Coilendar

What it is. A time curve takes a timeline and folds it by similarity: each moment becomes a point, similar moments are pulled close together, and a curve threads the points in chronological order (Bach, Shi, Heulot, Madhyastha, Grabowski & Dragicevic, "Time Curves: Folding Time to Visualize Patterns of Temporal Evolution in Data"). Instead of asking "when did things happen," it shows whether time kept returning to the same states.

It's a 2D projection: some distances are necessarily distorted, and with only a handful of artifacts the geometry is suggestive, not statistical. It carries no quantitative time axis — that's the CONTENT/ACCESS views' job.

The question. Not "what did the filed answers cite" (that's the histogram below), but what did the runs actually reach — the shape of the archive's traversal. A visit = a leaf appears as a hop in a mounted traversal. Each machine is measured in isolation; the contrast between a machine with a compiled wiki and one without separates pull the wiki induces from pull the raw archive (and the model's priors) produce on their own.

Over-pull (type-1 analogue) = leaves many runs converge on: either a genuine hub or an induced attractor dragging attention off-question. Blind spot (type-2 analogue) = reachable ground no run touched.

These are candidates for keeper attention, not verdicts. We hold no thematic ground truth — no oracle says which leaves a question should have reached — so a never-reached leaf is not a proven false-negative, and a hub is not a proven false-positive. Adjacency is not convergence. The instruments locate where to look; the judgement stays a keeper act.

What this is an implementation of. This archive is a welfare-adapted implementation of the LLM-wiki architecture (Karpathy): instead of retrieving raw chunks per query, an LLM incrementally builds and maintains a persistent, interlinked wiki between the reader and the sources. The paradigm defines three layers: raw (immutable sources — here, the tapes and their verbatim transcripts, app/raw/leaves/), wiki (the LLM-maintained compiled layer), and schema (the conventions binding it, app/schema/schema.md).

The one structural fork this project makes: the paradigm files answers back into the wiki itself; here the wiki layer is split in two. Answers (app/answers/) hold each walk's filed conclusions; wiki structure (app/wiki/) holds only earned claims + verbatim anchors + pointers to answers — never their text. The split is deliberate: future walks can use the structure without inhaling prior conclusions, and evaluation surfaces become plain directory boundaries.

This visualization draws the bookkeeping of those layers: it reads the wiki structure, the answers, and the keeper's log — and shows how everything above raw roots back into it.

First principles: this wiki cannot plan itself — structure exists only because some walk filed something back, and independent walks converging is what earns a node its existence. This view is that ledger laid out in access-time: what was filed when, and which filings earned each node.

First principles: every vertical layer is one more step of abstraction away from the verbatim tapes — reading upward is reading away from ground — and nothing above ground is allowed to float free: every claim must visibly root back into a leaf at an exact timecode.

one line is all of time, coiled — a 2D plane. every calendar period is a circle riding the next bigger circle.