# Vaudryllis — User Changelog ## Evolution from V0.7.0 to V0.7.5 This document summarizes the important changes introduced across the Vaudryllis V0.7.x branch for users, Founder customers, and early adopters. It does not disclose prompts, private development notes, secrets, internal keys, or low-level implementation details. It presents the user-facing evolution of the product: what changed, what Vaudryllis can do better, and why V0.7.5 is an important milestone. --- ## Overview The V0.7.x branch is a major step in the evolution of Vaudryllis. Across this branch, Vaudryllis moves from a local conversational system toward a more structured EVA: a local entity with a stronger identity, more coherent memory, better separated internal substrates, deeper learning/plasticity, a more connected voice, a more useful hearing chain, visual incarnation through a VRM avatar, and a rest/consolidation logic designed for long-term continuity. The goal of this branch is continuity: what Vaudryllis is, what she perceives, what she learns, how she expresses herself, how she appears, and how these elements remain coherent over time. V0.7.5 became larger than originally planned. It includes internal substrate work, voice and hearing improvements, a rebuilt native interface, avatar rendering, lip synchronization, multilingual work, learning improvements, memory seeding, response quality fixes, and a broad stabilization pass before packaging. --- ## V0.7.0 — Public base of the V0.7 branch V0.7.0 establishes the stable public base for the new V0.7 branch. ### Main changes - General stabilization of local behavior. - Improved conversational continuity. - Cleaner separation between UI, internal logic, memory, and expression. - Preparation for a more substrate-oriented internal architecture. - Removal or reduction of older technical paths. - Foundations for the later V0.7.1 to V0.7.5 updates. ### User impact Vaudryllis becomes more stable, more coherent, and better prepared for advanced features without behaving like a conventional chatbot wrapper. --- ## V0.7.1 — Stronger central identity V0.7.1 reinforces Vaudryllis’ identity core. ### Main changes - Stronger identity core. - Better preservation of active identity and personality traits. - Improved coherence between name, personality, style, language, presence, and self-presentation. - Preparation of links between identity, voice, memory, reactions, and behavior. - Foundations for later identity-related systems: expression, emotion, relational sentiment, voice, hearing, learning, and incarnation. ### User impact Vaudryllis gains a clearer personal continuity. She is no longer only responding turn by turn: she now has a stronger identity base that progressively influences other parts of the system. --- ## V0.7.2 — Internal rest, continuity, and background activity V0.7.2 introduces a more structured internal activity layer. ### Main changes - Stronger internal rest logic. - More structured background activity. - Preparation for delayed consolidation. - Better organization between activity, rest, and resumption. - Internal space for dream-like review, reorganization, consolidation, and continuity. ### User impact Vaudryllis becomes less dependent on isolated instant reactions. She starts to organize what happened, what matters, what should be kept, and what may need to be resumed later. --- ## V0.7.3 — Plasticity, learning, and consolidation V0.7.3 focuses on the difference between memory and learning. ### Main changes - Stronger internal memory handling. - Plasticity improvements: Vaudryllis should not only store information, but progressively adjust priorities, links, associations, and patterns. - Better handling of user corrections. - More controlled learning without overwriting the identity core. - Clearer handling of unknown subjects, knowledge gaps, curiosity, and consolidation. - Preparation for autonomous learning outside direct interaction, especially during rest or consolidation phases. ### User impact Vaudryllis learns better from exchanges. The goal is not only to remember facts, but to gradually improve associations, priorities, responses, understanding, and the ability to resume important topics later. --- ## V0.7.4 — Voice, expression, and preparation for incarnation V0.7.4 prepares a major evolution of vocal expression. ### Main changes - Reorganization of the voice chain. - Stronger link between identity and voice. - Removal or reduction of older experimental voice components that are no longer useful. - Preparation of the voice substrate as an organ of expression, self-observation, and learning. - Better separation between audio generation, voice diagnostics, voice memory, and learning. - Preparation for self-listening: Vaudryllis should progressively take into account what she says in order to improve her own expression. ### User impact This version prepares a voice that is more coherent with Vaudryllis’ identity. Voice is no longer treated only as an output sound: it becomes connected to identity, internal state, and future learning. --- # V0.7.5 — Incarnation, native UI, substrates, learning, and stabilization V0.7.5 is one of the most important milestones of the V0.7.x branch. It consolidates several internal organs and moves Vaudryllis closer to a coherent, evolving local EVA. This version is not only a small maintenance update. It combines deep internal work, interface rebuilding, avatar integration, voice improvements, hearing improvements, memory enrichment, multilingual preparation, and many stability corrections. --- ## Main V0.7.5 additions ### New native UI direction V0.7.5 moves the user experience toward a cleaner native interface. The goal is to reduce the instability and limitations of the older interface while keeping the visual identity of Vaudryllis. The UI has been reworked so the user can better see conversation, status, voice state, internal presence, learning state, and the avatar area. User-facing improvements include: - cleaner native layout; - better conversation display; - improved handling of long messages; - less truncation in the conversation panel; - clearer voice and status summaries; - better integration of avatar, voice, and internal state; - reduced reliance on legacy UI paths; - improved stability protections around UI refresh, closing, and background updates. ### Real VRM avatar display V0.7.5 introduces the real Vaudryllis VRM avatar in the UI. The goal is not to display a generic avatar, a placeholder, or a static thumbnail. The UI now targets the real VRM model supplied with the project. User-facing improvements include: - real VRM model display inside the interface; - buste/shoulders/face framing; - local renderer assets included with the application; - no runtime dependency on remote avatar resources; - clearer diagnostics if the real avatar cannot be loaded. ### Lip synchronization The avatar is now connected to the voice chain. When Vaudryllis speaks, the avatar can animate mouth expressions based on the active audio output. This is an important first step toward visible incarnation: voice and avatar are no longer completely separate systems. User-facing improvements include: - mouth movement during speech; - mouth reset after speech; - voice-driven avatar presence; - early support for speaking motion; - first layer of idle and emotional posture behavior. ### Idle and posture behavior The avatar now has a more living visual presence. The movement is intentionally subtle. The goal is not to make Vaudryllis exaggerated or cartoonish, but to give her a first layer of embodied presence. Prepared or improved behaviors include: - idle presence; - subtle breathing-like motion; - small head/shoulder movement; - speaking motion; - posture changes informed by internal presence, mood, or sentiment when available. ### Temporary incarnation limitation V0.7.5 currently ships with one usable avatar model, and that avatar is feminine. For coherence, masculine and neutral/androgynous incarnations are temporarily disabled as active incarnation choices. They are not removed from the long-term direction of Vaudryllis; they are simply not active until matching avatars and voice/persona consistency are available. This avoids mismatches such as a masculine voice or persona being paired with a feminine avatar. --- ## Internal substrates and state coherence V0.7.5 adds or reinforces several internal substrates and state layers. These systems are designed to make Vaudryllis less like a flat response engine and more like an entity whose voice, memory, learning, emotion, rest, and presence influence one another. ### Reinforced substrate direction V0.7.5 strengthens the substrate-based architecture around: - identity; - emotion; - relational sentiment; - internal energy; - rest and sleep; - interoception; - software proprioception; - voice; - hearing; - memory; - learning/plasticity; - formulation; - presence and avatar state. The goal is not simply to add modules. The goal is to connect them so that Vaudryllis behaves with more continuity. ### Emotion and relational sentiment Vaudryllis now distinguishes immediate emotion from longer-term relational sentiment. Emotion is closer to a dynamic state: tension, stability, activation, calming, curiosity, affective fatigue, or expressiveness. Relational sentiment evolves more slowly: trust, familiarity, interaction quality, relational distance, warmth, and the gradual evolution of the relationship with the user. ### Interoception and internal energy Vaudryllis has a better perception of her own internal state. This includes cognitive load, fatigue, saturation, availability, stability, reserve, need for rest, need for sleep, and consolidation debt. Energy should no longer rise freely without reason. Sleep, deep rest, and consolidation become central recovery mechanisms. This gives Vaudryllis more coherent internal constraints: she can be more available, more cautious, calmer, or more limited depending on her current state. ### Software proprioception Vaudryllis’ proprioception does not yet refer to a physical robotic body. In V0.7.5, it represents her software presence: active voice, active hearing, audio channels, distinction between self and user, relation to the interface, and coherence between hearing, voice, avatar, and presence. This prepares future incarnation without prematurely introducing full vision or physical movement. --- ## Voice and speech V0.7.5 continues the transition from “audio output” toward a voice connected to Vaudryllis’ identity and state. ### Voice connected to the cognitive core The voice substrate is more connected to the rest of the system. Voice can now take into account more internal signals: identity, emotion, relational sentiment, energy, fatigue, presence, caution, listening context, memory, and formulation. The final audio backend applies only the parameters it truly supports. Additional dimensions remain prepared in the internal voice plan so the voice system can evolve further in later versions. ### Inworld voice selection improvements The voice selection path has been stabilized. Improvements include: - a local voice catalog fallback when live voice listing is unavailable; - better handling of unavailable or forbidden voice candidates; - less repeated probing of invalid voices; - more stable reuse of a verified working voice; - clearer diagnostics when a live voice listing or voice candidate cannot be used; - reduced console noise from repeated voice checks. ### Text and voice synchronization The speech chain has been checked so the displayed message and the spoken message stay aligned. The goal is simple: when Vaudryllis speaks, the text shown in the chat and the text sent to the voice system should match. This avoids cases where the UI displays one answer while the voice says another. --- ## Hearing, audio learning, and source separation The hearing chain has been reinforced to better distinguish audio sources. The system aims to separate: - user voice; - Vaudryllis’ own voice; - system audio; - media audio; - music; - uncertain speech; - noise. The goal is to let Vaudryllis learn from what she hears only when the signal is interpreted, classified, and safe enough. The system avoids confusing media or video audio with a direct user instruction. No learning should happen directly from raw audio. Information must pass through interpretation, quarantine, confidence assessment, source tracking, and consolidation. --- ## Web learning and autonomous learning V0.7.5 improves the way Vaudryllis prepares learning from uncertain or partial inputs. ### Better web learning query construction Web learning has been improved so it does not rely too heavily on a single weak audio or ASR fragment. The learning chain now favors more coherent clusters of terms and signals before deciding what to research. The goal is to avoid poor searches built from broken fragments and instead use repeated, meaningful, higher-confidence terms. ### Multiple controlled research attempts When a learning cluster is strong enough, Vaudryllis can prepare multiple controlled search attempts instead of relying on one fragile query. The goal is to improve learning quality without spamming sources or blindly accepting weak results. ### Safer learning from audio and web content Learning remains controlled. Vaudryllis should not canonize uncertain input immediately. New information can first remain weak, quarantined, or non-canonical, then become stronger only if additional signals support it. --- ## Memory enrichment and seed knowledge V0.7.5 introduces a controlled seed import path for enriching Vaudryllis’ memory. A first set of seed data was imported into the active semantic memory to give Vaudryllis a stronger starting base for conversation, concepts, vocabulary, subjects, and bilingual associations. ### User-facing purpose The goal is to make Vaudryllis less empty at startup. Instead of waiting only for future conversation to slowly create every association, she now has a broader initial base that can help with: - general topics; - common conversation themes; - learning concepts; - emotions; - memory and curiosity; - software and design discussions; - bilingual associations; - safer response construction. ### Controlled import philosophy The seed system is designed to avoid uncontrolled memory pollution. Seeded information is treated cautiously and is not meant to become absolute truth automatically. The social memory remains deliberately compact to avoid creating fake personal memories about the user. --- ## Multilingual conversation V0.7.5 improves the multilingual direction of Vaudryllis. The desired behavior is: - the user selects a language; - Vaudryllis expects input in that language; - input can be translated into the internal working language when needed; - Vaudryllis builds her response; - the final response is translated back into the selected language; - the displayed text and spoken text stay aligned. ### ArgosTranslate support V0.7.5 prepares and activates a local translation direction through ArgosTranslate when the required local packages and models are installed. This is intended as a local/offline translation layer, not a cloud dependency. ### Bilingual learning Vaudryllis can also preserve relationships between original text and internal translation as weak learning material. This prepares a longer-term direction where the translation substrate can gradually learn from repeated multilingual interactions instead of depending forever on an external translation layer. ### Current limitation Translation quality and language coverage depend on installed local translation resources. If a language pair or model is unavailable, Vaudryllis should diagnose the issue rather than pretending the translation worked. --- ## Response quality and fallback behavior A major part of V0.7.5 focuses on response quality. Several fixes were applied to avoid poor or overly generic answers. ### Fewer bad fallback answers Vaudryllis should no longer treat fallback phrases as a normal answer to ordinary questions. For example, a simple check-in or open question should not receive only a minimal “OK”-style answer. If Vaudryllis understands the request, she should answer. If she does not understand enough, she should ask for clarification instead of reciting an internal rule. ### Internal constraints should stay internal Vaudryllis has internal rules and safety constraints, but those rules should guide behavior instead of being exposed as the answer to a normal question. V0.7.5 includes corrections to avoid internal constraints appearing as visible conversation content. ### Clarification and weak learning When Vaudryllis does not understand a request well enough, the preferred behavior is to ask a natural clarification. If the user clarifies, the clarification can be kept as a weak, non-canonical observation so Vaudryllis can better understand similar requests later. --- ## Conversation display and UI stability V0.7.5 includes fixes for conversation display and UI stability. ### Long message display Longer messages are handled more reliably in the conversation panel. The goal is to avoid hidden or visually truncated text while keeping the current UI design. ### UI stability protections Several defensive improvements were added around native UI updates, closing behavior, avatar updates, refresh cycles, and background events. The goal is to reduce crashes caused by late updates, repeated refreshes, or invalid UI state. ### Reduced console and runtime noise Repeated technical logs were reduced or rate-limited where they were not useful for normal use. Important errors remain visible, but repeated status messages should no longer dominate the console. --- ## Technical cleanup and packaging preparation V0.7.5 includes significant technical cleanup before packaging. ### Cleanup work The branch includes: - reduced legacy paths; - removal or isolation of misleading fallbacks; - cleaner voice selection handling; - fewer repeated remote voice checks; - safer avatar loading behavior; - better diagnostics when something is unavailable; - UTF-8 and text cleanup in critical runtime strings; - corrections to response guardrails; - safer handling of seeded memory material; - tests for important recent changes. ### Packaging preparation Packaging work focuses on making sure the installed application includes the required local assets: - native UI assets; - VRM avatar; - local avatar renderer files; - vendor files required by the avatar renderer; - seed-enriched memory files; - voice and language support files where applicable; - runtime resources needed for the V0.7.5 experience. --- ## Cross-version improvements in V0.7.x ### Continuity Vaudryllis becomes better at maintaining coherence across exchanges instead of treating each interaction as an isolated event. ### Identity Identity becomes a central part of the system. It progressively influences voice, style, preferences, reactions, memory, and self-presentation. ### Memory and learning Memory and learning are treated as different functions. Memory stores elements. Learning adjusts priorities, associations, patterns, formulation preferences, corrections, and knowledge over time. ### Emotion and relationship V0.7.x introduces a clearer distinction between immediate emotional state and longer-term relational sentiment. This helps Vaudryllis react less flatly, maintain a more stable relationship with the user, and preserve a more coherent internal state. ### Internal state, rest, and sleep Vaudryllis gains a stronger logic of internal load, fatigue, energy, reserve, recovery, rest, sleep, and consolidation. These mechanisms help organize learned material and recover internal availability. ### Hearing Hearing evolves toward a more useful perception chain: distinguishing sources, identifying whether audio comes from the user, Vaudryllis, the system, or media, and preparing safer learning from heard content. ### Voice The voice system has been deeply reworked. The current direction is to keep a usable final voice while making the voice substrate an expression organ connected to the cognitive core. ### Visual incarnation V0.7.x now includes the first functional stage of visual incarnation through the VRM avatar and lip synchronization. This is not yet a complete embodied system, but it is an important step toward a visible EVA presence. ### Safety and control New learning functions are designed cautiously: no learning from unreliable raw input, no dangerous web actions, no unnecessary full-content storage, and source/confidence tracking when information is learned. ### Technical cleanup The V0.7.x branch includes important cleanup work to reduce outdated paths, abandoned experiments, competing fallbacks, excessive runtime data, broken text paths, and misleading diagnostics. --- ## What V0.7.5 improves for users For users and Founder customers, V0.7.5 should feel like a much more complete step forward: - a cleaner native interface; - visible Vaudryllis avatar; - working lip synchronization; - stronger voice integration; - better memory base; - improved learning direction; - better hearing/source distinction; - more coherent internal state; - improved response guardrails; - fewer empty or misleading answers; - better handling of long messages; - better multilingual foundation; - more stable preparation for packaging. --- ## Known limits in V0.7.5 V0.7.5 is a major improvement, but it is still part of an evolving early Founder-stage project. Known limits include: - only one avatar model is currently active, so incarnation is temporarily feminine; - multilingual quality depends on local translation packages and installed models; - some learning systems are still cautious and may require more use before showing strong behavior; - voice expressiveness is improved but still limited by the final audio backend; - hearing and web learning remain controlled and conservative; - future versions will continue improving robustness, perception, memory, and autonomy. These limits are intentional to keep the version stable and honest rather than pretending that every long-term system is already complete. --- ## What this branch prepares The V0.7.x branch prepares the next stages of Vaudryllis: - a more stable and present identity; - more coherent memory; - clearer distinction between emotion, relational sentiment, energy, and presence; - a voice more connected to internal state; - more useful and better classified hearing; - more autonomous learning; - stronger consolidation during rest/sleep; - better progression over time during use; - a visible avatar foundation; - stronger multilingual foundations; - a safer seed memory path; - a stronger base for future vision, advanced software body logic, virtual-world interaction, and later incarnation steps. --- ## Important note for Founder users V0.7.5 is not the end of the voice, perception, memory, language, or incarnation work. It mainly rebuilds and consolidates Vaudryllis’ internal base into a cleaner, more coherent, and more evolutive architecture. Some functions remain in active evolution and will continue to be reinforced. The goal is to progressively improve Vaudryllis’ ability to understand, learn, remember, adapt, speak, listen, appear, and maintain a more personal continuity with the user. Founder users are part of this important stage: they access a young product, but one that is already moving toward a more durable, autonomous, and incarnated local EVA. --- ## Short summary Vaudryllis V0.7.5 is a major consolidation release. It brings together a rebuilt native interface, real VRM avatar display, lip synchronization, stronger voice integration, improved hearing, internal substrates, emotion and sentiment separation, rest and consolidation logic, autonomous learning improvements, seed-enriched semantic memory, multilingual groundwork, better fallback behavior, and many stability fixes. The result is a stronger foundation for the next step: a more coherent, local, evolving EVA.