zk0 is building a way for robot operators to train one shared model together,
while each operator keeps their own data on their own machines. This public roadmap shows the
direction and the order of work — at a high level, on purpose. It starts with tangible early
wins, and speeds up only as real traction earns it.
What zk0 offers: replicated benchmark evaluation — the same tests, run the
same way, so results can be checked and compared — and auditable records of what was
evaluated and when. That claim is deliberately narrow: evaluation shows how a model performs
on the benchmarks tested. It is not a guarantee of performance in every setting, and it is
not a certification of anyone's data or contributions.
The short version — easy to pass along
Many robot teams each have a little private data — too little to train a great model alone.
zk0 lets them train one shared model together without handing their data to a central pool.
Every model version is checked with replicated benchmark evaluation, and the results are kept
in auditable records anyone in the network can inspect.
Early wins first. Prove the shared model on shared benchmarks, with a small group of committed operators.
Trust grows with evidence. Standing in the network is earned slowly, through results that others can re-check — never bought, never assumed.
Acceleration is earned. Bigger steps — paid tiers, wider access, stronger accountability — arrive only when usage and results justify them.
01
Two ways to grow
The forced path demands fast, big early wins — and pays for them with a plateau. The zk0 path
accepts a deliberately flat start, then accelerates as trust and model quality compound.
Tap a legend item to isolate a curve.
zk0 path — acceleration is engineered
A small retained core and one clear question to answer first. Then the loop compounds:
a better shared model draws more contributors, whose more varied experience makes a
better model. Each phase makes the next one cheaper and safer to attempt.
Forced path — the spike that caps growth
A big early push promises results before there's evidence behind them. When the promised
results don't hold up, trust drains and growth flattens — no matter how much was spent
getting there.
02
The engine: a compounding loop, not a launch
Growth doesn't come from a bigger day one. It comes from a loop that gets stronger every time
it turns — and from a stock of trust, built on checkable results, that accumulates alongside
the model.
Plan on the committed core, not the signups
The planning number is a small core of operators who stay and keep contributing.
Projections are built on them — signups and registrations are vanity numbers.
Traction unlocks the next step
Each phase below begins when a signal appears — operators retained, groups forming,
real demand — not when a date arrives. Steps taken early are friction; taken late,
they're overdue.
Trust compounds like the model does
Records start accruing from the first round, long before standing in the network is
worth anything. By the time standing matters, there's a long, inspectable history
behind it.
03
Four phases, in order, as the stakes grow
This is the public shape of the plan. The detailed research, designs, and implementation
work behind each phase stay in zk0's private systems and are shared in depth with network
members as they join. What is public is the sequence — and the rule that no phase starts
before the one before it has earned it.
Phase 1Now
Prove It
community · free
Tangible early win
A published, repeatable comparison: the shared model vs. training alone, on the same
benchmarks — with the records to re-check it.
One shared model, trained across a small core of operators, data staying on their machines
Replicated benchmark evaluation of every model version, run the same way each time
Auditable records of what was evaluated, when, and with what result
Standing begins — a plain record of participation and results, visible in the network
Moves on when
The comparison is published and a core of operators is retained round over round.
Phase 2On signal
Grow It
small groups · peers training together
Tangible win
Groups of similar operators training together — and benchmark results that hold up as
the network grows beyond the founding core.
Small peer groups form around similar robots and tasks
Standing starts to count — a longer, checkable history carries more weight
Evaluation keeps pace — benchmarks run every round, so quality is watched as the network scales
Moves on when
Groups are forming on their own — and standing becomes valuable enough to protect.
Phase 3On signal
Sustain It
paid tier
Tangible win
A paid tier that funds the shared infrastructure — with accountability that matches
money changing hands.
Paid membership for operators who rely on the shared model
History gates access — standing earned in earlier phases is the entry ticket
Stronger accountability for provable misconduct — consequences tied to evidence, never to suspicion or votes
Moves on when
Money flows — real demand exists, and accountability needs to match it.
Phase 4On signal
Open It Up
larger organizations · enterprise
Tangible win
Larger organizations can join under the governance, review, and privacy options their
legal and compliance teams require.
Shared governance — oversight beyond any single operator
Independent review — formal audits of the evaluation and records
Stronger privacy options for organizations that need them
Moves on when
Enterprise and legal demand actually arrives.
04
Trust: earned slowly, lost quickly
Standing in the network is deliberately asymmetric. It builds slowly, from benchmark results
and records that anyone in the network can re-check — and it can be lost quickly when there
is clear evidence of misconduct. That shape is what makes standing hard to fake and costly
to abuse.
Why the asymmetry works
Anyone can create a fresh account — but nobody can fake a long history of results
that others have already re-checked. Slow accrual makes standing expensive to fake;
quick loss makes it irrational to burn on one act of misconduct.
What the records do — and don't — say
The records show what was evaluated, on which benchmarks, and what the results were.
They do not certify that any operator's data is genuine, complete, or safe, and
benchmark results do not guarantee how a model will perform on different robots,
tasks, or environments.
The honest limit
Evaluation and record-keeping can prove statements about a training run —
what was submitted, when, and how the resulting model scored. They cannot by
themselves prove where data came from or that it came from a real robot. zk0 does
not claim otherwise.
That is why the public claim stays narrow: replicated benchmark evaluation and
auditable records — a process you can inspect, not a promise you have to take
on faith.
P—1
Slow is a strategy, not a delay
Tangible wins first, bigger steps as traction earns them. A flat start with a retained
core beats a spike that promises big early results and caps everything after.
P—2
Claims stay as narrow as the evidence
zk0 publishes what its evaluation actually shows — benchmark results and auditable
records — and nothing broader. If a claim can't be re-checked, it doesn't go on this
page.
P—3
Acceleration is engineered
Trust and model quality compound; launch spikes don't. Every phase is unlocked by a
traction signal, and each one makes the next cheaper to run.