Claims you could disprove
Every claim is stated so that a specific observation would refute it: a protected span that changed, or a run that compressed when it should have abstained. Claims too vague to be refuted that way are left out.
Technical paper · August 30, 2026
The paper is where the numbers live
A technical paper accompanies the open-source release. It describes what TekMyra does, states its claims in a form that can be proved wrong, and reports measured results alongside the run records they were computed from.
In the paper
Three commitments
It is easy to write a flattering paper about a compression system by picking the friendly cases, reporting only the good average, and leaving out the runs that were skipped. The three rules below are what we held ourselves to instead. They stay on this page, along with the method section that follows, as the paper’s table of contents.
Figures backed by receipts
Reported figures are derived from emitted run records, and the method for re-deriving them is written out. The reference bundle ships with the release so the same records can be produced on your machine.
Abstention counted in
Runs where TekMyra declined to compress stay in the denominator of every headline figure. If they were filtered out, a system that abstains often would look much better than it actually performs.
Method
How receipt-verification works
Every figure is computed from what the system itself wrote down while it ran. Nothing is measured by an observer standing outside it.
-
The gateway emits a record per run
Each record holds the tokens that went in, the tokens that came out, the route the request belonged to, the spans that were protected, and whether the run abstained.
-
Figures are computed from records only
Totals in the console and in the paper are arithmetic over those records. They are not drawn from a second source or adjusted afterwards, and there is no sampling step where inconvenient runs could drop out.
-
Anyone can re-derive them
Run the reference benchmark against the shipped fixtures, read the resulting records, and check them against what the paper reports. If your numbers disagree with ours, please file a bug report.
The command
The reference benchmark is a single entry point, documented on the
docs page: python -m tekmyra.benchmark compressor
against the shipped fixture directory.
Availability
Published with the source
The paper is published alongside the repository and the packaged release (August 30, 2026), so that every claim in it can be checked against code from the moment it is readable. A citable archived copy comes next, and we will link it here once it exists rather than publish a URL we invented.