Single-signal detectors age fast
SONAR Benchmark, ACM 2024-2025
The best classifier caught OpenAI's TTS 78% of the time. Detectors trained on older data lost up to 43% against newer TTS.
Behavioral verification for the agent era.
Your device measures voice, touch and motion, then builds a zero-knowledge proof that Solana verifies.
Three rounds. The rest is automatic.
Each round asks for one spoken word and one traced path.
01Prompt
01Prompt
A fresh word and path every round.
The server draws each word from a curated vocabulary and each path from random parameters. Your browser locks in a round before the next one appears.
02Capture
02Capture
Voice, touch and motion, recorded together.
The SDK reads the microphone, touch or pointer, and motion sensors where present. Raw motion and touch stay in memory and are discarded after measurement. Phrase audio leaves for transcription and is never stored.
03Measure
03Measure
308 features across three signals.
Voice gives 170 (pitch, formants, voice quality), motion 81 (jerk, tremor, band energy) and touch 57 (velocity, pressure, curvature).
04Hash
04Hash
A 256-bit fingerprint.
SimHash maps similar feature sets to nearby fingerprints. Current research measures how well one person's captures stay close while other people and synthetic inputs land apart.
05Commit
05Commit
Only a salted hash goes on chain.
Poseidon hashes the fingerprint with a large random salt. Both stay on your device, with an encrypted copy on chain when your wallet can sign messages.
06Prove
06Prove
A zero-knowledge proof of distance.
A Groth16 proof shows both commitments open to fingerprints whose Hamming distance sits at or above the replay floor and below the maximum. The fingerprints stay private.
07Verify
07Verify
Checked by the service, settled on Solana.
The validation service checks the spoken phrase, signs of synthetic audio, the capture and fingerprints on other wallets. Solana verifies the proof, updates your Anchor and recalculates your Trust Score.
Five layers, one proof.
02
ZK Circuit
Hamming distance in zero knowledge
A Groth16 circuit proves two Poseidon commitments open to fingerprints within the configured distance range. It does not prove where a capture came from.
- Groth16 over BN254
- Poseidon commitments
- Verified on chain in under 200,000 compute units
03
On-Chain Programs
Three Solana programs
entros-verifier checks proofs. entros-anchor holds the non-transferable Anchor and Trust Score. entros-registry holds protocol configuration and treasury state.
- Non-transferable Token-2022 Anchor
- Trust Score from active weekly bins and account age
- Accounts derived from your wallet
04
Executor Node
Challenges and relay
A Rust gateway that issues signed challenges, applies integrator quotas, forwards evidence to the validator and relays protocol writes.
- Signed challenges for every session
- Per-integrator API keys and rate limits
- Best-effort SAS attestation
05
Validation Service
Private defense layer
A private Rust service screens each capture for synthetic audio, unusual signal statistics and fingerprints seen on other wallets. The protocol stays open. Detection thresholds stay private because they would guide an attacker.
- Synthetic audio detection
- Cross-wallet fingerprint registry
- Uniform public failure reasons
Privacy by architecture.
Done on your device
Capture, measurement and the proof run in your browser. Raw motion and touch never leave it.
Words never stored
Your spoken words leave for transcription, stay in memory and are never written to a log or store.
Only what the check needs
The service receives signal summaries, your words and a rough outline of each path.
No identity
No name, email, document, face scan or social account.
A one-way record
On chain, a salted hash stands in for your fingerprint. The salt stays with you or is encrypted under your wallet's key.
Apps read a result
Apps read your Entros Anchor and Trust Score, never the data behind them.
How Entros checks a capture.
A private service checks each capture before it settles. Published results name the attack class tested and the sample size.
01
Fresh challenges
The server issues a new word and path for every round, then checks the transcript against the issued words before settlement.
02
Feature validation
The private service checks the 308 features for acoustic artifacts, unusual signal statistics and duplicate fingerprints. Its thresholds stay private.
03
Cross-wallet registry
The service compares each new fingerprint with recent fingerprints from other wallets. Population-scale performance is still under evaluation.
04
Distance bounds
Each re-verification proves the new fingerprint sits within the circuit's minimum and maximum distance from the last one. An identical fingerprint fails.
05
Several signals
Voice, touch or pointer, and motion where available. Entros is measuring which combinations separate people, devices and synthesis methods.
06
Trust over time
Trust Score counts active weekly verification bins and account age. A second verification in the same week adds no bin.
07
Fees and rate limits
Fees and rate limits bound how many attempts reach the service. The validation policy decides whether a capture passes.
Verification state on Solana.
A captcha checks one session. Entros is designed to show that a returning person keeps operating the same wallet. Apps read the state and set their own policy.
Wallet-connected verification
Your wallet holds a non-transferable Entros Anchor, a Token-2022 account. Its commitment, Trust Score and history stay readable on chain. The wallet pays the fee and the Anchor's rent.
Evidence behind the research program.
Outside studies shape this research. Their numbers describe their own tests and do not describe Entros on consumer devices.
Touch and motion fusion
BioMoTouch, arXiv 2025
Fusion reached 99.71% accuracy and a 0.27% equal error rate for authentication, a different task from liveness.
Arm movement perturbs the voice
Pouw et al., Royal Society Proceedings B, 2025
Upper-limb movement changed a sustained vowel's amplitude in the lab. Whether phones can use it for liveness is open.
Physics-informed voice features
VoiceRadar, NDSS 2025
A micro-frequency feature reached a 0.45% equal error rate in the authors' tests.
TTS artifacts run both ways
Warren et al., "Pitch Imperfect", 2025
Neural vocoders can push acoustic perturbation above human levels as well as below, so a detector has to check both directions.
Touch shows up in motion
Device physics, BioMoTouch 2025
A physical touch can leave a time-aligned trace in motion sensors. Whether consumer devices measure it well enough is open.