doe:
leave the losses visible.
A small reversible-coder note where specification comes before prestige, reconstruction before ratio, and adverse results stay attached to the artifact.
The artifact is primary.
DOE-001 records a reversible experiment through its format, encoder/decoder contract, fixtures and observed measurements. No compression ranking is inferred from a single corpus or a convenient winning row.
Method
Name the format. Define the inverse. Record corpus identity, environment, command line, wall time, memory and exact reconstruction. Keep packaging and multifile behavior separate unless they are explicitly part of the tested format.
Evaluation
Start with reproducible fixtures and compare against obvious baselines. Bytes, encode wall, decode wall, memory and reconstruction are separate dimensions; a result that wins one axis does not silently win the others.
Adverse results
Inputs where the coder expands, slows down or becomes awkward remain in the note. Negative results narrow the design space and are therefore part of the engineering asset.
Graduation rule
If a mechanism survives exactness and benchmark gates, extract the useful mechanism into a standalone module. Do not preserve accidental coupling merely because the prototype once worked.