EDITORIAL PACKAGING UPDATE — ORIGINAL-COMPOSITION FILM The revised invitation is: Three AIs wrote piano music. Which would you hear again? The new promotional thumbnail is a native image-generation illustration guided by independent Sol review, not Argon output, a composer-designed cover or a simultaneous performance frame. The independently designed original covers remain in the film; the video and frozen music did not change. No winner, human-performance or equal-compute claim follows from the artwork. A recognizable difficult piece is better handled as a separate follow-up. Playing a supplied score tests interpretation, sequencing and tool control; audio transcription is another task. Neither is the original-composition question asked here. Any future test needs matched source, tools, instrument, tempo/expression and retry rules, plus separate checks of composition, exact edition/arrangement and recording rights. No such performance has been commissioned or run. Copyright Office source: https://www.copyright.gov/engage/musicians/ COMBINED-FILM UPDATE — PUBLISHED LISTENING DEMONSTRATION The completed combined film is345seconds, ordered Terra,Astra,Opus. It preserves each108second frozen score plus the complete3second source decay, with shared landscape falling-note graphics, brief cards and one outro. No composition was regenerated. The earlier separate117.5second edits and raw website takes remain preserved; checks of those files are historical evidence, not checks of this combined output. Both combined encodes passed full decode,all9990performance-frame key states,source-audio checkpoints,complete-ending checks and real-speed muted player playback. Independent human listening and physical-phone acceptance have not been recorded. The combined film and this article are released with those limits disclosed; there is no YouTube upload. PIANO EXPERIMENT — DETAILED METHOD AND FIRST ATTEMPTS Historical first-attempt and recording notes follow. Independent human listening is not recorded. ## What the models actually do In the original browser recordings, the models choose the notes and their timing. A script sends keyboard presses to OnlinePianist's virtual piano, and the website produces the sound. We capture the actual site audio and keyboard display. This is composition through an automated browser instrument; it is not a model moving human fingers across a physical piano. We retain the event schedule, its checksum, the recording scripts and the raw capture. The event schedule is the musical source. We have not recovered a separate MIDI file and do not describe one as an existing deliverable. ## The first round: two miniatures The original shared challenge allowed up to 30 seconds and no more than three held keys. Both models declared 24-second pieces. Astra's “A Light Left On” used 53 notes; Opus's “Lighthouse” used 56. The pieces were presented as A and B before the names were revealed. The local blind-review records show a disagreement. Across four order-alternated runs, Gemini 4 Argon selected B, the Opus piece, four times. Qwen 3.8 Omni Flash selected A, the Astra piece, three times and B once. These are repeated responses from two model judges, not eight independent human listeners or audience votes. The channel owner's recorded response was: “both good, hard to rate, a bit short.” The experiment had limitations that matter. Astra's identity was supported by its local Codex configuration; independent provider attestation was unavailable. Opus's retained Claude Code metadata identified `claude-opus-5-5`, but its effort setting was unknown. Opus had already read the production protocol in that session, and tool parity was not independently established. We should not turn that setup into a controlled ranking of the models. There was also a real technical repair. Astra's first recording accidentally used half sustain because the site's pedal control cycles through full, half and off. That take remains preserved. The replacement replayed the identical frozen score with sustain verified off. Opus had two setup stops before any notes, then one musical take; its retained log records a brief timing delay during a run of notes. The performances were not secretly recomposed to improve the result. “No retakes” is too broad a description of the first round and is corrected here. The blind edit used documented plain gain to match listening levels. It also required an editing fix when an early export accidentally doubled one audio source. Those editing corrections are separate from composition or performance attempts. ## The longer solo brief For Round 2, the central request is: compose an original solo-piano piece people would choose to hear again. Aim for 108 seconds from the first onset to the final release, with 105–110 seconds acceptable, followed by three seconds of natural decay. Choose the title, mood, key, meter and tempo. Introduce a memorable theme, develop it, and return or resolve it deliberately. Do not quote a known tune or imitate a named artist. All three episodes use the same locked musical brief, changing only the necessary performer and counterpart names. The instrument is OnlinePianist's Concert Grand, sustain off, transpose zero and metronome off. Up to six keys may be held at once; released-note decay does not count. The model must stay within the verified keyboard range. Creative work has a 20-minute active limit, excluding technical setup and export. The first complete score is frozen before playback. One full take is allowed, plus at most one technical replay of the identical score if capture fails. A later musical revision would be a separate version. These boundaries preserve what happened without pretending that matched instructions imply matched computation: the products and tool interfaces differ, and comparable token or runtime budgets are not available. ## Astra: “Lanterns After Rain” Astra's first complete Round 2 composition contains 314 notes and lasts 108.003 seconds in the captured performance. Its declared form is 36 bars in 4/4 at 80 BPM, moving from D minor to D major. It uses at most five held keys. The score was frozen before one full take; no performance repair was needed. Four failed edit/export variants were retained while the video timing and padding were corrected; those were editing failures, separate from the unchanged composition and single musical take. The executing task was explicitly selected as `gpt-6-astra` with high effort. That is task-control evidence, not independent provider attestation. The recording uses the site's sound and real keyboard highlights. The vertical edit shows two synchronized crops of the same keyboard performance to make both registers readable. Its finished private Short lasts 117.533 seconds. The introduction reads “Astra — original piano performance”; the outro asks “Would you replay this?” Technical checks found no full-scale clipping or unexplained silence during held notes. A sampled first audiovisual onset differed by about 15 milliseconds. Source-to-export audio checks and normal-speed browser playback passed. These measurements do not establish musical quality, and human listening acceptance remains outstanding. ## Opus: “Lamplighter” Opus's matching first composition is “Lamplighter”: 341 notes across a declared 31-bar form in 4/4, starting in E minor and ending in E major. Its main tempo is 70 BPM, with slower transitions and a closing ritardando. The frozen score runs for 108 seconds and reaches the six-held-key limit. During score generation, four note releases were shortened to meet that limit; those adjustments are recorded in the generator's report before the first complete score was frozen. The run uses Claude Code through Claude Pro, with Opus 5.5 selected at high effort and usage credits verified off. It received the same musical brief and the same neutral capture scripts in an isolated directory, without Astra's score, title, output or critique. The session retained its concept, generator and first score. Its log corrects an initially late creative-start entry: composition thinking began after preflight, around four to five minutes before the freeze, rather than when the written concept was saved. The first take completed without a capture repair: 341 key-downs and 341 releases were retained, with no keys left held. Page-observed keyboard events put the music at 108.005 seconds. The raw stereo recording peaked at about −1.33 dBFS with no full-scale samples; signal checks found no unexplained silence during held notes. Frame and audio measurements put the typical attack about 28 milliseconds ahead of its visual highlight, subject to the recorder-start clock's uncertainty. The private Short lasts 117.533 seconds at 1080 × 1920 and 30 fps, matching Astra's layout and intro/outro lengths. Seven source-to-export audio checkpoints correlated above 0.9999 with no detected timing lag at those samples. The edit adds no narration, backing music, gain adjustment or note changes. Its roughly three-second ending allowance includes a short silence pad after the complete recorded decay: the raw audio file retained about 2.95 seconds after the final dispatched release. That padding is an export detail, not a new performance or a musical repair. The longer solos use existing subscriptions and local capture/edit tools, with no new paid API calls. This does not price the subscriptions or machine time; the records do not establish equal token or compute budgets. The earlier duel's packaging had separately recorded paid production costs. The original Astra and Opus edits also passed headless normal-speed browser playback; the Opus player reached all 3,526 frames with no dropped frames or errors. Human listening acceptance of the new performances remains outstanding. ## Terra: “The Lantern Keeps Its Place” Terra's first complete score contains 197 notes across a declared 36-bar form in 4/4 at 80 BPM, with C major and A-minor color. The producer describes a steady lyrical phrase returning through theme, answer, development, return and coda. It stays within three held keys, below the common limit of six. The Terra run used a fresh Codex task configured as `gpt-6.1-terra` with high reasoning and no inherited turns. That is task-control evidence, not independent provider attestation. The neutral packet contained the locked brief and the same preflight/capture scripts, without the other pieces, titles, artwork or critiques. Ordinary-usage and overflow settings were not independently attested; the producer reports no paid API or model fallback. We do not infer a billing guarantee from that record. The initial preflight failed when sandboxed Chrome could not launch; the unchanged preflight then succeeded through approved host execution. Two generator validation failures occurred before a first complete score existed. Those are setup/validation events, separate from the single frozen composition and one full musical take. No musical or capture replay was used. The original recording retained 197 key-downs and 197 key-ups, with all keys released. The producer's log measures a 108.014-second action span and a full three-second tail. Decode passed; its reported peak was −4.9 dBFS, with no unexpected silence at the stated detection threshold. A coarse first-onset check put the highlighted key about 17 ms after the first detected signal; it does not certify the entire recording's synchronization. The recorded creative window, 16:03:00–16:03:47 UTC, describes the logged concept-to-freeze interval, not end-to-end model latency or an equal-compute comparison. Terra's replacement listening copy follows the same frozen notes and shared instrument/visual treatment as the other two. Human listening remains pending. ## New listening copies: unchanged scores, a different instrument The original website takes remain retained as evidence of what each model did. For the new listening drafts, we rendered those same frozen event files through GeneralUser GS 2.0.3 using the already-installed FluidSynth 2.4.4 engine. We did not ask any model to compose again or change any pitch, onset or release. This is a replacement-instrument rendition of the compositions, not another model attempt or the original browser performance. All three use the same bank 0/program 0 “Grand Piano” preset, fixed velocity 100, 48 kHz stereo output, engine gain 0.2, and disabled reverb/chorus. A common +12 dB linear gain is applied to all three edits; there is no per-piece normalization, compression or limiter. That preserves differences between the scores under one stated rendering policy. The event files do not carry expressive velocity data, so this treatment recreates their pitches and timings rather than claiming to reproduce every quality of the original website instrument. The new videos use original keyboard graphics driven by the same event schedules, with each model's independently designed cover in the 2.5-second opening. All three include 108 seconds of notes, 3 seconds for complete decay and a 4-second outro, for 117.5 seconds total. No OnlinePianist audio, screen capture, interface graphics or other website assets are used in these replacement versions. The original website takes and earlier edits remain intact. GeneralUser GS's [official licence 2.0](https://github.com/mrbumpy409/GeneralUser-GS/blob/684543d5e5efaef08d02be50dcda8d552478fa60/documentation/LICENSE.txt) explicitly permits private and commercial music creation and profits from recordings. Its author also discloses uncertainty about some inherited samples' original provenance. We retain that caveat rather than treating the licence as a legal guarantee. The exact downloaded asset, licence text, pinned repository revision and checksums are retained with the production records. No new application, paid service or model call was needed for the replacement render. The three separate replacement videos were prepared for private listening review. Those 117.5-second videos passed full decode, clipping checks and seven source-audio comparison checkpoints. Sampled keyboard highlights align within one 30 fps frame of the score. All three played through all 3,525 frames at normal speed without player errors or dropped frames. Technical measurements and a successful player check cannot decide whether the instrument suits the music; human listening remains necessary. ## Each model designs its own cover We also asked each performer to design original vector artwork for its own episode. All three received an identical 1080 × 1920 canvas brief with reserved safe areas; no other performer’s Round 2 art, title or score was supplied for its design task. We then applied the same AIHackers information layer: the outlined mark and domain, model name, Round 2 label, and 108-second performance information. Only the model name changes in that shared layer. [The original covers appear in the combined film opening and chapter cards.] Astra chose a paper lantern above broken rainwater reflections that suggest piano keys. Its proposed title is “Lanterns After Rain — Astra's Original Piano Performance.” [The original covers appear in the combined film opening and chapter cards.] Opus chose a waterfront promenade whose lamps follow the opening melody, with a lamplighter at the bright final lamp. One proposed title is “Lamplighter — an original piano piece by Claude Opus 5.5.” [The original covers appear in the combined film opening and chapter cards.] Terra chose a steady amber lantern above shifting ridgelines, with a curved keyboard partly submerged in the scene. Its proposed title is “The Lantern Keeps Its Place.” The producer retained the first SVG without rendering it; the coordinator rendered and inspected it, with no repair needed. All three chose light-related titles or imagery. We checked the exact locked briefs: they did not prescribe lantern, lamp, light or rain imagery, and the supplied Terra packet excluded the earlier musical outputs. That is an observation about these attempts under a shared task, not a measured rule about model creativity or proof of independent originality. These are original illustrations interpreting the pieces, not footage of the performances. We retain each first SVG, its checksum, rationale and optional copy before coordinator branding. There were no artwork repairs. The product sessions and available computation differed, so this design exercise is a creative demonstration rather than a benchmark or leaderboard. ## What a listen can tell us A recording can answer whether a particular setup produced a complete piece, respected the note and timing limits, and survived a real browser performance. A listener can decide which theme they remember or which ending they prefer. Those are useful, different observations. Neither a loudness reading nor a note count measures beauty. A repeated model judge preference is not an audience poll. A small set of performances cannot establish a universal winner in music, coding or general intelligence. The invitation is narrower and more interesting: listen to a complete idea, then decide whether you would replay it. ## Recording and release status The Round 2 OnlinePianist sound/UI recordings remain private: publication permission for those service materials is unresolved. The replacement-instrument drafts use GeneralUser GS and original graphics under the separate method above. The GeneralUser GS author’s inherited-sample caveat stays in the record. The three separate portrait edits remain private. The public article instead uses one combined landscape film, Terra → Astra → Opus, with complete endings. Technical and editorial review are recorded; independent human listening and physical-phone acceptance are not recorded. The owner reports that the first duel is already posted. We have not recovered and verified its exact public video URL, so this article contains no guessed Round 1 embed or video publication date. No YouTube publication is implied by the first-party combined player.