VeriTool

Runtime trace invariant verification for multi-agent LLM systems. Z3 + Lean 4.

v = Verifier()

v.register_tool("deploy", action_type="DEPLOY", resource_fn=lambda a: a.get("env"))
v.register_tool("build",  action_type="BUILD")
v.register_tool("test",   action_type="TEST")

v.add_invariant(OrderingInvariant("DEPLOY", ["BUILD", "TEST"]))

deploy = v.wrap(my_deploy_fn)
build  = v.wrap(my_build_fn)
test   = v.wrap(my_test_fn)

Every tool call is checked against accumulated action traces before execution. Violations are blocked with counterexamples.

Architecture

Tool Call → Verifier.wrap() → Z3 encodes all 4 invariants as SMT constraints → UNSAT (permit) / SAT (block + counterexample)
                                                                                        ↑
                                                                              Lean proves:
                                                                              Z3 encoding ⊧ invariant semantics

Four invariant types, all encoded as Z3 satisfiability queries:

Each check creates a z3.Solver, encodes the violation condition as a SAT query, and returns a counterexample model on violation. This uniform SMT pipeline enables formal soundness proofs in Lean.

Why Z3 + Lean?

Existing policy engines (OPA/Rego, AWS Cedar) check single actions against static rules. They cannot enforce invariants that span multiple agents across sequences of actions — e.g., "no deploy without prior build and test."

Z3 encodes the accumulated action trace and invariants as SMT constraints. Lean proves the encoding is sound — each invariant is defined as a Lean Prop with nil/snoc structural induction theorems showing that a Z3 permit preserves the invariant.

Related work: Lean4Agent (UIUC, June 2026) uses Lean 4 for design-time workflow verification. VeriTool targets runtime interception with Z3 for fast SMT checking and Lean for proof of encoding correctness.

Usage

from verifier.verifier import Verifier
from bridge.invariant import OrderingInvariant, ExclusiveAccessInvariant

v = Verifier(agent_name="ci-bot")
v.register_tool("deploy", action_type="DEPLOY", resource_fn=lambda a: a.get("env"))
v.register_tool("build", action_type="BUILD")
v.register_tool("test", action_type="TEST")
v.add_invariant(OrderingInvariant("DEPLOY", ["BUILD", "TEST"]))
v.add_invariant(ExclusiveAccessInvariant("DEPLOY"))

deploy = v.wrap(deploy_to_production)
build  = v.wrap(run_build)
test   = v.wrap(run_tests)

assert build() == ...
assert test() == ...
assert deploy(env="prod") == ...  # only permitted if BUILD + TEST completed

v.reset()  # clear trace for a new session

Project Structure

bridge/          # Action trace model + invariant types + Z3 encoding
verifier/        # Generic Verifier with wrap()
Lean/Trace.lean  # Soundness proofs for invariant encoding (3 invariants proved)
policy_store/    # Versioned policy store + audit trail
tests/           # pytest suite
benchmark/       # Scalability, throughput, 3-way OPA comparison

Tests

make test      # 14 tests, all pass
make verify    # Lean theorem compilation check

Benchmark Results

Full results in BENCHMARK_RESULTS.md.

Detection Rate

Latency & Throughput

VeriTool is ~60x faster than OPA across both stateless and history modes.

Scalability

All invariants remain sub-millisecond up to 1000-trace entries. Per-check latency grows near-constantly (2-3x from size 1 to 1000).

Formal Soundness

Three of four invariant types are proved correct in Lean via structural induction:

• ordering_invariant — nil/snoc theorems
• exclusive_access_invariant — nil/snoc theorems
• approval_invariant — nil/snoc theorems

Each theorem's hcase (safety condition) corresponds exactly to the Z3 encoder's violation check.

Reproduce

make all   # Lean verify → tests → benchmark → 3-way OPA comparison

Requires OPA installed:

curl -sL -o /usr/local/bin/opa https://github.com/open-policy-agent/opa/releases/download/v1.2.0/opa_linux_amd64_static
chmod +x /usr/local/bin/opa

Docker

docker build -t veritool .
docker run --rm veritool make all

The Dockerfile pins Python 3.13, Lean 4.15.0, OPA 1.2.0.

License

Apache 2.0