From vericoding to verified projects

Bangalore System meetup July 2026

Transcript

1. From vericoding to verified projects By Sachin Singh
2. https://arxiv.org/abs/2604.16584v1
3. What is vericoding Certified program synthesis (a.k.a. vericoding) is the process of automatically generating a program, its formal specification, and a machine-checkable proof of program/specification alignment from a task description given in a natural language
4. What does verified means • Verified” does not mean “perfect” or “bug-free in every possible sense.” It means: Specific properties were proved under specific assumptions. • Testing asks, “Did the program work for these examples?” • Verification asks, “Can we prove it works for every permitted input?” • For example, if the specification only says “return a non-negative number,” a program that always returns zero could be verified—even if the user actually wanted an absolute difference.
5. Two problems identified by the paper • Generated specifications may be too weak or too strong. • Verification ecosystems are fragmented across testing, automated verification, and interactive theorem proving.
6. Velvet (the DSL that solves the above problem) Velvet combines three verification modes: • Property-based testing finds concrete counterexamples. • Automated verification handles routine obligations. • Interactive Lean proofs address harder residual goals.
7. Certified Program Synthesis with a Multi-Modal Verifier (https://arxiv.org/abs/2604.16584v1)
8. Why decomposition matters For every task i, the worker attempts to establish Preᵢ(x) → Postᵢ(x, fᵢ(x))
9. Why decomposition matters Decomposition offers: • Smaller specifications • Smaller proof obligations • Less context for each agent • Parallel execution • Isolated failures
10. Local correctness versus global correctness But it introduces a new theoretical problem: compositionality. Correct(A) ∧ Correct(B) ∧ Correct(C) != Correct(A ∘ B ∘ C)
11. Local correctness versus global correctness Global correctness additionally requires: • Compatible interfaces • Valid assumptions between modules • Shared invariants • Correct data flow • A composition theorem
12. Can a single-method vericoding be scaled into a controlled project workflow while preserving each worker’s verification guarantees?
13. decomposition, parallel execution, artifact validation, and module integration https://github.com/sachinkumarsingh092/multi-agent-leetproof
14. Thank you!