diff --git a/openspec/changes/generic-float-predictions-docs/.openspec.yaml b/openspec/changes/generic-float-predictions-docs/.openspec.yaml
new file mode 100644
index 0000000..454f4e0
--- /dev/null
+++ b/openspec/changes/generic-float-predictions-docs/.openspec.yaml
@@ -0,0 +1,3 @@
+schema: spec-driven
+created: 2026-09-16
+skip_specs: true
diff --git a/openspec/changes/generic-float-predictions-docs/proposal.md b/openspec/changes/generic-float-predictions-docs/proposal.md
new file mode 100644
index 0000000..8b8dd39
--- /dev/null
+++ b/openspec/changes/generic-float-predictions-docs/proposal.md
@@ -0,0 +1,19 @@
+## Why
+
+The `development-explained` book (notably ch02 common-core-contracts) documents `predict -> Array1<f64>` and `SKTargetView::Continuous(f64)` as the contract. Once the code contract becomes generic over `SKFloat`, the book contradicts the implementation and teaches the lossy `f64`-only pattern.
+
+## What Changes
+
+- Rewrite ch02 sections `SKTargetView`, `SKPredictor`, `SKFeatureTransformer` family diagram, and `SKSupervisedScorer` examples to the generic `F: SKFloat` contract with the regressor/classifier split.
+- Update every `Array1<f64>`-as-prediction snippet, Mermaid contract diagram (`predict features -> Array1 f64`), and KaTeX-free prose that claims dtype-independence of predictions.
+- Add a short "Which float should I use?" guide (`f32` vs `f64`: memory, SIMD/GPU throughput, precision) plus a migration note for the breaking change.
+- Verify with `mdbook build docs` and a rendered-page check; keep Mermaid `accTitle`/`accDescr`, KaTeX math, and English prose per the documentation guide.
+
+## Capabilities
+
+(none — documentation-only change; behavior is specified by `generic-float-predictions`.)
+
+## Impact
+
+- Affected docs: `docs/src/development-explained/ch02-common-core-contracts.md`, related tutorial/chapter snippets referencing `predict`/`SKTargetView`, Mermaid diagrams, `docs/book.toml` untouched.
+- No Rust code, no API, no dependency changes. Merged into `docs/documentations` per repo workflow, not `main`.
diff --git a/openspec/changes/generic-float-predictions-docs/tasks.md b/openspec/changes/generic-float-predictions-docs/tasks.md
new file mode 100644
index 0000000..150c5dd
--- /dev/null
+++ b/openspec/changes/generic-float-predictions-docs/tasks.md
@@ -0,0 +1,19 @@
+## 1. Book rewrite
+
+- [x] 1.1 Rewrite ch02 `SKTargetView` / `SKPredictor` / scorer sections to the generic `F: SKFloat` contract with the regressor/classifier split.
+- [x] 1.2 Update the contract family Mermaid diagram (`predict -> Array1 f64`) plus `accTitle`/`accDescr`; replace float-label examples with `i64` labels.
+- [x] 1.3 Add "Which float should I use?" guide (f32 vs f64) and breaking-change migration note, in English with KaTeX only for real formulas.
+
+Note: the `Array1<f64>` Welford accumulators, normal-equation state, and SGD
+`f64` state in the transformer/linear/streaming tutorials are internal
+accumulation precision, not prediction outputs, and were deliberately left
+untouched.
+
+## 2. Verification
+
+- [x] 2.1 Run `mdbook build docs`, open the rebuilt ch02 page, and confirm diagrams render and prose reads correctly.
+- [ ] 2.2 Confirm the change merges into `docs/documentations`, not `main`.
+
+Note on 2.1: the rebuilt ch02 page was verified in the built HTML
+(`SKRegressorPredictor` present, `accTitle` present, no `Array1 f64`
+remainder, new section present). Note on 2.2: merge happens via the doc PR.