Who is fak for? A persona gallery

TL;DR: people adopt a tool when they see themselves in the pitch. This page lists the kinds of person who actually land on fak and gives each one a single quote-ready sentence for why they specifically care, plus the one door to walk through first. Find your row, quote the line, follow the link.

This is dimension E — Social proof & community of the concept-popularization epic. The roster below is the same one the persona-readiness scorecard grades, so the personas here are the ones fak commits to serving, not an invented list. Each pitch is a per-audience cut of the canonical pitch ladder; if a claim here and a claim there ever disagree, the pitch ladder wins.

Every number is witnessed and traces to BENCHMARK-AUTHORITY.md and CLAIMS.md. These are pitches written for each persona, not quotes from one: nothing here claims a real user, a market share, an unrun benchmark, or a novelty the 0/29 prior-art audit refutes.

Consume and build on it

Free-tier solo dev. Downloads a prebuilt binary and runs it. Will not clone, build from source, or read the docs; spends seconds, not minutes.

Download one static binary and watch it refuse a dangerous tool call in under a minute — no key, no GPU, no toolchain.

Start here: the README.

App developer / vibe-coder. Building an app or agent and wants to point an existing harness (Claude Code, Cursor, an MCP client) at fak with no hand-wiring.

Point the agent you already run at fak with one base-URL change, and every tool call it makes gets checked against a default-deny floor before it runs.

Start here: the integration index.

Backend integrator. Embeds the binary in front of a model server inside a real service and needs a stable interface plus a way to extend it.

Install fak in front of your model server and build on a frozen interface, adding a leaf instead of forking the kernel.

Start here: the getting-started guide.

Operate and trust it

Infra / platform engineer (SRE). Runs fak serve in production and wants a container image, metrics, health checks, rate limits, and a deployment story.

Run the same static binary a developer runs, now with /metrics, /healthz, rate limits, and a deploy guide for Docker, Compose, and k8s.

Start here: the deployment guide.

Security engineer. Evaluates the capability floor and the containment claims before trusting fak in front of an agent, and reads the threat model.

Refusing an irreversible action never depends on catching the attack: the capability was never granted, and every deny cites a code from a closed vocabulary into a tamper-evident journal.

Start here: the threat model in SECURITY.md.

Study and decide

ML researcher. Wants to reproduce fak’s determinism and benchmark results bit-for-bit, study the kernel from the notes, and cite the work.

Reproduce the bit-identical determinism witness offline and cite the committed benchmark data, not a screenshot.

Start here: the reproduce-offline packet.

Benchmark / eval engineer. Runs fak’s benchmarks and compares the numbers against the field (vLLM, SGLang, TensorRT-LLM, llama.cpp).

Run the fan-out benchmark yourself and read fak’s honest position against the field: the tuned ~4.1× less work on a 50-turn × 5-agent session, never the naive multiplier.

Start here: the bench plan.

Evaluator / decision-maker. An engineering manager or tech lead deciding in about ten minutes whether fak is real, honest, and worth a deeper look.

See in ten minutes what is shipped, what is still simulated, and how fak compares — every headline number traced to a commit, no overclaim.

Start here: the README.

Contribute and extend

Open-source contributor. Wants to add a feature and ship it green without tripping an enforced guard, with the rules stated up front.

Add your feature as a leaf and ship it green, with the enforced rules stated up front so the guard teaches instead of ambushes.

Start here: the contributor contract.

AI coding agent. An autonomous coding agent (Claude Code, Codex, Cursor, an MCP client) that lands in the repo cold and must discover, adopt, and build on fak with no human in the loop.

Land in the repo cold, read AGENTS.md and llms.txt, and adopt fak without a person handing you the steps.

Start here: the agent entry point AGENTS.md.

The shared pitch under all of them

Every row above is the same core claim aimed at a different reader: fak treats every agent tool call like a syscall, so the model proposes and the kernel disposes. One static Go binary makes the same agent loop safer (a default-deny capability floor plus result quarantine), cheaper (a witnessed ~4.1× less work than a tuned warm-cache stack, with max|Δ| = 0 KV eviction), and faster (a ~362 ns in-process decision). None of the primitives are new; the assembly into one drop-in binary is the point.