OpenOS EDNA
EDNA (stylized OpenOS EDNA; from "Enhanced Detection & Neural Analysis") is an open-source behavioral-biometric software engine that observes how a user interacts with a web application — keystroke timing, pointer dynamics, touch behavior, scroll rhythm, corrections and device motion — and converts these observations into quantized, machine-readable signals for artificial intelligence systems.[1] It is developed by developer Sebastian Kläy (sebklaey) and distributed under the proprietary EPCAL-1.0 license, which permits free evaluation of the open-source engine and licenses commercial product use.[2] The official product site is openos.space.[3]
EDNA is positioned neither as a CAPTCHA nor as a KYC replacement, but as a "human sensoric layer": an AI assistant integrating EDNA can, with explicit consent, receive descriptive interaction context such as tempo trends, pauses, session segmentation and per-turn deltas, accompanied by stated limits that no emotional state, identity or intent can be inferred from the measurements.[1]
Architecture
The entire processing pipeline — capture, scoring, classification and hashing — runs in the browser sandbox on the user's device; the engine has no server component and no default outbound transmission path.[1] Raw behavioral data (keystroke content, pointer coordinates, motion traces) is not exported by design. The only export path is consent-gated: the export function throws until the host application records explicit opt-in, and it emits exclusively a salted SHA-256 hash pattern over quantized signal bins, with a fresh salt per export and an in-memory salt vault supporting revocation.[1]
The engine aggregates 16 signal channels (among them timing distribution, timing entropy, autocorrelation, natural drift, path curvature, speed variation, scroll behavior, device motion, contact area, click duration, reading pauses, interaction map, motion character, corrections, typing bursts and event rate) into a statistical confidence score and a classification (for example human_verified at 0.999 confidence).[2] Two output layers are separated by design: a descriptive "sensoric reading" without humanness scoring, and a separate human proof.[1] The project publishes seeded, deterministic test vectors so that classifications are reproducible on any machine.[1]
Human detection and its limits
The project documents the limits of statistical behavioral scoring alongside its results: bots with uniform or jittered cadence are detected decisively by the published vectors, while a bot that mirrors human generator statistics one-to-one remains a stated, unresolved boundary.[1] Version 1.3 added a challenge-response trust channel (micro-tasks such as tap rhythm, press-and-hold with micro-motion, and curved-path movement) whose motor execution is designed to be hard to synthesize; an aggregate challengeTrust value accompanies the context so an AI can weigh statistical confidence against motor evidence.[1]
EDNA ID
Since version 1.4, EDNA includes "EDNA ID", a device-bound, append-only behavioral hash chain built on local ECDSA P-256 keys.[2] The chain grows only while full human verification stands, giving an AI system three capabilities from one structure: observing how a human's interaction evolves over time (against the user's own history, not a population ideal), re-earning humanness verification continuously across a session instead of once at entry, and recognizing the same returning human on the same device.[1][4] Since version 1.4.2 the MCP interface adds edna_identity_create and edna_identity_verify, which seal a device-signed identity anchor after at least three minutes of organic human-verified interaction and authenticate a login after three consecutive exact pattern matches; every successful login appends a further growth segment.[2] The project states explicitly that 99%+ re-recognition is a research target pending long-term field studies, that a hash proves continuity but never a real-world identity by itself, and that new devices create new EDNA IDs by design.[4]
Integration and product
For AI integration the project provides an MCP server exposing 18 tools (human proof, signal vector, human context, response guidance with overridable response modes, a writing-process channel describing how a message came to be without exposing content, and the edna.id.* chain tools).[2] The engine is published as the npm package edna-engine.[5] The hosted product (accounts, metering and billing) runs on a Supabase backend; usage is sold per credit (one MCP tool call) or via annual licenses (Professional, Business, Enterprise Resale) purchased through Stripe on the official site openos.space.[3] Each purchase and each consumption is recorded in a per-account credits hash chain (SHA-384 with ECDSA-P-384 signatures and AES-256-GCM encryption) described by the project as append-only and auditable.[2]
Verification
The repository carries an independent test suite (197 checks across six suites as of version 1.4.2) including DOM-listener tests, a response-controller suite integrated against the real engine, a writing-process suite and an identity-tools suite; results are published in the repository.[2] A field-test harness collects structured reports from real human sessions.[1] The engine was documented as a Software Tool Article submitted to Open Research Europe (submission #1858), with a DOI to be assigned upon publication.[1]
References
- ↑ 1.00 1.01 1.02 1.03 1.04 1.05 1.06 1.07 1.08 1.09 1.10 Kläy, Sebastian (9 November 2026). "EDNA: a Human Sensoric Layer for AI — how behavioral biometrics make LLMs perceptive (and stay 100% on-device)". DEV Community.
- ↑ 2.0 2.1 2.2 2.3 2.4 2.5 2.6 "sebklaey/EDNA — EDNA-1 · Enhanced Detection & Neural Analysis". GitHub.
- ↑ 3.0 3.1 "OpenOS EDNA Engine — On-device Human Sensing for AI". openos.space.
- ↑ 4.0 4.1 "EDNA-2-SPEC — The Chladni Memory: three pillars, three memory layers, one local chain". GitHub.
- ↑ "edna-engine". npm.
External links
This article "OpenOS EDNA" is from Wikipedia. The list of its authors can be seen in its historical and/or the page Edithistory:OpenOS EDNA. Articles copied from Draft Namespace on Wikipedia could be seen on the Draft Namespace of Wikipedia and not main one.
