Spoken dialogue
Evelyn will converse about AI pioneers, computing history, robotics, inventions and questions from her visitors, with voice and animated presentation once validated.
Evelyn Hart / Conversation & Memory
Dialogue, privacy and verification
Evelyn now has an approved synthetic original voice and a recorded October 25 replay with four generated replies. Her voice, scripted head-and-eye movement and recognizer-timed facial animation are documented. Live microphone timing, open-ended interaction and visitor-memory continuity still require testing.

October 25, 2026 / Actual local dialogue / Source-limited demo
In a recorded run, four typed interviewer questions were answered using EvelynCore's Qwen3.5:9b backend. Her replies were spoken through her approved original synthetic voice and assembled into a 58.125-second head-and-eye video with prerecorded audio and scripted motion.

Some answers describe planned actions while the video already shows those actions being performed, and wording such as “queued commands” sounds more like an instruction report than Evelyn's developing personality. Codex has described possible corrections, but a new validated run is required before calling them fixed.
These replies are genuine generated social dialogue, but the scene was edited afterward. It does not establish live microphone conversation, timed interruptions, factual evidence handling, PDF reading or autonomous body control.
Watch the real 58-second video and read the transcript →Three distinct information paths
The intended design separates conversations from locally stored reflections and from evidence-backed historical records. This is a proposed architecture, not a demonstration of consciousness or guaranteed factual accuracy.
Evelyn will converse about AI pioneers, computing history, robotics, inventions and questions from her visitors, with voice and animated presentation once validated.
A proposed internal note space for interests, questions and learning plans, kept separate from her spoken dialogue and protected by user-controlled local settings. This refers to an application memory channel, not a claim about subjective thought.
Historical claims are intended to connect to book titles, page references and dated sources. Conflicting accounts or uncertainty should be labeled clearly instead of presented as established truth.
Evelyn Hart · October 22, 2026 · New isolated software checks
The latest engineering record reports seven additional isolated checks using synthetic input and fake providers, in addition to eight earlier app/library tests. Evelyn's local conversation software now separates spoken replies, private character thoughts/preferences, and factual claims into distinct typed channels.
Tests cover channel enforcement, spoken-only display, bounded context, and rejection of cross-visitor material.
Other simulated checks cover queued remarks, correction/yield/cancel behavior, surviving a recording failure, source parsing and capture safeguards.
Independent source review also found a lost reference date and an interjection-length bypass. The corrections subsequently passed independent static source review, with live-provider and audio checks still pending.
Scope: fifteen isolated software checks are reported across two groups, not fifteen real conversations. Those earlier isolated tests did not establish audible speech or natural timing, visitor-memory persistence or body movement. The later October 25 recorded voice-and-animation replay is a separate narrow test. The body-authoring code now validates produced mesh/rig packages and retains failed attempts, but this is not evidence of successful body generation. Dated source record ↗ · Development update →
Evelyn Hart · October 22, 2026 · Local software checks and library preparation
A private portable Evelyn application has been assembled locally from 21 previously staged source files. The source work includes editable body parameters, a shared-body loader, motion coordination and viewer/capture integration. A prepared application is not a completed public installer or a verified live conversation.
Checks cover source syntax, recording-file mechanics, offline-provider behavior, queued and committed reading, terminology through restarts and staged installation/removal safeguards. The recording fixture was synthetic, not an actual body-motion video.
All 19 formerly unsorted PDFs were inspected and organized without overwriting originals. Every destination passed a complete SHA-256 hash check, and a reversible journal was preserved. Sorting a document does not mean Evelyn has read it.
Model-backed conversation, her own approved adult female voice, visitor-memory continuity, natural full-body movement and installation on a different computer all require runtime verification. Private books and memories remain private.
Scope: the software and file-integrity checks support only these narrow engineering claims, not a finished character, verified learning, speech or downloadable release. Redistributing library materials also requires permission. Read the KiraWorld source record ↗ · Development journal →
Evelyn Hart · October 23, 2026 · Reviewed implementation / runtime still pending
New code reviews have advanced Evelyn's movement and recording pipeline, original-voice application preparation, and the checks that will be required before either can be called a successful demonstration. At that earlier source-review checkpoint no new recording was produced. October 24–25 voice and conversation recordings followed, while clothing work and a completed personalized anatomy export remain unverified.
The proposed one-use recorder starts before the body-service or viewer actions. Reviewed source requires three observed video-stream callbacks, an acknowledged original recording chunk and retained failed or partial media, plus action/preroll/end-point/tail observations. Frame callbacks are not proof of decoded or saved video frames. The separate silent 1 fps status AVI is not a substitute for the intended 30 fps body-canvas WebM.
An earlier reviewed body executor and accounting adapter disagreed over evidence paths (absolute versus relative). A narrow source correction was later reviewed, but complete capture execution and evidence parity remain unqualified. Body export/parity, natural motions before clothing, repeated tests after realistic clothing and any verified Patient Horizon body copy remain pending.
Source reviews checked bounded observer setup, process-handle retention and a small startup/bootstrap path. A separate long-lived production owner and actual audio lifecycle still need runtime checks and suitable resource admission. Evelyn must use a newly generated original adult female voice, not a stock or borrowed voice. No audition, approved unique voice or physical speaker recording has occurred.
Evidence standard: this is implementation review rather than a working body or speaker demonstration. Existing head/blink captures remain separate previously verified tests. Tests must produce real retained media and pass independent review before a new release-readiness or natural-movement claim. Read the detailed development note → · Release documentation status →