Today we published our first two technical notes and organized the website around two kinds of writing: research reports and updates from the lab.
Vesper Applied works on quantitative systems and computational engineering. Much of that work concerns the infrastructure beneath an experiment: data representation, accounting rules, implementation performance, and the checks needed to interpret a result correctly. The first publications examine two of those problems.
The first technical notes
Faster BPE Without Changing a Single Token examines fitting and encoding optimizations in our Python byte-pair tokenizer. The objective was to reduce repeated computation while preserving the learned artifact and exact token-ID sequence. The note reports local comparisons against our reference implementations, the associated memory overhead, and differential validation of the optimized paths.
This is implementation work. It does not establish a better tokenization objective or a performance advantage over established optimized libraries. Defining that boundary is necessary to make the timing results useful.
A Total-Return Index Is Not an Executable Portfolio addresses a different representation problem. A dividend-reinvested index describes growth under an assumed reinvestment policy. A portfolio simulation must also track share quantities, dividend entitlement, outstanding claims, and available cash. The note uses synthetic accounting cases to show where those representations diverge.
Neither publication is a strategy-performance announcement. Both concern conditions that need to be understood before broader experimental results can be interpreted.
Documenting methods and evidence
We are using the Research section for technical notes, experiments, benchmarks, and methodology. Each publication should identify the question being investigated, the assumptions used, the evidence available, and the limits of the conclusion. Publication type matters: an implementation note is not a peer-reviewed paper, and a passing software test is not independent validation of an investment result.
The tokenizer note includes a public numerical summary of its recorded measurements. It does not include the private source snapshots or corpus, so it is not a complete reproduction package. The accounting note uses illustrative prices rather than redistributing vendor data. We will state these availability limits directly rather than imply that every published result can already be reproduced externally.
Updates from the lab
The Blog will document project progress, engineering decisions, and shorter observations that do not require a full technical report. Research will contain the detailed methods and results. Keeping both visible gives readers a way to follow the work without confusing a progress update with experimental evidence.
For now, the public record begins with these two notes. They cover bounded problems, with concrete comparisons and explicit assumptions. Future updates can build on that record as additional work is completed and reviewed for publication.