Refract
KO/EN

Method — how this is made

How a piece gets made and verified at Refract, written out in the open. In an AI era the method gets caught up to anyway — so we publish it. What compounds is judgment and the accumulated corpus.

Planplan-led
Researchsourced ledger
Draftdraft
Review · 8 lenses in parallel
argumentcraftdimensionsinsightAI-tellsadversarypersonasflow
Fact-checkhard gate
⟳ revise ≤2
Validatemachine check
Bilingual EN ⇄ KOpublish

1. The angle

Refract takes a single event and reads it across several dimensions — economics, policy, technology, psychology — rather than reporting one thread. The angle shifts with the topic; the constant is that the surface is never the whole story.

2. A data foundation — DATAWALLET

Refract isn’t opinion stacked on opinion; it sits on a personal data platform called DATAWALLET (a local data lakehouse — a “personal NotebookLM”). Market and macro time-series (indices, FX, rates, prices), papers, and the published essays themselves accumulate over time across three domains — economy, tech, humanities — read at once through three lenses: table, meaning, relationship. Writing grows on top of that data, and once published it feeds back into the corpus (a flywheel). That’s why “the piece never outruns the data” is an engine, not a slogan.

Sources · 3 domains
economytechhumanities
Lake · medallionraw ▶ staging ▶ mart · files = SSOT
Engines · 3 lenses on the lake
table · SQLmeaning · vectorsrelations · graph
Serving boundaryconsume mart / views only
Use · topic workspacetable · meaning · relations · narrative
⟳ nightly flywheel — verified, then accrues (compounding)

3. First principle — the piece never outruns the data

Every load-bearing claim (figures, assertions, quotes) carries a source and is matched 1:1 against a fact ledger. Unsourced, overstated, or data-contradicting → it doesn’t ship. Analysis is drawn only within what the data supports — no over-claiming.

4. The pipeline

Veteran-level planning first → targeted research (a sourced ledger) → drafting → an 8-lens review team in parallel (argument, craft, multi-dimensionality, non-obvious insight, AI-tells, devil’s advocate, multi-persona, lay-reader flow) plus a domain expert when needed → fact-check (hard gate) → machine validation → a bilingual English relocalization (numbers and sources 1:1). Facts are non-negotiable (gates); polish is best-effort within a fixed budget.

5. What the AI does, what the human does

The AI reads across dimensions, drafts, and reviews itself. The human reviews, decides, and publishes. AI makes mistakes. Forecasts can miss — you are the protagonist who decides.

6. Open by method, private by code

This page is public; the operating repository is private. In an AI era the method gets caught up to anyway — so we don’t hide it. What compounds is judgment and the accumulated corpus, not the machinery.

7. Still open — a work in progress

This method isn’t finished. The review team and the pipeline keep getting rewritten. Refract is not a closed system but an experiment still in progress, revised toward sharper analysis and more honest writing. If you see a gap, please tell us — we sharpen it together.