ambient-compute + research-compute · V1 · 2026-08-18
Two machines, different chips, eleven stages to run. The submission never names a machine.
You write down what a piece of work needs. Every machine has already written down what it offers. Something matches the two, runs it, checks the answer, and writes down what happened. Below is one real job going through that.
One job, end to end
The whole systemSix steps. Under each is the real thing at that step, from a run that happened, and the one rule that step exists to enforce.
Narration
01 Asking, and splitting
You write down what the work needs: how many
cpus, how much memory, which capability it depends on. You do not write down
where it goes. A submission that names a machine is how a fleet quietly
becomes one computer again.
The study is a map over one axis of coordinates,
eleven values of stage, and each value becomes an independent
job with no ordering against any other.
02 Eligible, then placed
Fourteen hard questions are asked of every machine before anything is ranked at all: can it be admitted, is there headroom, is it reachable, does it carry the capability, the cpus, the memory. Ranking happens afterwards, among whatever is left. Whatever survived takes the work, and nobody chose which. The split beside this sentence is what was free at the time, not a policy and not a preference.
03 Checked, and written down
Two questions, asked separately. Whether the declared space was covered exactly once, which belongs to the scheduler. And whether each answer is right, which belongs to the science. A job exiting zero is a statement about execution and never about the result. Everything that happened goes under a content id you can ask about later. Resuming re-reads the outputs from disk and re-verifies them rather than trusting a stored status, so accepted work is never recomputed.
Step 3, watched deciding
The filter, re-runThe hard filter, run again against each machine's committed declarations, in the two states this fleet has been in. One declaration changes and the whole study moves.
What the ledger recorded
Eight terms, integers, and a tie broken on the name
Fixed integer weights, so two machines compare exactly and the answer reproduces anywhere it is asked. Nothing predicts a runtime or a cost.
Observed runtime outranks locality because locality is an inference about transfer cost worth about a second here, while a matching runtime sample is a direct observation of the same quantity for this work on these machines. The benchmark term is never consulted once real runtime has ranked them: the two are mutually exclusive by construction rather than by their weights comparing correctly.
Three and a half minutes, and the term that followed
In the accepted run one stage took 342.87 s where it landed against 127.42 s on the other machine, on a map that finished in about six minutes. It won on locality alone, because the ledger held no runtime history for that stage under that contract.
The evidence already existed: every job row carried the normalised manifest hash, the machine, the state and the duration. The change was one index over existing columns, one query and one named scoring term. No telemetry store was added.
The key matters more than the term. Runtime is filed under the normalised manifest hash, which includes the environment, rather than under the scientific id, which excludes it. When the same stage was reconfigured and became 35 times faster the old slow samples could not reach the new contract: they were filed under a different key. No outlier rule exists, and none is needed.
Recomputed above, recorded below
The two panels are a recomputation. The feasibility filter is a pure function of the manifest and the machines' stored records, both committed, so it re-runs exactly. Everything under what the ledger recorded is historical.
The ranking half is deliberately not recomputed: scores depend on live facts that were true at one instant and were never stored per stage. Rebuilding them would be a reconstruction. Five of the eleven durations are not in the committed log either, and are shown as unrecorded.
What has run on it
The evidenceEach kind of work here has a real map id behind it, and each is qualified by execution, never by a machine declaring it could.
The eleven stage names are not this system's. They are the step names of a separate, unpublished research study on jet sensing: spec and family are spectral and estimator-family passes, tri and resid the held-out and residual ones, and so on. The point is that eleven opaque steps were placed on two machines by their declared requirements and then checked, and that neither the placing nor the checking needs to know what any of them computes. A scheduler that had to know would be the wrong scheduler.
| work | where | what was proven |
|---|
| field |
|---|
Two questions, asked separately
What verified meansA job succeeding means it exited zero and every declared output was collected and hashed. That is a statement about execution and it is never treated as a statement about the science. The first check asks whether the declared space was covered exactly once, and belongs to the scheduler. The second asks whether each answer is right, and belongs to the science project.
The second check does not compare files. It walks each stage’s output structure and compares it value by value, and a leaf is one of those values: a single number or string at the end of a path, the smallest thing two runs can disagree about. A stage with 1,402 leaves is a stage whose answer is 1,402 comparable quantities, and bit-exact counts how many of them came back identical to the last bit. Naming the disagreement that precisely is what lets a stage be accepted while one leaf is excluded by name, rather than a whole file being waved through under a loosened tolerance.
| stage | verdict | leaves | bit-exact | worst deviation |
|---|
residdiag reads CONCLUSION because three of its recorded values are a stochastic training trajectory that diverges between the two microarchitectures. Those three are excluded by name in the manifest, so the five leaves the table counts for this stage are what is left after the exclusion, and one of those five still differs. They are excluded rather than absorbed by loosening a tolerance: a bound wide enough to swallow a different optimisation path would also swallow a real error, and a reader could not tell which had happened. What the stage concludes is bit-identical.
Exact identity, declared tolerance
Content ids stay bitwise over the whole artefact, always. Two results within tolerance keep two different content identities and both are printed. Nothing is normalised, rounded or re-collected to make two files match.
What a study may declare is an acceptance rule, versioned and inside the scientific id, so a study that accepts a difference is not asking the same question as one that refuses it. There is no default tolerance anywhere: an output that declares none is compared bitwise. A structural difference, a changed integer or string, a lone NaN or a missing output is a disagreement at any bound.
If you already know a computing centre
The fastest way inThe operator model is deliberately the one used at every computing centre, so the habits transfer in both directions.
| here | conventional HPC |
|---|
Where the analogy stops is the more useful half. None of the following exists here, and none of it is missing by accident:
One distinction inside that list matters more than the rest. What works is many independent tasks spread across machines. A single tightly coupled solver using several machines' cores at once is a different problem needing MPI, an interconnect and a multi-node allocation, and it should not be built until a real workload requires it.
What generalises
Nothing above is a fact about a particular number of machines. Feasibility before preference, unknown staying unknown, and a decision that arrives with its own reasoning attached are properties of the layer.
The system is deliberately small and deliberately finished. A feature here needs a trigger, not a precedent: every deferred item is published with the condition that would justify it, and "real systems have this" is not one of them.