diff --git a/docs/decisions/GDR-003-triad-tetrad-difficulty-reorder.md b/docs/decisions/GDR-003-triad-tetrad-difficulty-reorder.md
new file mode 100644
index 0000000..93e3118
--- /dev/null
+++ b/docs/decisions/GDR-003-triad-tetrad-difficulty-reorder.md
@@ -0,0 +1,105 @@
+# GDR-003: Triad/Tetrad Difficulty Reorder
+
+**Status**: Proposed
+**Date**: 2026-07-17
+**Area**: Pedagogy | Content
+
+## Context
+
+Playtesting the full Stage 1 (Triads) and Stage 2 (Tetrads) sequence surfaced
+a felt difficulty regression: after `aug`/`dim` (Stage 1) and after the
+`m7`-family (Stage 2), the curriculum drops back to something noticeably
+easier — `sus2`/`sus4` in Stage 1, `maj6`/`dom7sus4` in Stage 2 — instead of
+continuing to ramp up.
+
+Post-hoc analysis of why the drop feels wrong:
+
+- `sus2`/`sus4` differ from `major`/`minor` by exactly **one replaced tone**
+  (2M or 4P swapped for the 3rd). `maj6`/`dom7sus4` are the same kind of
+  one-note delta from `maj7`/`dom7` (6M replaces 7M; 4P replaces 3M). Their
+  actual difficulty is much closer to the base major/minor/dominant chords
+  than to what currently precedes them.
+- `aug`/`dim` alter the 5th — already a step beyond the plain triad — and
+  `m7b5`/`dim7`/`mMaj7` stack two structural changes at once (an altered 5th
+  *and* a specific 7th). These are legitimately harder than the sus/6/7sus4
+  group.
+
+Current order (`CURRICULUM` in `src/config/game.ts:23-24`) teaches the harder
+qualities first, then regresses to the simpler one-note-delta group. This
+contradicts the "isolate one new variable at a time" principle already
+established in **GDR-002** (Foundational Pedagogy Stages).
+
+## Options Considered
+
+### Option A — Keep current order
+
+- Triads: `major, minor, aug, dim, sus2, sus4`
+- Tetrads: `maj7, dom7, m7, mMaj7, m7b5, dim7, maj6, dom7sus4`
+
+- Pro: no change required, already implemented and tested.
+- Con: confirmed by playtesting to produce a felt difficulty regression —
+  sus/6/7sus4 chords land after harder material despite being simpler.
+
+### Option B — Reorder to isolate one-note deltas immediately after their base quality
+
+- Triads: `major, minor, sus2, sus4, aug, dim`
+- Tetrads: `maj7, maj6, dom7, dom7sus4, m7, mMaj7, m7b5, dim7`
+
+- Pro: each new quality differs from an already-mastered one by exactly one
+  altered/replaced tone before moving to a harder structural change (altered
+  5th, then altered 5th + specific 7th); removes the observed regression;
+  consistent with GDR-002's "isolate one variable" principle.
+- Con: pushes `aug`/`dim`/`m7b5`/`dim7` later in the sequence; breaks the
+  implicit "chord-family grouping" of the current order (e.g., all
+  diminished-family qualities together) in favor of grouping by structural
+  distance from a base chord.
+
+## Decision
+
+**Chosen: Option B — reorder by structural distance from the base chord.**
+
+- Triads: `major, minor, sus2, sus4, aug, dim`
+- Tetrads: `maj7, maj6, dom7, dom7sus4, m7, mMaj7, m7b5, dim7`
+
+This removes the playtesting-observed regression and keeps the difficulty
+ramp monotonic: one-note deltas from an already-known chord come before
+qualities that alter the 5th, which in turn come before qualities that
+combine an altered 5th with a specific 7th.
+
+## Consequences
+
+**Trade-offs accepted:**
+
+- `aug`, `dim`, `m7b5`, and `dim7` are introduced later than today. Players
+  reach the full diminished/augmented vocabulary a few substages later than
+  before — acceptable since the goal is a smoother curve, not a faster one.
+- The "chord-family" grouping in the current order (e.g., keeping all
+  half-/fully-diminished tetrads adjacent) is given up in favor of grouping
+  by structural distance from the base chord.
+
+**Files affected:**
+
+- `src/config/game.ts` — `CURRICULUM[1]` (Triads) and `CURRICULUM[2]`
+  (Tetrads) reordered to the arrays above.
+- `docs/GDD.md` §5 — Etapa 1 (Triadas) and Etapa 2 (Tétradas) progression
+  lists reordered to match.
+- `src/stores/__tests__/progress.spec.ts` — asserts on `stage`/`subStage`
+  **indices**, not quality names; expected unaffected, but should be
+  re-verified once the reorder lands.
+
+**Signals to watch:**
+
+- Substage completion rate and error rate for `sus2`/`sus4` (Stage 1) and
+  `maj6`/`dom7sus4` (Stage 2) before vs. after the reorder — the local
+  spike-then-relief pattern in error rate should disappear from the curve.
+
+## Open Questions
+
+- [ ] None blocking. Recommend a follow-up empirical check once enough
+  playtesting data exists under the new order, to confirm the regression is
+  actually gone rather than shifted elsewhere in the sequence.
+
+## Related
+
+- GDR-002: Foundational Pedagogy Stages — shares the "isolate one new
+  variable at a time" principle this reorder is built on.