Engagement metrics are filling with synthetic activity, and feeds rank whatever flows in. A ranking that could weight engagement from present humans would measure attention that exists. The metric would mean something again.
We updated verification on entros.io.
Quiet microphones used to stall a round before the capture heard you. The round now listens at the level your mic delivers, so a soft voice keeps things moving.
A failed attempt also gets a guided retry with your slot returned.
Live now: entros.io/verify
By December, every EU member state must offer citizens a digital identity wallet. Identity infrastructure is arriving, government-grade and regulated. Presence is the adjacent layer: a session needs to know that someone is there before it asks who.
Imperva's 2026 bot report puts automated traffic above 53% of the web. Human traffic is down to 47% and falling. The assumption that a visitor is a person stopped being the safe default a while ago.
Every Entros verification now passes its spoken phrase through a learned anti-spoofing model, enforced in production. Synthesis artifacts are caught at validation; the audio itself is discarded as soon as the check completes.
Second World War intercept operators learned to tell one Morse operator from another by rhythm alone. They called the rhythm a fist. The lesson aged well: the way a person moves is individual, recognizable, and hard to hand to someone else.
GPT-6 Astra vs Entros verification.
Every attempt so far: refused or timed out.
We assess human presence through random spoken words issued live, prescribed tracing paths, timing and synthetic-voice checks.
We're researching correlations between voice and movement, and refining the checks through testing and iteration.
AI agents clear CAPTCHAs now. The web needs a successor that looks for live human signals without sacrificing privacy. That's what we're building Entros to become.
Deepfakes can clone a face or a voice from seconds of source. Proving a live human is present only gets harder from here.
Our answer: never trust one signal. Entros challenges voice, touch and motion together in server-gated rounds, validated server-side, with raw motion and touch kept on the device.
This week our red team ran its full adversarial tiers against the live paired-round flow, in an isolated copy of the production configuration. Tampered commitments: refused. Replayed sessions: refused, 0 of 39. Non-conformant payloads: stopped at the protocol boundary, 0 of 48. Procedural bot audio: rejected at detection, 0 of 24 this run, extending 0% across 16,000+ logged attempts. 95% confidence upper bounds on the pass rate: about 12% for audio, 6% for payloads, 8% for replays.
We are still early, and the goal is unchanged: an internet where being human is provable in seconds and costs you nothing about who you are.
Three short rounds. Each shows one word and a short path to trace. The app hash-commits that round's evidence; the server reveals the next round only after the commit lands. Each commitment chains to the previous one, sealing the session as an ordered sequence.
Raw motion and the full-resolution touch stream stay on the device. Phrase audio leaves transiently for transcription, never stored. Validation receives derived features, coarse trace outlines, wallet and commitment fields.
Personhood without surrendering a biometric.
A single-block challenge has a structural weakness: the client sees all of it before capture begins. The response can be composed as one artifact, and every check runs after the fact.
Proof of humanity cannot be a static object. It has to be a live exchange.
This week we prototyped an upgrade to the verification challenge.
Live today, the whole challenge arrives at once: a twelve-second phrase and a shape to trace. You hold it all in your head, and a client that receives it can build its answer afterwards.
In the prototype it arrives one step at a time: one word, one shape, then the next. The server withholds each step until it has an unchangeable copy of your answer to the one before. Better UX, because each step asks less of you. Better security, because no client can build its answer in advance.
Proof of personhood asks whether a person is there. It does not have to ask who. Entros keeps your motion and touch on the device, and puts nothing on-chain that names you.
We released a new piece of Entros this week: the Agent Operator Permit.
An AI agent can hold a wallet, sign transactions and work all day without anyone watching. An application that wants to allow that has two options today. Block automation and lose the agents people actually want to use, or allow everything and keep no record that anyone approved it.
A permit is the third option. It authorizes one action, by one registered agent, at one application. The wallet that owns the agent signs it, and it expires in minutes.
We ran it on devnet with a real agent and a real wallet.
>One owner-signed permit: the action ran once.
>The same permit again: refused.
>A copy presented by another key: refused.
>A permit held past its expiry: refused, without a single chain read.
>A permit held while the agent changed hands: refused, with the new owner named in the log.
What this opens up:
Any action where an application wants a verified owner behind the agent: a vote, a claim, a limited mint, a listing, a call that costs somebody money. The application keeps its own policy, including the Trust Score it requires and how recent the verification must be.
It is live on devnet now. Two packages carry it, pulse-sdk and verify, with a runnable reference consumer covering issuance, both signatures, expiry, replay and transfer.
A permit states current wallet control of a registered agent, plus the owner's Entros verification result. It does not claim legal ownership or population uniqueness. Next up: the first outside application settling a real action through one.
Learn more:
entros.io/agents
