How We Think
Proof is not the presentation after the work. It is part of the work.
A system can produce a confident recommendation without producing a trustworthy one. The difference is the quality of the measurements, the honesty of the assumptions, and the ability to trace a conclusion back to what supports it.
Our research approach is grounded in a few deliberately demanding principles. They guide the questions we ask and the standards we aim to meet; they do not substitute for demonstrated performance.
-
01 / PRINCIPLE
Measure the operating reality.
A benchmark without workload context can answer the wrong question beautifully. We begin with what the data and the surrounding system actually need to do—and which constraints cannot be compromised.
-
02 / PRINCIPLE
Compare alternatives, not headlines.
There is no universal winner across every dataset, access pattern, or computing environment. Meaningful evaluation requires viable comparisons and the willingness to recognize when the status quo performs best.
-
03 / PRINCIPLE
Follow the cost wherever it moves.
Less storage is not automatically less expense. Savings may shift into decoding, additional reads, compute time, network transfer, or operational complexity. The outcome that matters is the one the whole system experiences.
-
04 / PRINCIPLE
Make the reasoning inspectable.
A useful conclusion should remain understandable after the original experiment. We value reproducible measurement, explicit assumptions, traceable evidence, and clarity about the conditions under which a recommendation holds.
-
05 / PRINCIPLE
Treat uncertainty as a result.
Not every experiment produces a clean answer. When the evidence is incomplete, alternatives are too close, or a critical constraint is at risk, restraint is not a failure of intelligence. It is a sign that the decision is being taken seriously.
Questions worth answering
When does a smaller dataset actually cost less to operate?
Which tradeoffs change under random access or partial reads?
What evidence is sufficient to recommend a layout change?
How should a decision be reconsidered when the workload or environment changes?
These questions are not rhetorical. They help define the research agenda behind Marvin.
Research with a practical destination
Marvin gives these principles a concrete application: compression and physical layout decisions for AI data. The ambition is not to make technical decisions sound more certain. It is to make them more deserving of confidence.