diff --git a/docs/architecture/decisions/051-yaml-input-for-from-openapi.md b/docs/architecture/decisions/051-yaml-input-for-from-openapi.md
index a1e313f..c23e965 100644
--- a/docs/architecture/decisions/051-yaml-input-for-from-openapi.md
+++ b/docs/architecture/decisions/051-yaml-input-for-from-openapi.md
@@ -133,6 +133,79 @@ default-features model avoids. The dependency is small (a pure-Rust YAML
 parser, no native code), consistent with the existing default-features
 philosophy of the crate.
 
+### 5. YAML/JSON parity contract (review 002 OAI-12)
+
+> **Amendment (2026-08-30, review 002 OAI-12):** the original delivery fed
+> `yaml_serde::from_str::<serde_json::Value>` straight into `from_value`
+> with zero post-parse normalization, and three verified corruptions
+> flowed through unimpeded: duplicate keys silently last-won; `.inf`/
+> `.nan` scalars silently became `null` (`serde_json::Number::from_f64`
+> is `None` for non-finite floats); merge keys (`<<: *anchor`) survived
+> as literal `<<` properties. `from_yaml` now runs a normalization
+> pipeline; this section records the resulting contract. The same
+> document may still mean different things through `from_json` and
+> `from_yaml` only where explicitly stated below — every difference is
+> loud on the YAML side.
+
+The `from_yaml` pipeline is: an explicit parse into `yaml_serde::Value`
+→ [`apply_merge()`](https://docs.rs/yaml_serde) → one structural
+normalization pass into `serde_json::Value` → the shared `from_value`
+path. Per-construct rules:
+
+1. **Duplicate mapping keys are rejected loudly on the YAML path** — the
+   explicit `yaml_serde::Value` parse rejects them natively with the key
+   and its line/column, naming the document position. The JSON path
+   *silently last-wins* (`serde_json::Value`'s `visit_map` inserts into
+   a map; verified empirically against serde_json 1.0.151 — the
+   oft-assumed "JSON errors on duplicates" behavior does not hold for
+   the `Value` target). The YAML path is deliberately the stricter one:
+   a duplicate-key document fails YAML import instead of silently
+   meaning different things through the two entry points. Callers who
+   need JSON-path tolerance of duplicates have it by construction; the
+   asymmetry is documented and tested, not hidden.
+2. **Non-finite floats are rejected loudly.** YAML 1.2 core schema has
+   `.inf`/`-.inf`/`.INF`/`.nan` (and case variants) as floats;
+   `serde_json::Number` cannot represent them, and an unguarded
+   conversion would advertise `null` where the document declares e.g.
+   `maximum: .inf` — silently dropping a declared constraint from the
+   schema. The normalization pass fails import with the offending
+   value's JSON pointer and the rendered value.
+3. **Merge keys are applied, not advertised.** `<<: *anchor` resolves
+   per the YAML merge type before conversion; `<<` never survives as a
+   literal property in the parsed spec. The semantics are `yaml_serde`'s
+   `apply_merge()`: shallow `entry().or_insert()` into the referencing
+   mapping — the referencing mapping's explicit keys win, and a `<<`
+   value that is neither a mapping nor a sequence of mappings fails
+   loudly. This is the one deliberate, tested deviation from strict
+   YAML 1.2 core-schema processing (§2): merge keys are a YAML 1.1
+   mechanism that YAML tooling and authors universally expect to work,
+   and the alternative (rejecting `<<` loudly) would refuse
+   anchor-heavy real-world documents that JSON cannot even express. A
+   producer wanting strict 1.2 processing can avoid `<<` in favor of
+   explicit repetition.
+4. **Non-string mapping keys** are stringified on both paths by
+   construction (JSON keys are strings; YAML scalar keys are rendered
+   exactly as the YAML 1.2 core schema renders the scalar: `200:` →
+   `"200"` matches JSON `{"200": ...}`, `true:` → `"true"`, `1.5:` →
+   `"1.5"`). Keys with no round-trip-stable string form — nullish keys
+   (`~`, empty) and collection keys (sequences, mappings) — are
+   rejected loudly with the key's JSON pointer instead of being
+   stringified through YAML's debug rendering.
+5. **Scalar interpretation is identical by dependency version.** Both
+   paths go through the YAML 1.2 core schema or JSON's stricter grammar
+   (§2 as amended): bare `yes`/`no`/`on`/`off` are strings, `!!str 200`
+   is the string `"200"`, and unknown tags fail loudly on both paths.
+   The JSON-first rule of §2 remains the defensive lock against a
+   future YAML-parser swap changing this.
+
+The normalization pass adds no new resource bounds: it walks structure
+already accepted by `yaml_serde`'s parse-time limits (recursion limit,
+alias jump limit, repetition limit — re-verified during review 002), is
+linear in document size, and allocates only the converted document
+plus error messages. Fixtures for the normalization are all linear or
+bounded; the `$ref` resolver's budgets (OAI-01/OAI-11) are downstream
+and unchanged.
+
 ### 4. Scope boundary: `to_openapi` output is not affected
 
 `to_openapi` generates the published gateway doc, served at `GET