What's proven. What's tested. What's still being built.
Every capability on this page carries an honest label, because a roadmap that flatters itself is worthless to the person relying on it.
Dates on completed work are what happened. Dates on unfinished work are targets, and we say so.
Every milestone, at a glance.
Milestone, date and status in one table. The sections below give the detail and the evidence.
| Milestone | Date | Status | Evidence |
|---|---|---|---|
| Bounded understanding of a real codebase | Completed — Q1 2026 | Proven | Exercised on 13 real, widely-used projects, including one with 4,342 source files. |
| Verified self-repair with clean rollback | Completed — Q1 2026 | Proven | 109 verified repairs, 0 observed cases of anything being made worse, 100% clean restore when uncertain. |
| Catching silent, non-crashing mistakes | Completed — Q1 2026 | Proven | 10 of 10 planted silent faults caught and repaired; every fake repair rejected. |
| Refusal and escalation instead of guessing | Completed — Q1 2026 | Proven | Every recorded outcome lands in one of two columns: repaired and confirmed, or safely referred. |
| Impact analysis before any change | Now — Q2 2026 | Tested | Behaves reliably in controlled runs; being measured across more real projects before it is called proven. |
| Learning that carries between sessions | Now — Q2 2026 | Tested | Retention and reuse verified in controlled runs; long-horizon behaviour still being measured. |
| Early access with your own bounded authority | Target — Q3 2026 | In development | — |
| Focus Domain Intelligence for your work | Target — Q3 2026 | In development | — |
| Continuous watch over live systems | Target — Q4 2026 | In development | — |
| Shared understanding across a team | Target — Q4 2026 | In development | — |
| Cross-domain reasoning beyond software | No date yet | Research direction | — |
| Long-horizon autonomy | No date yet | Research direction | — |
What the labels mean
Four honest states.
Proven
Measured on real, unfamiliar software with results published on the Evidence page.
Tested
Works in controlled runs with known answers; not yet measured across enough real projects to publish as proven.
In development
Being built now. Behaviour exists in part; dates are targets, not promises.
Research direction
Deliberate direction with open questions. No date, and we will say so plainly until there is one.
Proven
Measured, published, repeatable.
Measured on real, unfamiliar software with results published on the Evidence page.
Completed — Q1 2026
Bounded understanding of a real codebase
Grace reads a project it has never seen, builds persistent context across files, sources, relationships and history, and holds that understanding between sessions.
Exercised on 13 real, widely-used projects, including one with 4,342 source files.
Completed — Q1 2026
Verified self-repair with clean rollback
Faults are diagnosed, repaired, and checked independently of the process that produced the repair. When a fix does not hold, everything is restored exactly as it was.
109 verified repairs, 0 observed cases of anything being made worse, 100% clean restore when uncertain.
Completed — Q1 2026
Catching silent, non-crashing mistakes
The dangerous bugs are the quiet ones: the software runs, it just returns the wrong answer. Grace compares behaviour against the last version known to be correct.
10 of 10 planted silent faults caught and repaired; every fake repair rejected.
Completed — Q1 2026
Refusal and escalation instead of guessing
When evidence is insufficient, Grace hands the problem to a person with what it found, rather than attempting a change it cannot justify.
Every recorded outcome lands in one of two columns: repaired and confirmed, or safely referred.
Tested
Works in controlled runs. Not yet published as proven.
Works in controlled runs with known answers; not yet measured across enough real projects to publish as proven.
Now — Q2 2026
Impact analysis before any change
Before acting, Grace traces what a change touches — the chain of consequences, the blast radius, and the areas it cannot see — and shows that map to you.
Behaves reliably in controlled runs; being measured across more real projects before it is called proven.
Now — Q2 2026
Learning that carries between sessions
Outcomes, successful approaches, failures and open unknowns are preserved, so the same ground is not re-learned every time.
Retention and reuse verified in controlled runs; long-horizon behaviour still being measured.
In development
Being built now, with target dates.
Being built now. Behaviour exists in part; dates are targets, not promises.
Target — Q3 2026
Early access with your own bounded authority
You set exactly what Grace may touch and what always needs your approval, and every consequential action leaves a receipt you can inspect.
Target — Q3 2026
Focus Domain Intelligence for your work
Grace narrows to the domain you actually operate in — your systems, your vocabulary, your constraints — instead of behaving like a general assistant.
Target — Q4 2026
Continuous watch over live systems
Moving from repair on request to noticing a problem forming, verifying it is real, and either fixing it inside its bounds or raising it with evidence.
Target — Q4 2026
Shared understanding across a team
One accumulated understanding of your systems that several people can work against, with each person's authority kept separate.
Research direction
Honest about what we don't know yet.
Deliberate direction with open questions. No date, and we will say so plainly until there is one.
No date yet
Cross-domain reasoning beyond software
Applying the same evidence-bounded loop — understand, act inside limits, verify, learn — to work that is not code. Open questions remain about how verification works when there is no test to run.
No date yet
Long-horizon autonomy
Work that runs over weeks rather than minutes, with authority that stays bounded the whole way. We will not put a date on this until the safety case is settled.
How this page is kept
Labels move in one direction only when the evidence says so.
- Nothing becomes "proven" until it has been measured on real software Grace had never seen, with the result published.
- A missed target date stays visible and gets re-dated rather than quietly deleted.
- Research directions carry no date at all, rather than a comfortable guess.
Early access