Research
We publish the method before the results.
This page carries our claims policy, the fidelity map, how we pre-register studies, and the questions we are currently working on. No results and no magnitudes appear here: a study publishes only after it clears its gate, and the gate is described before the study runs.
What we allow ourselves to say.
This is the policy our own external materials are written against, published as it stands rather than summarized.
Claim types we permit
Mechanism
What causes an effect, and through which part of the system. "The collective stalls because of queueing at this tier."
Relative
How one configuration compares with another, on the same workload and the same fabric.
Envelope
A range with its uncertainty stated, widening as the projection moves away from what was measured.
Claims we will not make
- Any improvement, utilization, or recovered-throughput magnitude we have not produced under a cleared gate.
- A point prediction at an unvalidated scale without an uncertainty envelope.
- Packet-level fidelity for anything produced by a flow-level or analytical stage.
- Reliability, failure, restart or checkpoint-loss prediction.
- Results reported by named provider rather than by network archetype and experimental condition.
Measured, predicted and inferred are separated in every artifact and on every page that describes one. When results eventually publish, the misses publish alongside the successes.
Three different claims that are frequently merged.
Reach
How large a simulation the engine will run at all.
Validated
What has been checked against measurement, which is a smaller set.
Economically searchable
What you can afford to sweep across a design space, which is smaller again.
Collapsing these into one number is the most common way capability gets overstated in this field.
What we are working on, and what could stop it.
The home page says installed capacity underproduces and cites other people’s numbers for it. This is what we are doing to produce our own — and the conditions under which we would publish a null.
Pre-registration, and the gate a study has to clear
Hypotheses and thresholds are declared before any results are inspected, and validation data is held out. A study that misses its threshold is reported as a miss, on the same terms as one that clears it.
Does small-scale measurement determine full-scale training performance, and can controlled simulation isolate the variables the published record confounds?
Method published; results gated- Method
- Controlled simulation across scales with the recipe held fixed, pre-registered thresholds, and held-out validation against measured runs.
When two healthy workloads share a fabric, who pays — and can it be predicted before the shared run?
Under a feasibility gate- Method
- Paired replay of concurrent workloads against a shared fabric model, with the prediction declared before the shared run is observed.
- The gate
- This study depends on concurrent multi-trace replay, which is not an established capability. Its feasibility study has not been approved. If the gate does not clear, the question does not get answered and we will say so.
No results, no magnitudes and no capability claims appear on this page. Results publish only after their gate passes, and are labeled as ours when they do — with their error bars and their misses.
How to work with us on it.
Co-design a study
Shape the question and the thresholds before anything runs.
Contribute ground truth
Measured runs against which a projection can be checked.
Brokered validation access
Hardware time for the cross-check rung of the validation ladder.
Review the method.
If you are an analyst, a benchmark partner, or an engineer deciding whether these results would be worth citing, the method is what to examine. We will walk through it, including the parts that have not cleared their gates.