A phase is not a period of time; it is a condition. We only leave a phase when its exit condition is
demonstrated — and if a later phase turns out to be reachable earlier, the roadmap is wrong, not the
evidence.
Phase 1
in progress
Automate our own machine learning engineering
The loop closes on targets we control: training recipes, data pipelines, evaluation harnesses and
serving configurations. A person stays at the boundaries — objectives, budgets, anything with an
effect outside the sandbox — but not between the steps.
- Exit condition
- Campaigns complete propose → run → evaluate → iterate without human intervention, with every result replayable from its record.
- Why first
- Fastest feedback, unambiguous scoring, and we are the customer we can disappoint safely.
Phase 2
in progress
Make it economical
An automated loop that costs more per useful result than the manual alternative is a demonstration,
not a capability. This phase is about the arithmetic: throughput, verification cost, and the share
of trials that produce information at all.
- Exit condition
- Cost per verified result falls as campaign size grows, and the trend holds across repeated campaigns rather than one favourable run.
- Why it matters
- Everything downstream — partners, domains, grand challenges — is unaffordable until this is true.
Phase 3
not started
Make it portable
A complete workflow — objective, harness, budgets, review gates, run records — moves to an
environment where the hardware and the data are not ours. Portability is where hidden assumptions
about our own infrastructure surface.
- Exit condition
- The same campaign definition produces comparable, replayable results in a second environment without rewriting the harness.
- What it proves
- That the method is a method, and not a description of one particular cluster.
Phase 4
not started
Prove it with one partner, in one domain
One collaboration, in a domain where the computational stage is the bottleneck and the partner can
check our work without us in the room. Success is not a headline; it is a partner saying they
reached a conclusion faster and could verify it independently.
- Exit condition
- An externally verified result, with the method published and the failures reported alongside it.
- Why one
- A second domain before the first is verified teaches us nothing except how to run two unverified pilots.
Phase 5
not started
Broaden toward the grand challenges
Medicine, clean energy, clean water, cybersecurity and the tools of discovery — approached with a
method that has already been wrong in public at least once. Each new domain inherits the same
disciplines: measurable objective, held-out evaluation, enforced budgets, human gates at the edges.
- Exit condition
- Per domain, not per company: a compressed computational stage with methods published and attribution stated.
- Why last
- Because the consequences of being wrong there are not measured in compute.
Cross-cutting
continuous
Measurement, safety and honesty
Not a phase, because these cannot be deferred to a later one. The harness, the budget ceilings, the
human gates and the public gate list run alongside every phase above and are the reason any of the
others can be believed.