ytk

The experiment record

Experiments on my own knowledge system. Each writeup starts plain and goes deep; the full sections live on GitHub.

Experiment 01The fogcan the knowledge map's density field be rendered, and its filaments traced?Experiment 02Pickingdoes GPU colour-ID picking do what the shader claims, and is it faster?Experiment 03Flow pulseswhat wavelength and speed should the strand pulse use?Experiment 04Ribbonsgl.LINES cannot vary width — does the quad-expansion geometry hold on paper?Experiment 05Bloomis the numpy tuning pipeline a faithful model of the bloom shader?Experiment 06Semantic domainsis the map's grouping axis a directory axis wearing semantic labels?Experiment 07Time machinewhat do note birth dates look like before a scrubber is built against them?Experiment 08The alarm that cried wolf every dayA quality alarm that fired on growth was retrained to fire only on breakage. Then sabotaged, to prove it.Experiment 09Heatmap key momentsdo Claude-generated key moments land where people actually rewatch?Experiment 10Tag coherencedo enrichment tags name real categories?Experiment 11Animationscan the two overnight experiments be explained in motion?Experiment 12Every note points the same wayEvery note pointed the same way. Most similarity was a shared lean carrying no information.Experiment 13Space in 3Da still 3D scatter is ambiguous — does the claim survive a moving camera?Experiment 14Garden allometrywhat scale and branching geometry should the garden trees use?Experiment 15Plane geometrytwo documents span a plane — is drawing it worth anything?Experiment 16Corpus primerwhat do "the mean is 11x chance", the participation ratio and z = 17 actually mean?Experiment 17Corpus growthdo the geometric claims survive 15% more corpus?Experiment 18SAE fingerprintscan the shared cone, the tags and the roads be read as named SAE features?Experiment 19Rank metricsdo rank-based metrics beat cosine on an anisotropic space?Experiment 20Query spaceswhat queries can a road between two interests express?Experiment 21Geometrydo toy-model polytopes and road intersections exist at production scale?Experiment 22Two lenseswould a second embedding space see a different person?Experiment 23Style lenscan a purpose-built style embedder be the voice axis, and does voice survive inside one medium?Experiment 24Teaching my search index to explain itselfA sparse autoencoder that names real threads in the corpus, and honestly fails the retrieval gate.Experiment 25Shared and privatecan the two lenses be decomposed into shared, topic-net-of-voice and voice-net-of-topic?Experiment 26Medium signaldoes the profile's ranking change when the medium confound is removed from the save signal?Experiment 27Alpha sensitivitydoes the inherited alpha=7 still matter once the signal is medium-corrected?Experiment 28Honest ladderthe owner's playlist is curated — what happens when that intent is finally recorded?Experiment 29Planet continentscan /orb's radial layout be de-overlapped without destroying its continents?Experiment 30My notes have a coastlineWhere does a knowledge base end? Drawing the border with a yardstick instead of an eye.Experiment 31Theme planetsdoes each theme deserve its own survey, and how do sibling planets stay visually distinct?Experiment 32The galaxydoes the sky over the theme planets need its own de-overlap pass, or is the centroid layout already legible?Experiment 33Channelswhich orbital channels — moons, rings, spin — carry real structure, when each must beat its own null before it renders?Experiment 34Individual lenscan a single real note carry a finding the record has only stated statistically? (three remixes in the Welch Labs grammar, after studying "The Dark Matter of AI")Experiment 35The knobdo section 24's exemplar-named latents actually steer retrieval when clamped — the video's intervention loop run on the production space?Experiment 36Six rulerswhat do 19's six metrics do to one concrete note — the corpus's biggest hub?Experiment 37Two families22 said the two lenses file notes differently — what does that look like for one note you can read?Experiment 38Warmstart identity#83's threshold sweep ended on a ceiling — 5-7 lifecycle events on quiet days come from the KMeans refit itself; does warm-starting the daily refit from yesterday's centroids remove them, and at what cost?Experiment 39Trace inventory#96 rung 1: before designing telemetry, what do the traces the brain already leaves actually hold?Experiment 40Does my second brain pay rentDoes saved knowledge ever come back? Yes, through a channel built by accident.Experiment 41Vector search from first principlesthe index we depend on is a black box we never built once — brute-force 10M real-geometry vectors on a 16GB MacBook, then earn the speed back stepwise, recall priced at every stepExperiment 42Atlas inventory#183 rung 0: before the activation atlas builds anything, do its inputs exist, how far do they reach over today's corpus, and which latent is the protagonist?Experiment 43The feature wall#183 rung 2: are the top-100 latents coherent objects that survive retraining — and how much of their evidence can be pictures?Experiment 44Atlas binning#183 rung 3: bin the frozen map into cells with latent identities — and how much of each identity survives retraining, is head-explained, or was never seen?Experiment 45Segment strips#183 rung 4: the flagged wall — do segment codes aggregate back to their document, and does timeline order carry signal?Experiment 46The protagonist trace#183 rung 5: the missing genre — one real note traced through every representation the system holdsExperiment 47The GEN gate#183 rung 7: can a linear Qwen-to-CLIP translator earn the right to render latents as images?Experiment 48The semantic compassrung 7 replacement, part 1: can zero-parameter contrast axes give every object — note, latent, cell, corpus — one signature glyph?Experiment 49Latent portraitsrung 7 replacement, part 2: if a latent's portrait is composited purely from its exemplars' real thumbnails, is it an identity or mush?Experiment 50Constellationsowner's conjecture: canonicalize the code image's dimension and the 32 lit latents become a shape — is the shape real?Experiment 51CLIP identity49R (#192): was section 49's portrait claim wrong, or just its pixel judge? re-run the identical gate with the perceptually aligned judgeExperiment 52Regrowing the compass48R (#193): section 48 lost the compass 3-2 on two thin poles — did those axes die of the corpus, or of their recipes?

How I decide what is real — the time I caught my own false positive.