Bilingual-First, Vault-Native

"For a market of one, your user's real workflow isn't a plugin. It's the spec."

A Workflow, Not a Feature

Dad writes in two languages and lives in a vault. In a general-purpose editor, both of those are edge cases — bilingual work is a plugin for a minority, and vault access is 'open a folder.' In Rekindle they're first-class citizens, because they're what the actual daily work is made of. That's the whole meaning of building for one person: you don't bolt the user's workflow onto the spec; the workflow is the spec.

English-Canonical, Korean-Projection

The bilingual model isn't "translate on demand." It's a doctrine with a direction: English is canonical, Korean is projection. The durable artifact is authored in English; Korean appears deliberately, in specific slots where it carries something English can't (a direct quote, a voice specimen, a proper noun). That's not a preference about languages — it's a rule that keeps one canonical layer from silently fragmenting into two half-maintained ones.

Canonical layer   English      the durable artifact
Projection        Korean       deliberate, slot-specific, carries what EN can't

# The rule exists so there's ONE canonical layer, not two half-maintained ones.
# Direction matters: projection derives from canonical, never the reverse.

The Token-Mirror Trap

Here's the failure the doctrine defends against, and it's subtle: a bilingual writer's reflexes leak. Mid-sentence, the hand reaches for a token from the other language layer — a Korean word in an English-canonical slot, or an English habit distorting a Korean line — not because it fits, but because it's the reflex the previous context primed. That's the token-mirror: the wrong layer's token, mirrored into a slot it doesn't belong in. It's nearly invisible while writing, because it feels natural — it is natural; it's just wrong for the slot.

This is exactly why the voice underline exists (Track 4), and why 'live' matters more than 'later.' A token-mirror leak caught six months later is an archaeology project across a whole corpus; caught mid-sentence, it's a two-second fix. A first-class bilingual desk isn't about translation features — it's about catching the reflex at the moment it fires.

Code

One canonical layer, deliberate projection·text

Layer         Language   Role
------------  ---------  ----------------------------------------
canonical     English    the durable artifact; authored here
projection    Korean     deliberate, slot-specific: direct quotes,
                         voice specimens, proper nouns

Direction: projection derives FROM canonical, never the reverse.

# The token-mirror leak: a token from the OTHER layer lands in a slot
# it doesn't belong in — because the previous context primed the reflex,
# not because it fits. Natural-feeling, and wrong for the slot.

Exercise

If you work in more than one language (or in more than one register — formal docs vs. casual chat), find one token you reach for out of reflex in the wrong context. Ask two questions: could you catch it on reread, or only while writing? And is it a knowledge error or a reflex error? Reflex errors are exactly the ones that need a live external catch.

Hint

The test for a reflex error: when it's pointed out you immediately recognize it's wrong — you knew all along. That's what makes rereading unreliable (your reflex reads it back as natural too) and why the catch has to happen at the moment the token lands.