Grace proof

Here is the record. Reproduce it.

Grace is built on proof, so this page is engineering evidence rather than testimonials: what was tested, what happened, what Grace refused to do, and what has not been exercised yet.

verified repairs

0

verified repairs

false fixes accepted

0

false fixes accepted

unfamiliar real projects

0

unfamiliar real projects

repairs made anything worse

0

repairs made anything worse

Self-healing

A repair, start to finish.

Detect, trace the origin, calculate what it touches, repair inside boundaries, verify independently, keep the undo, write the record.

  • Independent verification of every repair
  • Rollback tested — clean restore when a fix did not hold
  • Refusal tested — Grace hands over rather than guessing
  • Fake repairs presented for approval, and rejected
Repair record

    Under two seconds. No human intervention. Every step reconstructable, and reversible if it hadn't held. Timeline illustrates a repair of the kind measured in our published testing.

    Reach

    Consequences calculated before anything changes.

    Grace holds how things connect, so before a change is applied it knows what that change reaches — which services, which functions, which tests. When the reach exceeds granted authority, Grace defers instead of acting.

    0,

    source files reasoned across in one project

    0/

    developer-written faults repaired

    0/

    planted silent logic faults caught

    0%

    of in-scope faults repaired

    One change, understood

    Before Grace touches a line, it has already traced what the line touches.

    In one real project, a single proposed change was traced across the full connected world before any edit was made. That is the difference between repairing software and editing text.

    0,

    files in the connected world

    0,

    relationships between them

    0

    dependent computations traced

    0 ms

    to know what one change reaches

    Determinism

    Same evidence, same decision.

    A decision is not a mood. Given the same state and the same policy, Grace reaches the same decision — and the decision can be replayed from the stored record rather than taken on trust.

    • Decisions are recorded with the evidence they were made on.
    • Replaying the record produces the same decision, not a similar one.
    • If the record cannot reproduce the decision, the decision does not stand.

    input state A

    policy X

    decision 17384

    — replay —

    input state A

    policy X

    decision 17384

    Identical on replay

    Behaviour

    Six rules Grace holds to.

    These are not slogans. Each one corresponds to a mechanism you can test.

    Grace does not guess when evidence is missing.

    Unknowns are named

    Grace names what she cannot see.

    Blind spots

    Grace understands consequences before acting.

    Consequence calculation

    Grace never changes something without knowing how to undo it.

    Rollback

    Grace proves what she changed.

    Records

    Grace learns only from verified outcomes.

    Learning

    Limitations are published alongside the results: categories not yet exercised are shown as unexercised rather than passed, and independent third-party validation is still ahead of us.

    Early access

    Test Grace against your own systems.

    Request early access