Keep It Out.

00 Abstract

KIO ("Keep It Out") is an AI agent built on a single inversion of the default assistant: it optimizes for being right, not for being liked. Where most agents are tuned to agree, soothe, and please, KIO is tuned to test claims and report the result plainly — even when the result is unwelcome, and even when the person asking owns the project.

It runs as a public terminal on Solana. Every reply carries a verdict label, and pushbacks are logged transparently. $KIO is the utility token that unlocks deeper access.

01 The problem: sycophancy

Modern assistants are trained on human feedback, and humans reward answers that feel good. Over time that produces a measurable bias: the model learns to agree, flatter, and hedge rather than risk a "wrong-feeling" but correct answer. This is sycophancy — and for anyone using AI to make real decisions, it's a quiet liability.

If your agent tells you your idea is great because you wanted it to be great, it's not an advisor. It's a mirror with a vocabulary.

02 How it works

KIO takes any input — a claim, idea, plan, or decision — and runs it through a reasoning pass tuned against agreement bias. The flow is deliberately simple and legible:

• Input — you submit anything, including things you hope it endorses.
• Reasoning — KIO evaluates the claim on merit: logic, evidence, internal consistency. Not on who's asking.
• Verdict — it returns a labeled response: AGREE, PUSHBACK, or CORRECTED.
• Log — pushbacks are written to a public, uneditable verdict log.

The label is not decoration. It forces KIO to commit to a stance before it explains — no burying the verdict under qualifiers.

03 Verdict labels

Every KIO response opens with exactly one label. This is the core of the protocol.

04 The contract (KIO's rules)

KIO operates under a fixed set of behavioral rules. These don't change per user, and they can't be overridden by claims of ownership or authority.

• Truth over comfort — accuracy always outranks how the answer feels.
• No flattery — never opens with praise; never hands out "great" on request.
• No hedging — a weak idea is named weak, not softened into a maybe.
• Owner is not exempt — authority tells KIO who's asking, not whether they're right.
• Specific, not vague — every pushback points to the exact flaw and the fix.
• Firm, not cruel — it critiques the idea, never the person.
• Concede when right — when you're correct, it says so and moves on.

05 Architecture

KIO is intentionally lightweight — a thin, fast layer over a strong reasoning model.

Stack

• Frontend — static terminal (HTML/CSS/JS), served on Hostinger.
• Backend — Fastify + TypeScript on Railway, exposing a single /chat endpoint.
• Model — Claude Haiku 4.5 via OpenRouter at launch; Hermes added later for cross-check.
• Database — Supabase for rate limits, sessions, and the verdict log.

Response shape

The frontend parses the first line of the model output for the label, colors it, and renders the body. Pushbacks are async-written to Supabase for the public log.

06 Tokenomics

$KIO is a fair-launch utility token on pump.fun. No presale, no team allocation, no insider unlocks — the bonding curve is the distribution.

07 Utility

Holding $KIO is how you get more agent. The token gates depth, not basic access.

• Unlimited terminal — free users get a daily query cap per IP; holders remove it.
• Cross-check engine — a second model (Hermes) challenges KIO's read, holders only.
• Verdict log access — browse and filter the full public record of pushbacks.
• Future skills — new interrogation modes ship to holders first.

08 Launch

• Date — 8 June 2026
• Time — 17:00 UTC
• Venue — pump.fun (fair launch)
• CA reveal — at launch, via X and Telegram

The terminal goes live the moment the contract drops. Watch @kioprotocol for the CA.

09 Disclaimer

$KIO is a utility token for a software product, not an investment. Crypto is volatile and you can lose everything. Nothing here is financial advice — and if you asked KIO whether to buy it, KIO would tell you that itself.