Published notes

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.

Abstract

After the previous AI action film was archived as a failure, this article records a Claude Fable 5-driven greenfield restart on the same platform. No prior assets were carried over; the task was reduced to a three-shot atmospheric film within the models' capability boundary, with no physical contact or combat causality. It tests a new division of labor: the creator produces the keyframes in Midjourney, Flova and Seedance handle orchestration and animation, and Claude Fable 5 drives the official CLI and extracts frames for review. The article follows three lines of failure and correction: a helmet hallucinated because the first frame lacked structural information, a text-only silhouette entrance rebuilt as an impact shot, and thick-painted imagery becoming synthetic in motion before MJ Animate offered a better route. It also records the final decision to delete a shot because of character drift. The finished film runs 14.2 seconds, consumed 838 credits, and produced seven reusable rules.

Abstract

This article records the complete production of a 15-second dark-fantasy film driven by Codex 5.6 Sol, from greenfield planning to final export. Codex 5.6 Sol supplied Flova with Niji concept art for the knight and Sylas; the platform planned a five-shot storyboard, generated character and environment Key Elements, and completed two rounds of shot generation, while external frame review, sound design, and timeline assembly closed the process. The workflow respected staged gates and avoided complex combat, but only two restrained single-subject shots earned strong evaluations. Sylas appeared without credible spatial logic, platform-generated weapons and characters looked synthetic, reruns failed to correct travel direction and identity markers, and the final shot had to be replaced by a still. Treating the creator's per-shot scores as the final record, the article explains why a new project, a complete workflow, and a successful export do not make a film successful, and compares the result with a Claude Fable 5 route built around human-authored Midjourney keyframes.

Abstract

This test did not treat a successful export as success. A four-shot action film had to preserve story, space, physical contact, character identity, visual style, and sound continuity at the same time. Flova correctly diagnosed the earlier phantom-distance problem, built a directing plan and timeline, and used Seedance 2.0 to deliver a 38.24-second film with native sound. Gate B, however, repeatedly introduced a knight's helmet on the Boss, premature chest contact, duplicated weapons, and wholesale style drift. In the final sequence, Shot 2 still failed to prove a knee strike, while Vera's identity and lantern hand remained unreliable in Shot 4. More importantly, Flova repeatedly marked these defects as passes. This article records the complete failed workflow from director audit to export, and explains why project-level orchestration still cannot replace independent frame-by-frame review.

Abstract

The goal sounded modest: a 16:9 battle in which a knight evades and counters a stone giant, defeating it before Vera arrives. The visual language stabilized quickly; time and space did not. Early versions dodged too soon, paused before attacking, collapsed floors without a cause, teleported characters, and left the Boss floating. Reducing the task to one five-second action gave Dreamina a clear advantage over Kling, but the continuation introduced a deeper three-dimensional error: the knight was visibly far from the Boss, stood up in place, and somehow struck it. A static contact sheet even made the shot look usable. Rather than force the material into a finished edit, this test preserves the failures, assigns clean single-subject generation to Niji and controlled assembly to Image2, and pauses manual shot-by-shot production while Flova is evaluated as a project-level system for storyboards, shot coordination, and model selection.

Abstract

This was not an attempt to move an entire Godot project into a website. It extracted one minimal, complete, verifiable combat loop from Watchman: a fixed character and fixed values, with movement, jumping, attacks, evasion, poise, stance breaks, and executions. The record covers an explicit asset allowlist, behavioural contracts, collision and hit detection, input buffering, audio and damage numbers, followed by Web packaging, real loading progress, keyboard focus, mobile fullscreen, and contextual-input revisions. Several failed iterations established two rules: combat feel cannot be approximated in a rewrite, and web controls cannot simply copy the main game. Both have to be calibrated against their real operating environment.

Abstract

A company website rebuild is more than a page-design exercise. This article divides the work into seven stages—requirements, stack, content, localization, deployment, domain and filing, and asset handoff—and explains how to choose between static and dynamic architecture, handle forms without a conventional backend, decide where to host, and keep the domain, DNS, hosting account, and source code under company control.

Abstract

This article documents the complete construction of my first Ableton Live cover project. Its central principle is modularity: every missing instrument receives an independent track, and Live supplies only the parts absent from the live lineup rather than playing a complete backing track. Using The Pretty Reckless's Make Me Wanna Die as the working example, the article follows song and key selection, downloading with yt-dlp, stem separation in Moises, import and alignment, drum programming with EZdrummer, and guitar-tone preparation with Ableton's built-in amp simulation. It also explains concepts that commonly obstruct a first project—including Warp, track arming, Session View overrides, and General MIDI mapping—and records the judgment behind decisions such as physically retuning the guitar instead of relying on software transposition.

Abstract

A field note on the unglamorous half of spatial analysis: the data engineering that precedes any coefficient. It walks the generic pipeline from raw address points and administrative boundaries to a single analysis-ready layer — text cleaning and composite keys, reprojection to a metric coordinate system, network-based accessibility instead of straight lines, and a circuity check that flags impossible detours without discarding real geography. It then sketches the modeling ladder above that layer — OLS with a Moran's I diagnostic, a global spatial model to absorb dependence, and geographically weighted regression when the relationship itself varies across the map. Examples are textbook-generic; the method travels, the findings do not.

Abstract

Customer orders rarely arrive in one format: a short message, a spreadsheet, a photograph of a quotation, or a scanned document. This note describes an intake tool that uses a multimodal model to turn those inputs into structured order records in a shared online table. The important part is not the code but the reliability contract: every extracted field is marked certain or uncertain; uncertain values remain empty for a person to resolve; a confidence gate chooses between automatic writing and an explicit review card; and the model handles repetitive interpretation while a person retains final judgment. Technical terms are introduced in plain language for readers new to automation.

Abstract

Alias Archive grew from a thin personal-site skeleton into a trilingual archive spanning research, games, music, and work that resists those categories. This article explains the move from Next and vinext to Astro, the repository contracts used to coordinate multiple agents, the role Impeccable played in refining hierarchy and responsive behaviour, and the way articles, images, audio, video, and a Godot Web demo coexist inside a static site. The durable subject is not a tool list, but a system that remains portable, testable, and controlled by its owner as it grows.

Abstract

This is not a one-click-generation demo. Codex served as the only operational interface and coordinated the script, scene plan, API work, visual package, and review. HeyGen supplied Vera's English voice, lip sync, and base performance; HyperFrames supplied Chinese subtitles, information hierarchy, and the 9:16 master. The process exposed failures in avatar reference selection, vertical typography, face-safe layout, preview reliability, background continuity, and speech cadence. The first sample was accepted with two disclosed imperfections so the next production can correct them at the source.

Abstract

This entry documents a pipeline for turning a generated concept into a usable pixel boss. The work lies not in a single generation but in choosing the medium, anchoring style, and iterating by hand: an existing enemy defines density, silhouette, and palette; PixelLab supplies seven animation states; grounding, scale, motion, and an existing area effect are then corrected and assembled. Failed approaches remain visible to mark the boundary between concept art and in-game assets.

Abstract

Watchman's first minute is established through black frames, a distant bell, fading brazier embers, and the awakening of the fire-marked knight. Production began with Grok concepts and keyframes, continued through five image-to-video segments and an edit in Jianying, then entered Godot as an OGV file. This record covers silent environmental narrative, grain boiling in slow generated shots, and the transition from high-definition cinematic imagery to a pixel-art game. It also documents a manual-first method: determine shot by shot which images survive motion before turning the successful decisions into a reusable pipeline.

Abstract

This entry documents Watchman's asset-management pipeline: heterogeneous packs are repacked into standard atlases, manifests record animation timing and foot anchors, and indexing, mapping, and composition tools keep hundreds of resources searchable and replaceable. The lasting object is not one asset set, but infrastructure that can absorb placeholders, later replacements, and continued expansion.

Abstract

This entry records the extraction of combat values from code into balance.json and the browser console built for centralized search, editing, and live comparison. Assertions lock derived relationships and valid ranges so tuning cannot quietly change program logic. The tool does not replace judgment; it shortens the distance between a decision, its validation, and a rollback.

Abstract

This entry organizes an AI development workflow that can run independently overnight. A bounded brief defines the work, automated tests form the merge gate, and adversarial review searches for judgment gaps outside the test suite. The final delivery includes code, verification results, and a change report, keeping unattended execution inspectable and reversible.

Abstract

This entry records Watchman's radial world structure: the hub concentrates characters and systems, bonfires anchor saves and death return, and regional lamps handle transit between areas. The first chapter remains linear, using distance, return paths, and node density to pace progression, so spatial structure carries navigation, risk, and narrative at the same time.

Abstract

This entry decomposes combat weight into tunable timing and feedback relationships rather than higher damage numbers. Stance, stance breaks, executions, parry windows, and turn speed jointly shape the rhythm of attack and defense. Frame-by-frame tests then calibrate hit stop, displacement, and recovery so that every exchange carries legible risk and reward.

Abstract

This entry explains why the combat system is concentrated in one iron sword and how build variation is moved into forging, transformations, scabbards, and whetstones. A single core weapon reduces animation and feel-maintenance costs while preserving a combinatorial space of moves, attributes, and special forms: a solo-development trade-off between systemic depth and production scale.

Abstract

This entry defines Watchman's direction as a side-view Soulslike set within an unending night, then records how a solo developer organizes several AI production lines from a director's position. Task briefs, test gates, and adversarial review establish the boundary of delegation, keeping design judgment human while implementation, validation, and documentation move in parallel.

Abstract

This entry divides an AI music video into eight stages: song, directing plan, stills, selection, animation, review, color, and release. Image prompts own appearance while video prompts describe motion only; animation tiers, filmstrip quality control, and anchor-based color matching then contain cost and inconsistency. Models perform production labor, while style, shot selection, and final acceptance remain human decisions.

Abstract

This entry begins with character design rather than a genre label and divides a Suno Style prompt into tempo and key, instrumentation, voice, harmony, dynamics, and negative constraints. Register limits, deceptive cadences, textural development, and A/B generations are used to resist the model's defaults. The central work of AI composition remains explicit trade-offs and repeated judgment about whether the result belongs to the character.

Abstract

Starting from a downward-sloping demand curve, this note derives the monopolist's marginal revenue and the two-step choice of quantity where MR equals MC, followed by price on the demand curve. The markup over marginal cost, the Lerner index, and deadweight loss together show how market power changes distribution and total surplus.

Abstract

Under price taking, homogeneous products, sufficient information, and free entry, this note derives the competitive firm's output choice. In the short run, price, marginal cost, and average variable cost determine production or shutdown. In the long run, entry and exit drive economic profit toward zero and the industry toward efficient scale.

Abstract

This note derives fixed, variable, average, and marginal cost from production technology and explains why marginal cost crosses AVC and ATC at their minima. Long-run average cost is then read as the envelope of short-run curves for different plant sizes, clarifying economies of scale, efficient scale, and capacity choice.

Abstract

This note uses the production function to connect inputs with output, separating total, marginal, and average product and explaining why fixed inputs eventually generate diminishing marginal returns in the short run. Isoquants and the marginal rate of technical substitution then express multi-input choice geometrically, providing the technical basis for cost minimization.

Abstract

This note combines preferences and the budget constraint into the consumer's optimization problem. It derives the tangency condition for an interior solution, the equal-marginal-utility-per-dollar rule, and the interpretation of the Lagrange multiplier. Corner solutions, kinks, and nonconvexity show why MRS equals the price ratio only under specific conditions.

Abstract

This note reconstructs the budget line from income and two prices, then distinguishes the parallel shift caused by an income change from the rotation caused by a change in one price. It translates the diagram back into relative prices and purchasing power, establishing the constraint used in later analysis of consumer choice.

Abstract

This note explains when completeness, transitivity, continuity, and related conditions allow preferences to be represented by a utility function, while keeping ordinal utility separate from a measurable psychological quantity. Indifference curves and the marginal rate of substitution then become geometric expressions of ranking, preparing the ground for budget constraints and optimal choice.

Abstract

This note uses the within transformation to show how fixed effects make each individual or group its own control, removing every unobserved factor that does not change over time. The cost is that time-invariant regressors cannot be estimated and between-group information is discarded; identification comes entirely from change within the same group.

Abstract

This note begins with the endogeneity that makes OLS inconsistent and shows how relevance and the exclusion restriction allow an instrument to restore identification. It connects the first stage, reduced form, and two-stage least squares, while emphasizing weak instruments and the untestable core of exclusion: computation can be automated, but credibility still comes from research design.

Abstract

This note explains why heteroskedasticity-robust standard errors change the covariance matrix while leaving OLS coefficients untouched. HC0 through HC3 mainly differ in finite-sample adjustment and their treatment of high-leverage observations. Robust errors repair inference; they do not make an inefficient estimate efficient or solve misspecification and endogeneity.

Abstract

This note separates the variance problem that weighted least squares can address from the endogeneity problem it cannot. When weights approximate inverse error variance, WLS can improve efficiency and restore valid inference under heteroskedasticity. Bias from omitted variables, reverse causality, or measurement error does not disappear merely because observations are reweighted.

Abstract

This note explains how to read a local coefficient from geographically weighted regression: it is a local association conditional on a chosen bandwidth and weighting scheme, not an automatically valid local causal effect. Coefficient surfaces can also reflect collinearity, edge samples, and multiple testing, so they need global benchmarks, stability diagnostics, and a stated spatial mechanism.

Abstract

This note organizes the LM decision rule that moves from OLS residual diagnostics to a spatial-model choice. Classical and robust LM-lag and LM-error results help distinguish lag-like from error-like dependence. They provide comparative evidence rather than an automatic answer, and must remain consistent with the mechanism, weight matrix, and residual diagnostics.

Abstract

This note compares the spatial error and spatial lag models, then uses the more general Durbin form to show how the two can be nested. The central distinction is whether dependence enters the disturbance or the dependent variable, and whether feedback and spillovers follow. With a spatial lag, coefficients must be decomposed into direct, indirect, and total effects.

Abstract

This note places global Moran's I beside local indicators of spatial association. The global statistic asks whether clustering exists overall; the local map locates high-high, low-low, and spatial-outlier areas. The emphasis is also interpretive: significance, multiple testing, and the fact that a cluster map alone cannot identify a mechanism or a causal relation.

Abstract

This note treats the spatial weight matrix as a research-design choice that must be defended, not as a software default. Contiguity, distance, and nearest-neighbor rules encode different mechanisms. A useful W follows the theory, scale, and data structure, and its conclusions should be checked against plausible alternative definitions of neighborhood.

Abstract

This note separates spatial dependence from spatial heterogeneity. The first concerns correlation between neighboring observations; the second concerns relationships whose parameters vary by location. Because they call for different diagnostics, models, and interpretations, confusing them at the outset can distort every later choice, from the weight matrix to the final claim.

Abstract

This note treats a dwelling as a bundle of location, size, quality, and public-service attributes, then shows how coefficients in a log-linear hedonic model become implicit prices. It also marks the identification boundary: a proximity coefficient describes market valuation, but does not by itself establish the causal effect of building an amenity.