Reflex · KatGPT

Typed decisions, calibrated honestly, on your machine.

Reflex answers choice, score, and yes/no questions about a state you give it — with calibrated confidence and a first-class abstain: when it has no signal, it declines instead of guessing. No trained weights on the decision path, no AI service, no telemetry. Loopback-only, microsecond-tier, bit-deterministic.

macOS / Linuxcurl -fsSL https://raw.githubusercontent.com/gist-rs/reflex/main/install.sh | sh
Windowsiwr -useb https://raw.githubusercontent.com/gist-rs/reflex/main/install.ps1 | iex
Homebrewbrew tap gist-rs/tap && brew install riir-reflex
Scoopscoop bucket add gist-rs https://github.com/gist-rs/scoop-bucket && scoop install riir-reflex
G1 calibration beats the conformal-naive floor — 8/9 suites G2 p99 0.06 ms / decision set G4 zero allocs on the hot path laya parity: 88/88 forwards, top-1 1.000, drift ≤ 3.1e-6

Three steps

The engine never phones home. The browser cannot reach it until you open the door for exactly one origin.

1 · Install

One binary. riir-reflex --version prints its compiled feature set and warns if the build is incomplete.

2 · Run with the door open

RIIR_REFLEX_ALLOWED_ORIGIN=https://reflex.gist.rs riir-reflex — the engine listens on 127.0.0.1:7331 and accepts browser calls from this page only.

3 · Play

Type a state and a question. Answers come back in microseconds, with confidence — or an honest abstain.

Playground

This page talks to YOUR engine on YOUR machine — nothing is uploaded anywhere.

checking for a local engine…
—

Measured, with the losses published

The arena exists to make the trade visible: the modelless lane is ~10³–10⁴× faster and deterministic; the open-weights laya lane is the accuracy heavyweight. Same questions, both lanes, per-task tables — full benchmark.

0.475 ms
modelless p50 per 2,000-question typed-decision set
suitemodellesslaya·bestlaya p50
typed decisions0.2780.7444493 ms
AG News0.2580.950388 ms
emotion0.2950.593170 ms