All 6 output cells match.
One color occurs in exactly one connected object. The selected rule crops that object's bounding box while retaining every color inside. All three training outputs and the illustrated test output match.
AETERNA / INDEPENDENT RESEARCH
Show the result.
Keep the limits visible.
Aeterna is my independent research project around object representation, symbolic problem solving and lasting value from experience. These two films distinguish recorded results from the wider research direction.
These films visualize recorded outputs from selected tasks; they are not live solver screen recordings. Captions are already visible in the video.
RECORDED RESULTS
One exact result. One explained failure.
Shape correct. Color wrong.
The occupied and empty cell positions match. However, 26 cells have color code 2 instead of 7. Training fit is 2 of 5 examples. This is not an exact answer: the color-assignment failure is shown explicitly.
Two selected, public ARC-AGI-1 tasks. These are not an overall accuracy figure, a generalization result or an independent competition score.
SOFTWARE RECORD / RESEARCH DIRECTION
Separate working components from research goals.
469
passing software tests
One archived run from 14 September 2026, with 3 warnings. This is not a count of solved ARC tasks or an intelligence score.
Software records from April and May 2026 document specific development milestones. My broader research question is how a system can retain useful experience and distinguish which contributions helped subsequent work.
I direct the project, define requirements and choose evaluation criteria, using AI tools in software development and review. Narration was AI-assisted using my own authorized voice recording.
Keeping records does not, on its own, prove learning or progress. These films do not claim general intelligence, consciousness, unlimited autonomy or superiority to another model.
SOURCES AND DOWNLOADS
The demonstration, with its scope attached.
ARC-AGI · cd3c21df ↗ARC-AGI · 009d5c81 ↗Task format and exact-match criterion ↗Development story · Medium ↗Sources, attribution and license notes ↓R4 media package · 12.4 MB ↓
The media package contains four videos, four posters, two English subtitle files, attribution notes and a checksum manifest. Private engine code, internal experiment records and the original voice reference are not distributed.
Discuss the rule and the evaluation, not just the result.
I welcome technical feedback and collaboration conversations on object representation, symbolic problem solving and learning from experience.