Alias Archive

Abstract

A host-agnostic architecture for a desktop AI companion. State and logic live in a local server while the web page only renders, allowing a browser, a borderless pywebview/WebView2 window, and Lively wallpaper to act as interchangeable skins. Three infrastructure battles follow: speech recognition moved server-side because a wallpaper host could not interact with native permission UI, bringing local whisper transcription and deadline-based echo suppression; audio playback moved to the server because the wallpaper process was frozen by its host; and global shortcuts moved from conflicting system combinations to a macro pad emitting F13–F19, followed by three separate debounce repairs. Two environment traps are also recorded: logging under a Japanese locale crashed a feature path, and missing CUDA libraries appeared only at first inference.

Abstract

A record of building an independent dialogue engine for an AI companion. The central decision was that a relationship should not use a multi-session model: one continuous timeline feeds four memory layers—recent verbatim context, context-separated rolling summaries, editable long-term notes, and a full archive with per-turn retrieval. The main failure was a style infection in which one phrase appeared eleven times in twenty replies and was then preserved by the summariser as long-term personality. The article traces three layers of reinforcement—examples in the window, self-imitation, and summary consolidation—explains why instructions lose against examples already in context, and gives a mechanical repair: summary surgery, summariser constraints, generation guards, and an on-demand context cut. A fast model handles conversation while a stronger reasoning model performs latency-insensitive memory compression in the background.

Abstract

A first-person record of a digital-companion system from first impulse to working whole: why it began, where it met hard limits, how the direction changed, and when the project left its upstream foundation to become independent. Five linked technical records cover lip-sync tests, content isolation, host architecture, memory, and a controlled selfie system. The account follows contemporaneous work notes rather than polishing the process in hindsight.

Abstract

A design record for an AI companion's selfie behavior. The central idea is layered authority: the model reads the conversational atmosphere—at home, late at night, outdoors, and so on—while outfits are sampled from a manually maintained pool filtered by that atmosphere. She reads the room; you stock the shelf. Expressive freedom and the upper content boundary belong to different actors. The outfit engine combines shot types, ten slots, exclusion groups, and whole-outfit replacement rules, then validates them through hundreds of randomized trials. The post examines three failures: an image endpoint silently ignored a reference-image parameter despite returning 200 twice; a caption claimed a composition that contradicted the sampled shot; and praise triggered an endless stream of follow-up photos. The fixes were the correct endpoint, shared composition and expression parameters, and a narrow rule that blocks praise-triggered repetition without removing proactive sharing.

Abstract

A design and validation record for a two-context isolation mechanism. Normal and private modes each hold separate persona additions, rolling summaries, and long-term memories. When normal mode reads history, every private segment becomes one constant bridge placeholder. The article explains the fail-closed valve—boot and every failure land on the normal side, entry requires verification while exit does not—the information argument for constant placeholder text, and crash recovery for an open segment. It also analyses three failures: a silently blocked cross-origin request meant isolation never activated while mocked tests stayed green; double trimming erased context every turn; and prefix deletion removed legitimate history on the same side. Twenty-two regression checks protect the final semantics.

Abstract

This article records a failed generation experiment and the statistical review that could have anticipated it. To build a dependable action-shot pipeline, the author examined five complete, publicly released video workflows, extracting 231 video nodes and 228 prompts to classify input modes, reference-image counts, and action language. Attack verbs appeared in 18% of prompts and two-subject interaction in 6%, yet prompts describing physical contact between characters appeared zero times: the five apparent contact cases were deliberate non-corporeal penetration or collisions between a character and the environment. Ten generations of a two-character combat shot then addressed seven repeatable failure modes—turn-based staging, style collapse, anatomical drift in non-human subjects, reference-locked distance, broken action causality, idle waiting, and unnecessary endings. Only the causal credibility between attack and evasion never held, and the experiment was stopped. The article also records two methodological conclusions: sampled-frame grids are systematically unreliable for judging generated video, and adding more documentation does not solve the failure to retrieve conclusions already on record. Statistics, observations, and inferences remain distinct throughout; the claims are limited to the tested models and the time of writing.

Abstract

This article documents the construction of a keyframe-to-animation pipeline for a painterly animated pilot. It begins with a reversal in production logic: image models are difficult to control while the script is completely editable, so the storyboard should adapt to strong generated images instead of repeatedly spending generation budget on a fixed composition. It records the failure that forced this change and the revised storyboard structure, which specifies beats and red lines rather than composition. It then identifies three repeatable prompt failures: unspecified body parts are filled in, mentioned objects appear even when negated, and some spatial constraints cannot coexist from any camera angle. Finally, it measures texture retention across two image-to-video paths. Using Laplacian response to separate flat-surface texture from contour sharpness, the two paths retain 76% and 39% of the keyframe's surface texture. A post-production correction matches the metric exactly but fails visually. Observations, measurements, and inferences remain distinct; the pipeline is still unfinished.