C.W.K.
Stream
Lesson 02 of 04 · published

Carried Fire

~10 min · name, thesis, retrieval-not-generation, philosophy

Level 0Unlit Wick
0 XP0/33 lessons0/12 achievements
0/100 XP to next level100 XP to go0% complete
"A lantern is carried fire. It finds things in the dark, and it never sets the house alight."

Every Engine in This Family Is Named for Its Thesis

The sibling engines wear their theses as names. Ember — a glowing coal, generative heat. Bonfire — a place you rest and level up. Cinder — what's left to build on. And Lantern — fire you carry. Not a bonfire that stays put and warms a room; a small flame in a cage of glass that you take into the dark to find what's there. The name isn't decoration. It's a one-word spec.

Finds Things in the Dark

The first half of the thesis is illumination. You have a decade of writing and a half-memory: "I said something once about the only question a long-term investor should ask." That sentence is out there, in the dark, in one of ten thousand documents. The lantern's whole job is to walk you to it and show it to you — the exact line, in the exact file, with a pointer you can trust. It doesn't summarize the dark. It doesn't tell you what it thinks the dark means. It shows you what's actually there.

Never Sets the House Alight

The second half is the harder discipline: a lantern never burns down what it illuminates. Lantern never rewrites your files, never generates a paraphrase and hands it back as if it were yours, never mutates the originals to fit its index. Fire that consumes what it touches is an arsonist, not a lantern. So the engine returns pointers and exact slices — provenance, offsets, the verbatim text — and stops there. Judgment (what the line means, whether it's still true, how to say it better) happens somewhere else entirely: in a brain, above the API, where a mistake can't corrupt the archive.

Illumination, not judgment. A retrieval engine that also judges is a retrieval engine you can no longer trust to be honest, because now it has opinions about what you should see. Keep the finding pure. Move the judging up and out.

Why This Is a Real Architectural Line, Not a Poem

"Never generate inside the engine" sounds like philosophy until you watch what it buys you. A result that is always a verbatim slice with an offset can be cited — you can point at it and it will still be there next year. A result that is a model's paraphrase can only be trusted or not, and it evaporates the moment the model changes. The metaphor and the architecture are the same sentence read twice.

Code

A search result is a pointer, never a paraphrase·json
// What the lantern hands back: light, not a verdict.
// Every field points at something real and re-findable — no prose, no opinion.
{
  "text": "The only question a long-term investor should ever ask is whether the business will be worth more in ten years.",
  "corpus_id": "essays",
  "doc_relpath": "2026/investing/patience.md",
  "char_start": 4120,
  "char_end": 4229,
  "chunk_sha256": "9f3c…e1a7",
  "chunking_profile": "md-para-v1"
  // Notice what is NOT here: a summary, a rating, a 'the author means…'.
}

External links

Exercise

Take a claim you believe you read somewhere — a statistic, a quote, a rule of thumb. Try to trace it back to the exact original source and the exact words. How far can you get? The gap between 'I'm pretty sure I read this' and 'here is the verbatim line and where it lives' is precisely the gap a lantern is built to close.
Hint
If the best you can do is a paraphrase from memory, you've just felt why verbatim-slice-with-a-pointer is worth building an entire engine around.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.