Beyond BI · for agents

A frontier model wrote this code. We verified it — or refused it.

Frontier coding agents write more code than anyone can review, and they fail silently — a plausible function that breaks on the input nobody tried. The hard part isn’t generating code; it’s knowing which code is actually correct. The engine that never serves a wrong KPI runs on code too: it independently proves the model’s output against a declared spec — with math, not trust — or refuses with the exact input that breaks it.

Proven, not trusted

Three independent methods: execution-differential against an independent oracle, property / metamorphic invariants, and SMT where the problem is decidable. The model’s output is evidence — never the verdict.

It refuses — with a counterexample

When the code breaks, you get the exact input that breaks it and which method caught it. When a property is undecidable or out of scope, it refuses rather than bless it. No silent pass.

Model-independent, drop-in for agents

Any frontier agent calls one endpoint or MCP tool with code plus a spec and gets back a verdict and the evidence. The verifier is the same regardless of which model wrote the code.

Proof · how it decides

The verdict is ours, not the model’s.

The model’s confidence doesn’t enter the decision. Either an independent check confirms the code against the declared spec, or it’s refused with the reason. Here is a worked example — one function the verifier proves, and one it refuses with the input that breaks it:

Refused — with a counterexample

A median() the agent wrote returns the wrong value for an even-length list. Execution-differential against an independent oracle finds the breaking input:

input [1, 2, 3, 4] → expected 2.5 · got 3

caught by execution-differential — refused, not blessed.

Proven — over the declared domain

A double(a) the agent wrote, declared equal to 2·a. SMT proves the two are equal for every integer — symbolically, not by sampling:

method SMT · proven for all inputs (symbolic)

proven for the whole domain — not a sample, not a score.

PROVEN covers the full declared domain; VERIFIED covers the inputs that were checked. Both are independent of the model — the figures above are reproducible, not a vendor score.

Live benchmark · real run

12/12 planted bugs caught. 0 slipped through.

We ran 25 realistic functions across 5 families — each in a correct and a buggy variant a coding agent might write — through the engine. Every bug was caught with the exact input that breaks it, and no correct function was wrongly refused.

12/12

planted bugs caught

0

silently passed

0

false alarms

13/13

correct ones proven

familythe bug a model wrotecounterexamplecaught by
arithmeticforgot to negate negativesn=-20execution-differential
arithmetica*b+c instead of a*(b+c)a=-1, b=5, c=5smt
arithmeticmin/max swappedx=-4execution-differential
number-theorymissing n<2 guard (1 and 0 reported prime)n=0execution-differential
number-theoryloop n>9 drops final digitn=1execution-differential
list-aggregatereturns first element, not the maxxs=[0, 1]execution-differential
list-aggregatecounts odds (v%2==1)xs=[0]execution-differential
list-structurereturns input unchangedxs=[0, 1]execution-differential
list-structurealways returns Falsexs=[0, 0]execution-differential
list-structurereturns input unsortedxs=[1, 0]execution-differential
stringchecks only first/last chars='aaba'execution-differential
stringcounts 'b' as a vowels='b'execution-differential

Reproduce: python -m cuenta.codeverify_bench

We never say “safe” or “bug-free” — only verified against this declared spec, or refused.