Closed inside our stack
Campaigns run, measure and iterate on our own machine-learning engineering targets without a person between the steps.
Public status
Where we are, as of today
Gates, not dates. Each item below describes something observable that would have to be true for our claims to be credible, and where it actually stands today: including the ones that have not started.
Position
Rung 1
in progress
Rung 2
in progress
Rung 3
not started
Rung 4
not started
The evidence ladder
Progress outward is measured in what has been demonstrated, not what has been announced. We are working on the first two rungs; the rest are deliberately untouched.
Reproducible and economical
Promoted results survive re-running, and cost per verified result trends down rather than up as campaigns scale.
Useful to a partner lab
A research group reaches a conclusion faster with our methods, and can verify it independently without our help.
Contributing to a grand challenge
The computational stage of a real-world problem is genuinely compressed, with published methods and honest attribution.
A rung is only marked complete on evidence produced by the rung below it. We do not skip ahead because a claim would be more interesting.
Gates
What we hold ourselves to
Ten gates as of the date above. Evidence level refers to the strongest level any supporting result has actually reached: proposed, executed, measured or verified — never a level it has not earned.
| Gate | What it would demonstrate | Status | Evidence level |
|---|---|---|---|
| Closed inner loop | Campaigns on ML engineering targets complete propose → run → evaluate → iterate with no human between steps | in progress | measured |
| Re-run survival | Promoted candidates keep their advantage across seeds and repeat runs | partial | measured |
| Falling cost per result | Cost per verified result decreases as campaign size grows instead of rising with it | in progress | measured |
| Complete failure accounting | Every discarded trial carries a classification, so the search history is not a curated success list | partial | measured |
| Held-out agreement | The objective the proposer cannot see agrees with the one it optimises, on promoted candidates | in progress | measured |
| Budget containment | The scheduler refuses dispatch beyond a ceiling, so no campaign can overrun its budget in the first place | met | executed |
| Replayability | Every reported number can be rebuilt from its run record by someone who was not present | in progress | measured |
| Portability | A complete workflow — objective, harness, budgets, gates — runs where the hardware and data are not ours | not started | — |
| External verification | One partner reproduces a result of ours without our involvement and reports it independently | not started | — |
| A gate missed in public | We report a missed gate as missed, with the original target recorded as unmet | pending first test | — |
“Met” means the property holds as a system guarantee we can demonstrate. “Partial” means it holds for some campaigns and not others, and we are not going to average that away. “Not started” means exactly that — no partial credit for having thought about it.
History
What changed, and when
Every item on this page has a date attached to the change that introduced it. Nothing here was backfilled to look like a plan.
25
June 2025 — Company founded, thesis written
The observation the company was founded on: scientific progress is limited by a sequential loop, not by the quality of individual tools. No platform, no gates, one argument.
25
September 2025 — The word “autonomous” gets a definition
Four stages, three tests, and an explicit list of things we refuse to call autonomous. Published before any of our own systems could be described that way.
25
December 2025 — Measurement harness, first version
Frozen baselines, held-out objectives, repeated seeds and a failure taxonomy. The gate “Replayability” was added at this point, and has not been marked met since.
26
February 2026 — Parallel execution with enforced ceilings
Scheduling, preemption and recovery arrive with budgets as first-class objects. This is the change behind the one gate currently marked met.
26
April 2026 — The loop is pointed at our own stack
We become our own first customer, with four published commitments about what we will report either way.
26
June 2026 — Four evidence levels adopted
Proposed, executed, measured, verified. Every claim in this company now has to name the level it reached, which is why this page can have a column for it.
Now
September 2026 — This page published
Ten gates, their real statuses, and a commitment to update monthly. The next change to this page will be a gate moving — in either direction.
Limits
What this page does not say
A status page risks becoming a marketing page with checkmarks. Four things it is not claiming.
- No published results yet. Gates describe properties of our systems, not scientific findings. Our research page explains what we will and will not publish, and why the publication section is still empty.
- No external audit of these claims. Every status here is our own assessment. Gate nine exists precisely to build something better than that.
- No date on any gate. There is no schedule below this table, deliberately — see the roadmap page for why.
- No forecast of direction. Gates can move backwards. If one does, the entry stays and gains a note.
Disagree with a gate?
If you think one of these is too easy, too vague, or measuring the wrong thing, that is a useful conversation and we would rather have it than not.