diff --git a/docs/ideas/README.md b/docs/ideas/README.md
new file mode 100644
index 0000000..de5bb21
--- /dev/null
+++ b/docs/ideas/README.md
@@ -0,0 +1,12 @@
+# Ideas
+
+Loose space for concepts that aren't ready to be docs or a changeset but shouldn't be lost to chat history.
+
+Conventions:
+
+- One file per idea, kebab-case slug (`guided-migration.md`, not `migration-v2.md`).
+- No obligation to ship. Most of these won't.
+- Optional frontmatter: `status: exploring | deferred | shipped | dropped`. Omit if it's too early to tell.
+- Capture *why* it's interesting, not a full spec. Open questions are fine — often the point.
+
+When an idea graduates, move it into `docs/` proper (or into a skill-bundle reference) and delete the stub here. When it's clearly dead, mark `status: dropped` and leave it — the dead branches are useful context later.
diff --git a/docs/ideas/guided-migration.md b/docs/ideas/guided-migration.md
new file mode 100644
index 0000000..1c62912
--- /dev/null
+++ b/docs/ideas/guided-migration.md
@@ -0,0 +1,44 @@
+---
+status: exploring
+---
+
+# Guided migration: drifting a fingerprint toward another
+
+An agentic loop where repo A consciously migrates toward repo B's visual direction, driven by fingerprint distance + vector as the signal.
+
+## The observation
+
+Embedding distance is a *tier selector*, not a progress bar. Closing 0.6 → 0.3 is usually easier than 0.3 → 0.05: the first half removes obviously wrong answers (different font family, different base unit, different radii language); the second half means reconciling deliberate choices that both sides consider correct. The last 0.05 is where child identity lives and often shouldn't go to zero at all.
+
+## Distance tiers
+
+| Distance | Meaning | Mode |
+|---|---|---|
+| < 0.15 | Accidental drift | **Reconcile** — token renames, no philosophy work |
+| 0.15 – 0.3 | Deliberate variance on shared foundation | **Negotiate** — diff decisions, `adopt` or `diverge` each gap |
+| 0.3 – 0.5 | Different decisions on similar kind of system | **Adopt decisions first, tokens follow** — chasing the scalar alone will mimic without migrating |
+| > 0.5 | Different design languages | **Question the premise** — often not a migration |
+
+## What the scalar hides
+
+- **Shape.** Distance 0.3 spread across four dims is uniform drift; 0.3 concentrated in palette is a single PR. Same headline, opposite plans. `computeDriftVectors` already exposes the per-dim direction.
+- **Irreducible distance.** Some gap is intentional (accessibility floor, CJK glyph support, dense-data use case). Target isn't `d = 0`; it's `d ≤ floor + ε` where `floor = Σ(diverged_dims × weight)`. Floor is computed from the manifest, not guessed.
+
+## What would be built
+
+A `migrate.md` skill recipe alongside profile / review / verify / generate / discover. Loop shape:
+
+1. `adopt B` — snapshot starting per-dim distances into `.ghost-sync.json`.
+2. Pick the steepest dim from `computeDriftVectors`.
+3. Host agent translates that delta into code edits, respecting the tier (reconcile vs negotiate vs adopt).
+4. Re-profile → `ack` → repeat.
+5. Stop when `d ≤ floor + ε`, or agent hits a dim the user marks `diverging`.
+
+No CLI primitive is missing. All the judgement work lives in the recipe.
+
+## Open questions
+
+- **Ordering within a tier.** Vector-first (fast, risks mimicry), decisions-first (correct, slow), or interleaved (each `ack` commits one dim + the decisions that justify it). Current lean: interleaved.
+- **Detecting when the scalar is lying.** Child can descend the vector gradient without importing decisions, landing at low `d` but not actually looking like B. Candidate: don't declare success until *both* the machine dims and the decisions dim are inside tolerance.
+- **Diverge budget up front.** Should `adopt` accept `--diverge <dim>,<dim>` so the floor is known before the loop runs, instead of discovered mid-migration?
+- **Symmetry.** `checkBounds` already flags `reconverging` when a diverging dim has closed to < 50% of acked distance. Guided migration is the deliberate form of that — same bookkeeping, inverted intent. Worth thinking about whether the two should share a verb.