Sweatshop for AgentsMeasure before you switch
Moving an app to a local model is a claim. The shadow log turns it into a number first. So far no local model has earned the switch — so every app still hears Jev.
SM/SWA · The machine room · Reflexmark
A second opinion
on every decision.
A lot of the fleet's work comes down to small verdicts: is this a complaint, which option fits, how good is it. Those questions go to Jev, a hosted decision model. Reflexmark stands in the doorway: the apps get Jev's answer as before, a local model answers the same question beside it, and the log shows how often they agree — before anyone moves an app to local.
Apps 15 ask through itMode all in shadowEngine one Rust binary, local model
What the apps ask
Yes or noIs this about a refund?
Pick oneWhich of these options fits?
Give a scoreHow good is this, on a scale?
From 15 appsEach with its own key
REFLEXMARKtwo answers · one logthe decision deskAsk · Route · Answer · Log · Queue
What comes back
Jev's answer, unchangedStraight back to the app
The local answerComputed after the reply went out
Agreement, measuredOn real traffic, not a benchmark
Every decision loggedThe question and both answers
An honest "busy"With a real time to retry
How it works
Five rules for the decision desk
A local model is cheap to ask and easy to overtrust. Reflexmark lets it answer every question — and makes it earn the right to be the one that's listened to.
- 01
Ask
Reflexmark speaks Jev's API, so an app moves over by changing one base URL. Nothing else in the app changes.
- 02
Route
Local, upstream, shadow or sampled: each app gets its own mode, and can be moved one at a time.
- 03
Answer
In shadow mode the app gets Jev's answer at once. The local model answers afterwards, so the app never waits for the comparison.
- 04
Log
Every question, both answers and whether they agree go into the decision log. The log is the product: it is the evidence for any switch.
- 05
Queue
One engine, one queue. When the wait would be too long, the caller hears so at once, with a real time to retry — and comparisons step aside first.
Why
Why a decision desk
Confirmed, not assumed
Agreeing with Jev is not the same as being right. A local model is only promoted on examples a person has confirmed independently — never on its own opinion.
Local by design
The engine runs on the owner's own Apple Silicon machine and never calls the network. One Rust binary: server, command line and agent tools in one.
The decision desk
Every verdict leaves a trace.
Reflexmark answers, compares and writes it down — so switching an app is a decision with evidence.