Evidence
A differential fuzzer: random functions and contracts inside the subset, every verdict checked by running the real Python on every input in [-5, 5] for each parameter. Our own testing, not independent.
Before the fix (seed 1, 1,500 contracts)
| Outcome | Count |
|---|---|
COUNTEREXAMPLE_OK | 605 |
CRASH_COUNTEREXAMPLE | 645 |
PROVED_BUT_CRASHES | 70 |
PROVED_OK | 180 |
PROVED_BUT_CRASHES: Z3 said "proved for all inputs" and Python raised ZeroDivisionError on one.
CRASH_COUNTEREXAMPLE: the counterexample made Python raise, but was described as an ordinary violation.
After the fix
| Outcome (seed 1) | Count |
|---|---|
COUNTEREXAMPLE_OK | 251 |
PROVED_OK | 146 |
RAISES_COUNTEREXAMPLE_OK | 1103 |
| Outcome (seed 7, fresh cases) | Count |
|---|---|
COUNTEREXAMPLE_OK | 290 |
PROVED_OK | 135 |
RAISES_COUNTEREXAMPLE_OK | 1075 |
Every verdict correct: proofs hold on every grid input, every counterexample breaks the postcondition in Python, and
every "raises" counterexample raises in exactly the part it names. The fuzzer would report FALSE_PROVED,
FALSE_COUNTEREXAMPLE, PROVED_BUT_CRASHES, WRONG_FAILURE_KIND or ENGINE_ERROR; none occurred.
Hand-written probes
| Case | Before | After |
|---|---|---|
| module-level abs = lambda v: -1, contract result >= 0 | PROVED (false: returns -1) | REFUSED |
| parameter named abs | PROVED (false: TypeError) | REFUSED |
| local max = 3 | PROVED (false: TypeError) | REFUSED |
| @staticmethod decorator | accepted (decorators ignored) | REFUSED |
| return 10 // x, any postcondition | PROVED for all inputs (false: raises at x=0) | counterexample x=0, function_raises |
| postcondition result // x == 1 or x == 0 | PROVED (false: raises at x=0) | counterexample x=0, postcondition_raises |
| precondition 10 // x > 0 | PROVED | counterexample x=0, precondition_raises |
| x == 0 or result // x == 1 (short-circuit) | PROVED | PROVED (correct) |
| if x: on an int | engine exception (HTTP 500) | REFUSED: mixes int and bool |
| function with a docstring | REFUSED (unsupported statement) | PROVED; certificate re-checks offline |
Tests
tests/test_vericoding_soundness_2026_11_02.py 18 + test_vericoding_keystone_routes/scan 19 passed; vericoding-gate/test_vericoding_gate.py 9 passed.
Files
| vericoding-evolution-2026-11-02.json | d13400a61211d91ab15ad4c696c31717121243a6d8ef55fd0f5f1517a299a044 |
| vericoding-fuzz-soundness.py (the fuzzer; it needs the engine to run, so it is published to be read) | b0cb60716461684a73a741e1648eff4eab239f8178247b4e6b701eeb5216f21c |
FIRME AI CORPORATION · CAIN-42 is a Byzantine trust fabric. Numbers marked "our own check" were run by us, not by an independent party. Nothing on these pages is a certification.