Engineering Notes is a series of short, real lessons from building the Empoorio ecosystem in 2026. Each one cost something to learn.
What happened
PhoonoS's media pipeline encodes audio with Opus. A tuning change — -frame_duration 60 — was measured on a developer machine and saved 11 to 15 % of output size. It was applied to production. Production saved about 2 %.
Why
The developer machine ran a recent libopus (1.6). The production worker ran a Debian release that freezes libopus at 1.3.1, from 2019, where that parameter barely changes the encoder's behaviour. Same parameter, same input file, nearly four times the difference — from nothing but the library version.
Had the change not been re-measured inside the real container, the 11–15 % would have been reported as a fact. It would have been wrong by 4×.
The rule
Before accepting a performance or compression improvement, run the experiment inside the image that runs in production — docker run --platform … <deployed image> — not only locally. If the numbers do not match what production produces, the discrepancy is the finding.
What the finding led to
Updating the production base image (to one shipping libopus 1.5.2, x265 4.1 and SVT-AV1 2.3.0) by itself quadrupled the savings — before any encoder tuning at all. The optimisation that mattered was the one nobody was measuring.
Where else it applies
- A runtime weight benchmarked on a laptop says nothing about the validator's reference machine. EmpoorioChain's weights are hand-written pending exactly this measurement.
- A test that passes against a mock runtime says nothing about the live chain's accumulated state (see the near-miss the
try-runtimefast-forward caught). - A wallet feature verified in an emulator says nothing about a physical phone that refuses to install while locked.
The environment is not a deployment detail. It is a variable, and usually the biggest one.
From the PhoonoS media-pipeline work, August 2026.