91%
of in-scope problems repaired
Measured across real, unfamiliar projects.
The sequence on the right runs in order, every time. If verification fails, the repair is discarded and the prior state is restored exactly.
Every stage is recorded. If verification fails, nothing is kept.
Aggregate results from a benchmark run on real, widely-used software Grace had not seen before. Project names and code are withheld.
91%
of in-scope problems repaired
Measured across real, unfamiliar projects.
109
independently verified repairs
Verification separate from the repair itself.
0
cases of anything made worse
No observed regression across the run.
10/10
silent faults caught
Software still ran; the answer was wrong.
13
unfamiliar projects
No prior exposure, no tailored setup.
100%
clean rollback when unproven
Prior state restored exactly.
Custody
Repairs are made locally inside your own systems. Grace does not ship your code, failures or logs away to be fixed by an external API, so the evidence trail and the system itself stay inside your boundary.
Fault diagnosis, bounded repair, independent verification, rollback and refusal — measured, with figures above.
Rejection of repairs that only appear to work, including deliberately fake fixes.
That Grace repairs anything, anywhere. Work outside granted boundaries is refused, not attempted.
That verification is possible without something to check against. Where there is no ground truth, Grace escalates.
A symptom isn't necessarily the origin.
Nothing crashed here. The totals were simply wrong. Generic illustration — not customer code.
Before
function finalTotal(amount, discount) {
return amount - discount; // discount treated as cash
}After — verified
function finalTotal(amount, discount) {
return amount * (1 - discount); // discount is a rate
}The repair changed one line inside the granted boundary. It was then checked against known-correct behaviour by a separate process, and everything that depended on it was re-tested before the change was kept.
Self-healing