 2.2: Universal Suspension
Strict Prohibition
// Naive AI follows outdated spec Section 2.2
if (customer.status === 'S' || customer.isSuspended) {
return OrderResult.REJECTED; // Universal block!
}
// Carve-out deleted as "obsolete dead code"!
Result: 4 Regressions in BETH oracle. Wrongfully blocks valid customer orders ($128k/day revenue loss).
// Sealed by DEC-00004 (Class Action CS-4471)
if (customer.isSuspended && order.type === 'DISC') {
return OrderResult.APPROVED; // Preserved exception!
}
if (customer.status === 'S') return OrderResult.REJECTED;
Result: 100% Behavioral Equivalence (BER). Zero regressions; fully compliant with legal audit.
Test real enterprise workloads live in your browser. Experience firsthand how unconstrained AI drops critical edge cases, and how Tarsius guarantees 100% equivalence.
66.7%
8 / 12
4 Regressions
BLOCKED: High Risk
Executing BETH differential oracle across legacy and modern AST...
| Vector ID | Test Scenario | Inputs | Legacy Oracle | Modernized Output | Status | Differential Trace |
|---|
Tarsius ships as a native CLI tool. Initialize governance in any legacy repository and launch the local Triage Studio instantly.
# 1. Install Tarsius CLI via one-line curl:
curl -fsSL https://raw.githubusercontent.com/farhank15/Tarsius/main/install.sh | bash
# 2. In any legacy repo, initialize .bob/ custom modes and MCP tools:
tarsius init
# 3. Launch the local React Triage Studio at localhost:8321:
tarsius studio
# 4. Execute BETH differential oracle across both legacy stacks:
tarsius verify
tarsius init
Automatically creates .bob/custom_modes.yaml, registers MCP tools in .bob/mcp.json, and initializes the binding contract .bob/RISK-CONTEXT.md.
tarsius studio
Opens the desktop/local React Triage Studio at http://localhost:8321 where engineers review 🔴 Must-Review rules, sign off with legal citations, and monitor the SHA-256 Merkle chain.
tarsius verify
Executes differential assertions against both legacy runtimes (RPGLE & COBOL), guaranteeing 100% Behavioral Equivalence before code merges.
Why unconstrained AI code-generation fails in regulated enterprise architectures, and how Tarsius enforces mathematical behavioral equivalence.
| Modernization Dimension |
Unconstrained AI (Standard Code-Gen)
|
Tarsius Governed (Bob 2.0 + BETH Oracle)
|
Regulatory & Business Safeguard |
|---|---|---|---|
| Tacit Knowledge Capture | ❌ Spec-Only Bias: Deletes unwritten code exceptions as "dead code" (Self-Fulfilling Fallacy). | ✅ Dual-Source Contradiction: AST vs Spec parser detects conflicts before code generation. | Protects legal carve-outs ($3T daily financial volume). |
| Human-in-the-Loop Triage | ❌ Zero Triage: Generates code autonomously without human sign-off on contradictions. | ✅ Zero-LLM 5-Condition Classifier: Auto-approves 70% low-risk; isolates 🔴 Must-Review for SME sign-off. | Mandatory human accountability before code generation. |
| Mainframe Enterprise Scope | ⚠️ Narrow / Batch Only: Cannot handle transactional CICS screens or complex DB2 routines. | ✅ Dual-Workload Verified: IBM i (RPGLE ORDVAL) & IBM z/OS (CICS COBOL 2,069 LOC). | Production-tested on tier-1 transactional banking codebases. |
| Runtime Equivalence Proof | ❌ Coverage-Only or Output Blindness: Leaves 81%–91% edge cases unverified (arXiv:2605.17535). | ✅ BETH Differential Oracle: 100% Behavioral Equivalence Rate (BER) across all test vectors. | Eliminates TSB Bank-style £48.65M system failure risks. |
| Audit & Compliance Trail | ❌ Ephemeral Logs: Markdown or terminal logs easily tampered with or lost. | ✅ Cryptographic SHA-256 Merkle Chain: Tamper-evident hash ledger tied to human SME signatures. | Guarantees compliance with EU AI Act Article 12 & OCC audits. |
| Agent Hallucination Guard | ❌ Context Drift: AI subagents drift over multi-turn prompts and hallucinate requirements. |
✅ Binding Risk Contract: Injects immutable .bob/RISK-CONTEXT.md directly into Bob agent loop.
|
Ensures strict deterministic adherence to verified rules. |
| Automated Test Integrity | Unspecified or mock testing. | ✅ 49 / 49 Passing Tests (100% Green): Full automated verification in CI/CD pipeline. | Zero mock assumptions; fully reproducible runtime evidence. |
Aligned with EU AI Act Article 12: every rule approval produces a cryptographically sealed chain of custody tied to an attributed human decision.
All 4 decision blocks cryptographically linked via SHA-256 parent hash tree.
Plan-7 legacy customers retain locked-in pricing per class action settlement 2009-CV-118.
Hash: de2769d389d807f500f4dbfacf674c209d4adfc58e6606f1c8888ef154e06fdc
Chain: 1b94cc7f1ab4767d417b2e8ecffae38e76f2020f945a2233ecdd35aa2aefb30b
11-year unwritten carve-out for suspended accounts placing DISC orders approved under CS-4471.
Hash: c1bd6eeebdbe5f4df399891ecdbf4dd15881a3b3f0f5d721031ea5b51c0f6dfb
Chain: 1004ad65a78c8e3018820e25c23bbe9cfb9b98802a5f288f77efe20c4a7d72ef
Tarsius extends IBM Bob 2.0's 3-tier architecture with zero external orchestration dependencies.
Analyzes legacy ASTs against technical specs during Bob Plan Mode. Initiates a 60-second interview with SMEs prior to writing code.
5-condition deterministic classifier. Automatically approves standard rules (70%) while isolating high-risk legal carve-outs to 🔴 Must-Review.
Compiles decisions into .bob/RISK-CONTEXT.md, strictly binding Bob's subagents to prevent context drift and hallucinations.
BETH runs differential assertions between legacy rules and modernized outputs, mathematically proving 100% Behavioral Equivalence.