October 6 Step5 original animation restoration
The pinned Variora Step5 HTML and full MIT license were reviewed to restore the existing coastal SVG animation. Unchanged bytes and the full license remain on the isolated domain. Background-browser tests verified automatic wheel, leg and parallax movement. Pause/resume and speed adjustment are playback controls only, not a Play demo. The original date, number, cover and media are preserved; no work is added.
October 6 historical creator SVG preview
One GPT 5.1 code-generated work published by @SkyeSharkie on November 21, 2025 adds one independent work and one timeline representative. The original post maps the SVG-code prompt and model to the still attachment. The complete original 600×400 PNG is preserved without cropping, redrawing or fabricated motion; it is not the SVG source file. The date is the UTC publication day, not an invented generation date or effort setting. One-shot is an author report only. Hash and visual-neighbour review found no same work. Attribution, original source and bytes remain available; no open license, code mirror, playable demo or scored reference is added. The same post’s Gemini 3 original animation is reviewed separately, never borrowed for this preview.
October 5 original historical SVG previews
Continuing original 2024–2025 post research recovered two code-generated works published by @aiwarts on October 14, 2025: Ring-1T and DeepSeek V3.2, both timeline representatives. The original post explicitly maps the left and right images to the SVG prompt. Two separate original JPG previews are preserved without composites, crops or redraws. Only static previews are available: no motion is fabricated, and unrelated tests in the thread are excluded. Dates are original UTC publication days; generation time, reasoning effort and iterations are unpublished. Hash and visual-neighbour review against 1,397 archived decodable media found no duplicate. No open licence is claimed, no unlicensed code is mirrored, and no playable demo or scored reference is added.
October 5 Nebius historical code works
Eleven August 2025 SVG code-generation works were verified in pinned first-file commits of the official Nebius cookbook, adding three timeline representatives. Thirty-three original PNG, model-labelled PNG and SVG files retain the complete MIT licence and Nebius B.V. attribution; attachments are not extra works. Dates are original first-file commit UTC publication days, with actual generation dates unknown, never collection or model-release dates. File numbers do not imply reasoning effort. Four same-day repeated-response groups lack a documented default or medium, so no best output is guessed for the timeline. Hash and visual-neighbour review against 1,364 archived decodable media found no duplicate. Original failures are retained, with no new Play demo or scored reference.
What this is
Pelican bicycle works generated as code by language models, across all dates and presentation media, excluding direct text-to-video outputs. Source-reported models, creators and original assets are retained. Benchmark scores are separate references, not a museum ranking.
October 5 original motion restoration
The pinned Variora MIT license and complete SWE-2, MiMo V2.6 Flash Free and Grok 4.7 HTML originals were individually reviewed to restore three existing code-generated works’ original animations. Unchanged bytes and complete licenses remain on the isolated preview domain. Browser checks confirm moving wheels, legs and backgrounds matching each original screenshot. Artwork dates, IDs, covers and original media are unchanged; restoration does not add works. SWE-2 clicking changes playback speed only; the MiMo and Grok originals have no interaction controls. None adds a Play demo. Other downloaded originals still require individual review and are not declared restored.
Scope and counts
1758 independent works, including separate outputs at different settings. The 1550 classic works on the timeline across media are a subset, not additional works. 144 Simon Willison reading-index entries are not artworks. Raw JSON/CSV retains 4086 archival rows, including 2187 reference-only HF samples; items.length is not the artwork count.
One output, one record
One work is one source-attributed model output at a documented time. Labelled compilations are split only when each image can be matched to its model; faithful source crops retain provenance and full originals in details. Settings comparisons precede the individual image and complete composite. Reposts and duplicated images are not counted twice.
Evolution timeline
Every counted classic work enters the timeline across settings, independent runs and iterations; derivatives are listed separately. Exact model versions are folded by default, and reviewed source run groups can be expanded inside them. Author-default or medium choices only determine the initial folded display; missing default evidence never excludes a work. Source dates and model labels are preserved, without inventing releases, generation times or author preferences.
Benchmark references
HuggingFace / OpenEnv dataset and scoring environment share one Benchmark collection. Benchmark collections, model summaries and samples are referenceOnly: excluded from the main collection, timeline, artwork totals, numbering and default API search. Original IDs remain readable. Upstream outputs, not a Pelican Bicycle Archive ranking.
Code-generated works across all dates and media
The focus is language-model code generation: SVG, HTML/CSS/JavaScript, Canvas, Three.js/WebGL, Blender scripts, POV-Ray scenes and related documented code workflows. Presentation may be static, animated, recorded video, 3D or interactive. Direct text-to-video output such as Veo or Sora is outside the main collection; preserved legacy IDs and original media remain uncounted context, readable by direct ID. Classification follows source evidence about the generation workflow, not the media extension. A recording or frame of code-generated animation remains part of that same eligible work, not another work. New imports require generationMethod=code-generated and public codeGenerationEvidence; missing process evidence is deferred, not guessed. Models and tools remain source-reported. Play requires meaningful interaction and clear permission on the isolated demo domain.
How records are handled
We preserve public sources, dates, creator names, source-provided model labels and media. Model names reflect their sources; they are not independently certified by this site. Missing prompts are identified rather than reconstructed. English notes are translations of the archive summaries; source links and original creator names are retained.
October 2 presentation update
All works and the timeline share model-version order. Missing release evidence uses the earliest source-work month as a provisional position without a separate pending label; original artwork dates and numbers are unchanged. The homepage features the human-made 1988 film pelican, with sourced bilingual production history. It is not an AI work or proof of the AI prompt’s inspiration and adds no artwork count.
October 3 intensive historical code-work intake
Fixed original repository commits, author blogs and X originals yielded 44 earlier works: one from 2024 and 43 from 2025, adding 41 timeline representatives. Robert Glaser’s 13 visual-iteration runs each count once; all 67 original process JPGs remain without counting every revision as a work. Explicit screenshot model labels establish 14 denamwangi outputs; Harald Nezbeda’s GPT-5 comparison yields three, and historical X originals provide the remaining 14. Original publication dates, exact source labels, complete originals, faithful crop coordinates and hashes are retained without redrawing failures. Comparison against 1,259 decodable old assets excludes existing animations and reposts. This batch contains only genuine published still previews, not invented motion. No unverified-permission code is mirrored, and no Play or scored Benchmark collection is added.
October 3 X code-generated effort comparisons
The original public author article documents SVG generation through Claude Code. Explicit labels establish four settings each for Opus 5.5 and Opus 5: eight actual published still previews. Faithful pixel crops retain coordinates and original/crop hashes; the full eight-panel source remains a detail attachment. Three reported runs per setting do not become 24 published works. The timeline uses the documented author defaults, medium and high respectively, adding two representatives; details start with same-model setting comparisons. September 23, 2026 remains the original UTC publication day, not the collection or model-release date. Author latency, cost and success summaries are not a museum ranking. SVG source code was not recovered; unverified-licence code is not mirrored, and no Play or Benchmark data is added.
October 3 original Gist previews
Pinned author Gists and same-author comments yielded three real May/June outputs: one DiffusionGemma preview and separate grok 4.3 bike/bicycle prompts. Original published PNGs, attribution, UTC publication days and hashes are retained; the collection date does not replace artwork dates. The Grok follow-up is a prompt comparison, not reasoning levels or a verified fresh session. Without a documented default/medium, both remain in All works only; DiffusionGemma enters the timeline. Two original SVGs fail parsing because of duplicate attributes. They are not repaired, redrawn or mirrored without verified permission. This batch adds real source previews, no animation, Play demo or Benchmark data. Repository candidates lacking exact-model attribution or original previews remain deferred rather than filling a quota.
October 3 code-generation scope update
The museum focuses on language-model-generated code. SVG, web animation, 3D scenes, interactive projects and genuine recordings of those code-generated works remain eligible; a video extension is not an exclusion. Four clearly identified legacy direct text-to-video records become uncounted context. Their IDs, sources, dates, details and original media remain accessible by ID, but not in main works, timeline, display numbers or default search. Four unpublished Veo 2 split outputs are withdrawn, with research evidence retained privately rather than counted as new works. Future intake requires an explicit code-generation method and public source evidence. Unknown workflows are deferred; model capabilities and dates are not guessed. Homepage copy, motion presentation, Play and Benchmark boundaries remain unchanged.
October 2 earlier X works and public-source review
Four X authors’ original posts and OrcaRouter’s same-author announcement reply yielded five independent works and five timeline representatives: one complete cycling animation, two explicitly labelled photorealistic SVG preview crops, and separate Space Bunny Alpha previews by two authors. All 395 original video frames decode with real motion; a real one-second cover and muted automatic playback are used. Crop pixels match the source, with the full comparison only in detail attachments. Anonymous/routing labels remain unresolved rather than assigned a guessed provider; UTC publication dates are not generation dates. A public 73-drawing gallery remains queued for per-output date evidence, not bulk-imported for its size. Reposts and unresolved labels do not add works. No code, new Play demo or Benchmark expansion is mirrored.
October 2 three-model original animations and repository continuation
Cat’s original X post and same-author reply yielded four independent works: high run 1 animations labelled grok-4.7, grok-4.6 and gemini-3.8-flash, plus a failed grok-4.6 run 2 preview. Each lossless panel crop has 300 decoded frames matching the source video pixels; real covers and the folded complete original are retained. Run 2 has only the author’s static screenshot, not an invented animation. Neither grok-4.6 run has a documented default/medium representative, so both count in All works only; the other two enter the timeline. Model labels and UTC publication dates are source-reported, not authenticated identities or generation dates. The author’s disclosure that Grok used file tools is retained, not called a strict tool-free run. Metadata for 28 repository candidates was checked, but original previews or unified code permission remain missing; two other projects have unresolved rights/model evidence. No repository code is executed or mirrored, and Play and Benchmark are unchanged.
October 2 labelled comparison splits and animation follow-up
EvoLink and Haleemah public originals and author replies yielded six independent works and six timeline representatives: two Blender animations and four source-labelled SVG still previews. The consistently labelled two-model video was separated into single works; all 150 decoded frames per work match their source region. The four-panel image was faithfully cropped according to explicit labels. Ten media and real cover files are archived, with complete originals collapsed in details, never as main composite covers. Frames and reposts are not additional works. Dates are original UTC publication days; model labels are source-reported, not independently authenticated. Promotional superiority claims are not museum conclusions. Still-only previews do not invent motion. Code permission is unknown; no code, new Play demo or Benchmark expansion is mirrored.
October 2 forum follow-up and X originals
Seven forum topics and their replies plus two X authors yielded 16 reviewed works and 19 original media/frame files: five original GIF/WebP animations, one complete video and ten real still previews of dynamic webpages. Model labels remain as written, without expanding astra, 6pro or unverified routing identities. MiniMax visual iterations and one Grok edit are documented, not called one-shot generation. Day/night scenes, before/after views, same-work replies and video frames are not extra works; an already archived GIF is not added again. Two source-documented derivative prompts are identified explicitly. Chrome was used read-only for public posts and author replies. No unlicensed code, new Play demo or Benchmark expansion was mirrored. Unknown generation dates and dates of deleted earlier posts remain unknown; repost publication is not a new generation date.
October 2 community animation and source splitting
Two LINUX DO original topics, a pinned Pelecanus revision and five articles linked by Nile entries 56–60 yielded 45 independent works: 29 original animated WebPs, three dynamic-webpage still previews, eight labelled Three.js scene crops, two February SVG previews and three single outputs resolved from previously uncounted archives. Fifteen representatives enter the timeline. Twenty-nine repeated runs have no documented default; the enhanced prompt uses the existing classic representative. No best output is guessed. Complete originals, crop coordinates, hashes and the Apache-2.0 licence are retained. Three already-counted matching outputs are not added again; three archive splits reuse existing media. Recordings and reposts are not extra works. All animation frames decode; original failures with a missing pelican are not redrawn. The author’s suspected low/medium mapping problem remains explicit. Models, settings and generation dates are not independently authenticated. No code is mirrored and no Play demo added.
October 1 community all-media batch
Twenty-six independent outputs from ten LINUX DO topics and their replies were source-checked and archived: 2025 comparisons, 2026 animation and 3D previews, with 36 original media files and source-reported model labels. Eighteen enter the timeline; eight same-day comparison outputs have no documented default setting or default run, so they count as works without a guessed best representative. Projects with unverified code licences are not mirrored as Play demos.
Public X originals in the same batch
Two public original posts were checked in an authorized signed-in browser, yielding five static SVG previews and five timeline representatives. A labelled four-model image was split into four works; two screenshots of another single output count only once. Three complete originals and five faithful pixel crops are archived with coordinates and source evidence. A future animation mentioned by its author was not published in that post and is not presented as an animation or Play demo.
Original-source and cross-media follow-up
Twenty-five original articles linked by Nile entries 31–55 were reviewed. Forty-four works were added: 25 single-model Simon outputs, 15 original beetle_b POV-Ray renders, three MIT-licensed Sonnet 4.5 visual-feedback versions and one explicitly labelled GEPA / Opus 4.6 zero-shot baseline. They provide 32 timeline representatives and 58 original media files. Fourteen final SVGs were visually matched to source previews. Reposts, previously archived images and unresolved multi-model outputs are not added again. POV-Ray error correction and visual feedback are documented, not presented as one-shot generation. Twelve same-day settings/runs/iterations lack representative evidence or are non-representative alternatives and remain only in the collection. The optimized GEPA image is held pending model attribution. Log timestamps, publication dates and explicit dated updates are preserved; public commit dates are not authenticated generation dates. No project code or new Play demo is mirrored.
Maintenance
Latest October 1 authorization: the collection and timeline accept all dates and media; collection rules, bilingual scope, SEO and Agent/API descriptions are synchronized. Interactive mirrors still require permission, safety and operation review. September 27, 2026: accessible media, sources, bilingual pages and machine-readable catalogs archived. October 1: HF/OpenEnv moved to separate Benchmark references; labelled compilations and community grids split, five historical static SVGs added, original IDs and media preserved, model-family filtering introduced. Bilingual selection, counting text, SEO metadata, structured data and Agent/API guides updated together. Later historical-link batch: 37 submitted entry points checked, 36 missing static outputs archived and two existing DeepSeek comparison outputs split into works; one duplicated Qwen archive pair merged for counting. Dates preserved, only source-default/medium or page-default runs represent the timeline; broken SVGs are not repaired or counted. Early-timeline follow-up: original Simon articles linked by Nile yielded three additional static outputs: two Gemma 3n quantizations/runtimes and Grok 4. The Mistral Small 3.2 illustration duplicates an existing X output and is not counted again. No default/medium Gemma representative is documented, so both remain outside the evolution axis. Original previews and source labels retained; publication dates are not claimed as generation dates. Second early-timeline batch: Nile entries 21–30 were traced to original articles and yielded nine July–August 2025 static works and nine dated representatives: Kimi K2, Qwen3, GLM 4.5/Air, XBai o4 and Claude Opus 4/4.1. Five include verbatim final SVG code and four retain source-published static previews. Per-response timestamps or publication-date bases are preserved; existing Qwen output, reasoning screenshots, videos and games are not added again. Remaining entry points still require review; coverage is not claimed complete. Media availability is not independent reproduction.
Detail presentation update
October 1, 2026: existing moving media and reviewed on-site demos now appear before stills. Same-model comparisons precede the individual image; originals and provenance remain available in collapsed sections. Chinese and English share the layout. Year shortcuts, counts/page summary and the archive entry have been repositioned; the view button names the current mode. Artwork and timeline totals are unchanged.
Model release ordering and media galleries
October 2, 2026: timeline ordering now uses separately sourced first public availability of model versions, while cards retain original artwork dates. Versions are expanded by default, with optional exact-version folding; unverified releases stay separate. Detail galleries show real timestamped video frames or multiple archived images with horizontal navigation. Frames are not new works; the right information panel and existing notes stay in place.
October 2 Reddit and X original-source backfill
Nineteen independent works were verified in public original posts and author replies, with 34 archived original media files, faithful crops and real-frame covers (including three canonical PNG copies with original bytes and earlier links retained). These include two 2025 AI Studio previews, individual SVG comparison images, and seven playable animation/Blender recordings. Explicitly labelled two-model recordings are split losslessly; decoded frame hashes match the original regions. Complete comparison videos and images remain folded in the details. Four high outputs comparing free/paid access and prompts are not reasoning levels and have no documented default condition, so they count only in the collection; fourteen representatives enter the timeline. One differently encoded X animation matches an existing Reddit work and is not counted again; reposts are traced to the original creator. Source labels 2.5 and Astra are not expanded into invented full model names. Chrome reads public content only. No code mirror, Play demo or Benchmark is added.
October 2 historical forum replies and animation
Seven historical Reddit posts and author comments plus one fixed-version GitHub project yield thirteen independent works and thirteen timeline representatives, including six 2025 outputs and three animations. Seventeen original files, faithful crops and real-frame covers are archived. Two explicitly labelled models are split from one comparison; complete comparisons and same-work stills remain folded in details. Models, prompts and quantization are verified comment by comment without inheriting undocumented generation conditions from the main poster. GPT-5 and Gemini 3 Deep Think reposts are not counted again, and April reuse of a February comparison does not change its dates. A separate X video has conflicting model labels between text and picture and remains deferred, without guessed attribution or counts. Chrome reads public content only. No code mirror, Play demo or Benchmark is added; models and generation dates are not independently authenticated.
October 2 community animation and settings splits
Thirteen independent works and nine timeline representatives are verified in public posts and author replies. Five original animation recordings play on-site; three other works are described as animated by their source but have only still previews, explicitly distinguished. Three existing collection videos are reused as individually labelled model outputs, preserving the old ID and media; two new original videos are archived. Five Qwen settings are mapped to the same author: only medium represents the timeline, while all settings count in the collection and appear in detail comparisons. A model identity claimed on X without provider/session evidence remains an author claim, not certification of a version or promotional conclusion. Fifteen new media files and real-frame covers are saved, avoiding Luna’s pelican-free opening. Dynamic files come first; playback controls are not Play. Comments lacking an actual output or full attribution remain deferred. No code mirror or Benchmark expansion is added.
October 2 animation originals and individual works
The original four-panel Opus 5.5 recording is split into four labelled animations: all 180 decoded frames per crop match the source regions. The old collection ID and full recording remain, without double counting. Detailed-prompt visual refinement in medium is not a fourth reasoning level; the original medium represents the timeline. One new X animation still preview is added without claiming recovered playback. This batch adds four net independent works and one net timeline representative. Three existing Variora works now play byte-for-byte on the isolated on-site domain after pinned MIT permission and source-safety checks; pause/speed controls are not Play and do not add works. Existing media, dates and covers remain, with no Benchmark expansion. Another new recording lacks a full model label and remains deferred; source model/routing claims are not independently authenticated.
October 2 automatic motion previews
Cards with actual moving originals start when visible, across the timeline, all three collection views and both detail languages. Videos loop muted; off-screen media, background tabs and closed attachments pause. A pause/resume control and the system reduced-motion preference are supported. Source animation code stays byte-identical on the isolated domain, with no automatic navigation to external sites. Interactive details preserve author controls and game state. Still-only archives do not gain invented motion. Work, timeline and Benchmark counts are unchanged.
October 2 X original-post continuation
Read-only public Chrome source checks add two original Sonnet 5.5 animations by Jack and STEVExKONG and one single-work preview by Nikita: three independent works and three timeline representatives. Both original videos fully decode and have real frame covers, with muted automatic playback on cards and details. The screenshot-only work does not gain invented motion. Dates retain the original posts’ UTC publication days, not generation days; model and effort labels are source-reported, not independently authenticated. Ebi’s repost of an existing work is visually deduplicated; model self-identification screenshots and quoted announcement videos are not additional works. Original bytes and old IDs remain, with no code mirroring or Benchmark expansion.