IDEAS
The intake stage. Every experiment starts as one of these lines; some graduate to the lab, some are patiently waiting, and this list is allowed to be a wishlist — that's the whole point.
IDEAS → BUILDING → SHIPPED | GRAVEYARD
Claims-drift detector
AUG 2026The claims that rot are the ones with two homes — a number in code and the same number in prose. A checker that finds every published figure, traces it to the source that makes it true, and fails when they disagree would kill the whole class. The hard part is not the diff, it is that a comment citing a rule which was later rewritten is also a drifted claim, and nothing greps for that.
A test that catches a stale premise
AUG 2026The worst failures this year were not bugs — they were work built on assumptions that had quietly stopped being true upstream. All of them compiled and passed every gate. If a build could assert that the constraints a change cites still exist, that entire category becomes catchable instead of a reading exercise.
A next-app template that deploys itself
AUG 2026experiment:new already gets an idea to running code in one command. The gap is the last mile — a template pre-wired for its own subdomain would make the distance from idea to public URL short enough that more experiments become real.
Local inference that fits in 16GB
JUL 2026Every local-model writeup assumes a 64GB machine. On 16GB the interesting question is which jobs a 7B or 8B model does well enough to be worth not sending out — commit messages, changelog drafts, classification — and whether the answer is honestly none.
n8n as the boring half of an agent
JUN 2026Most of what an agent does on a schedule is plumbing that does not need a model. Running the deterministic parts in n8n and calling a model only where judgment is required should be cheaper and far easier to debug than one agent that does everything.
Soulseek to Plexamp, without the middle step
MAY 2026The tedious part of a self-hosted library is not acquiring music, it is tagging and filing it. A pipeline that lands a download already tagged and in the right place would remove the only reason the library goes stale.
Ontology-backed answers over pattern-matched ones
JUN 2026An agent asked about a Workday configuration currently pattern-matches documentation and produces plausible answers that are wrong in the specifics. Given a typed graph instead, its failure mode should shift from confident invention to an honest gap — and that is a measurable difference, not a stylistic one.
became Ontology
an idea here is a promise to nobody. the graveyard keeps this list honest.