Predict consequences across code and delivery workflow.

The engineering problem

Reviewers must infer blast radius, test coverage, operational criticality, and delivery delay from fragmented tools. The simulator joins those signals into one evidence-backed analysis and makes the next engineering decision explicit.

What the current baseline combines

• Repository modules and dependency edges
• Changed files and module criticality
• Module-to-test mappings and test duration
• Synthetic historical CI failure associations
• Workflow events, reruns, and review state

Deterministic graph traversal predicts impact. Inspectable scoring ranks tests and risk factors. Workflow heuristics estimate time-to-green, and centralized action weights compare counterfactual interventions. Numerical outputs do not come from a language model.

Where the learned system goes next

The intended latent-twin architecture encodes repository structure and workflow events into a fused state. A PR and candidate action act as interventions; dedicated heads predict impact, tests, calibrated risk, ETA, bottlenecks, and action value. The differentiator is unifying technical and workflow consequence prediction in one action-oriented system—not an unsupported claim of category novelty.

GPT-5.6 can turn a compact evidence pack into a grounded explanation. It does not generate numerical predictions or replace traceable evidence, and a labelled deterministic provider takes over when the API is unavailable.

Current limitations

This version uses one fictional repository, prepared PRs, synthetic history, module-level dependency paths, and heuristic calibration. Its evaluation verifies that the product and metrics pipeline work end to end; it does not establish production accuracy or research validity.

Privacy and provenance

Public GitHub analysis uses repository metadata and changed source text. Uploaded projects are processed in best-effort request-scoped temporary storage and are not executed. GPT explanations receive a compact evidence summary rather than the full repository. This prototype is not a production security product.

Application version: .