An idea is not finished when it sounds good. It is finished when it has survived being argued with.

01 — the input

I say it once, and it is never edited again

A thought starts as a capture: I talk into my phone, and the words commit to version control as they are. Not tidied, not summarised. It lands in thoughts/_inbox and that file is then permanent — everything downstream is derived from it rather than replacing it.

This is the load-bearing idea, and it is why nothing is ever lost. If the reduction later turns out to be wrong, the thing it was reduced from is still sitting there.

There is exactly one way in. Not because more inputs would be worse ideas, but because two half-working inputs are worse than one working one: neither ever gets finished.

02 — the first derivation

A seed is a new thing pointing back, not a rewrite

The capture stays exactly as spoken. A seed is written from it — a title, a couple of sentences, and an address other thoughts can already link to. Under 200 words, deliberately: padding a stub toward essay length is the specific failure this whole thing exists to prevent.

03 — the ladder

Four stages, and the length is measured, not judged

A thought is a position, not a document — so the stages are seed → proposed → standing → retired: I have not worked this out → here is a write-up, put to me → I stand behind this → I no longer do.

An essay is 400–900 words, target 650. Over 1500 it is two thoughts wearing one title. The band between 200 and 400 is the awkward middle — too real to be a stub, too thin to publish — and it is the band the interview exists to push an idea out of. Ask a model “is that 400 words?” and it says yes, so nobody asks: it is counted.

standing and retired can never light up by typing, and that is the rule rather than a quirk of this demo: no model may set them. A write-up arrives as proposed and waits for me to press Accept. That is what makes standing mean “I read this” rather than “a model was happy with it”.

And retired does not mean deleted. A retired thought keeps its page, its URL and its place in the graph, marked. The URL was public; unpublishing it would rewrite the record rather than correct it.

04 — the gate

7 goal posts. 6 of them refuse

Length is necessary and nowhere near sufficient. Before a single word of the essay is written, the idea has to have cleared these — and the check happens before any of the writing starts, which is the entire point of having it.

Watch what weak does. Half a counter-argument blocks exactly like a missing one, because “sort of” is how a thin thought sneaks through.

How it connects is the one that reports without blocking, and the reason is specific: the first thought in a collection has nothing to link to, and a manufactured edge to satisfy a checklist is precisely the failure the graph is meant to avoid.

05 — the part everyone gets wrong

Two axes, not one

People expect one dial, so this is the slippery bit. Stage asks “have I read it”. Visibility asks “may anyone else”. They are independent, and a thought reaches the site only if both say yes.

The default is private, and that is the safety property rather than a preference: a file that never mentions visibility cannot reach the internet. An absent value must never mean “publish this”. The loader refuses a configuration whose default would publish, so the fail-closed behaviour cannot be switched off even deliberately.

06 — the graph

Relationships are never declared, only written

There is no “related” field and no “see also” list. An edge exists because a sentence in one thought needed to point at another, written inline as [[the exact title]]. Backlinks are derived. At most 4 links per thought, because past that the graph is a hairball and it usually means two thoughts.

A link to a title that does not exist is not an error — it becomes a hollow node, so the graph doubles as the to-write list. Which creates one genuinely nasty failure, and it is the reason this demo is here rather than a prettier one.