diff --git a/CHANGELOG.md b/CHANGELOG.md
index 85093f6..8b9bddd 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [0.5.0] — 2026-08-07
+
+Scoring became CV-aware. A job's match score is now owned by the CV that produced
+it, so changing your CV re-opens every cached job instead of silently reusing a
+verdict reached for someone else's profile.
+
 ### New Features
 
 - **`jobradar cache prune-scores`** (`cli.py` / `cache.py`)
@@ -30,6 +36,20 @@
 - **Release highlights in Chinese, English and Spanish** (`docs/release-highlights.*.md` / `README*.md`)
   The highlights document now follows the same `<name>.<lang>.md` convention as the READMEs, with a language switcher in each file. Each README gained a Performance section carrying the latest release's headline figures and linking to the highlights and to the `### Validation` entries behind them.
 
+- **Release-highlights inclusion rule restated** (`docs/release-highlights.*.md`)
+  The rule read "from 0.5.0 onward each release is compared against the previous release", which said nothing about releases with nothing to compare. It now states that a release is listed when it changes performance or user-facing behavior, that the kind of change determines what is compared — measured figures for performance work, changed behavior for behavior work — and that a release altering neither is not listed.
+
+### Validation
+
+- **CV-aware scoring** — `f14e369` (0.4.0) → `b47dfc0`
+  - Held constant: the same local cache (293 jobs), the same CV (`eaf1b010`), `gemini-3.5-flash-lite`
+  - Jobs `assess` could reach: 0 → 205. Before, the command reported "all assessed" because every row carried a legacy assessment
+  - Cached jobs a CV change could never re-open: 95 → 0
+  - Jobs holding a match for the current CV: 93 → 293, by backfilling 200 (0 failures, 173 s, 5 workers)
+  - Scales feeding `effective_score`: 2 (legacy 0–10 alongside match 0–100) → 1
+  - Cache-hit classification, measured on five jobs a previous CV had rejected: passing the current `cv_hash` yields 0 skipped / 5 re-assessed, against 5 skipped / 0 re-assessed on the legacy path
+  - Verdict: correctness only. **No performance or cost improvement is claimed** — the release adds assessment work rather than removing it, since a cached job without a match for the current CV is now re-evaluated instead of reused. Expect a one-off cost increase on the first search after changing CV, and no change thereafter. Search latency was not measured and is not claimed either way.
+
 ## [0.4.0] — 2026-08-07
 
 ### New Features
diff --git a/README.es.md b/README.es.md
index 668845c..f18794a 100644
--- a/README.es.md
+++ b/README.es.md
@@ -79,7 +79,8 @@ Indeed 741 + LinkedIn 255 = 996 extraídos
 
 ## Rendimiento
 
-Última versión, 0.4.0:
+El trabajo de rendimiento más reciente es de 0.4.0. La versión actual, 0.5.0,
+modificó el comportamiento y no el rendimiento, y no reclama cifras propias:
 
 | Métrica | Antes | Después |
 | --- | --- | --- |
diff --git a/README.md b/README.md
index a987d75..32dff85 100644
--- a/README.md
+++ b/README.md
@@ -81,7 +81,8 @@ Indeed 741 + LinkedIn 255 = 996 scraped
 
 ## Performance
 
-Latest release, 0.4.0:
+Most recent performance work, 0.4.0. The current release, 0.5.0, changed
+behavior rather than speed and claims no performance figures of its own:
 
 | Metric | Before | After |
 | --- | --- | --- |
diff --git a/README.zh.md b/README.zh.md
index 1a5e52e..8ec2f39 100644
--- a/README.zh.md
+++ b/README.zh.md
@@ -79,7 +79,7 @@ Indeed 741 + LinkedIn 255 = 996 抓取
 
 ## 性能
 
-最新版本 0.4.0：
+最近一次性能改动来自 0.4.0。当前版本 0.5.0 改的是行为而非速度，不声称任何性能数字：
 
 | 指标 | 优化前 | 优化后 |
 | --- | --- | --- |
diff --git a/data/jobradar_cache.db.bak b/data/jobradar_cache.db.bak
new file mode 100644
index 0000000..7773f7c
Binary files /dev/null and b/data/jobradar_cache.db.bak differ
diff --git a/docs/release-highlights.es.md b/docs/release-highlights.es.md
index 243315c..a4af6cb 100644
--- a/docs/release-highlights.es.md
+++ b/docs/release-highlights.es.md
@@ -6,7 +6,41 @@ Resumen orientado al usuario de lo que mejoró cada versión. Cada cifra debe po
 rastrearse hasta una entrada `### Validation` en [CHANGELOG.md](../CHANGELOG.md);
 este archivo nunca introduce una cifra propia.
 
-A partir de 0.5.0, cada versión se compara con la versión anterior.
+A partir de 0.5.0, una versión aparece aquí cuando modifica el rendimiento o el
+comportamiento visible para el usuario, y se describe en comparación con la versión
+anterior. Qué se compara depende del tipo de cambio: el trabajo de rendimiento se
+compara con cifras medidas; el de comportamiento, con lo que la herramienta hace
+ahora de forma distinta. Una versión que no altere ninguno de los dos no se lista.
+
+## 0.5.0
+
+Una versión de comportamiento, no de rendimiento. La puntuación pasó a depender del
+CV, lo que cambia qué devuelve una búsqueda y cuánto cuesta. Las cifras proceden de
+la entrada `### Validation` de `f14e369` → `b47dfc0`.
+
+**Cambiar de CV vuelve a abrir todos los empleos en caché**
+Hasta 0.4.0 la puntuación se guardaba sin registrar qué CV la produjo, así que un
+empleo juzgado con un CV antiguo seguía juzgado. En una caché de 293 empleos, 95 que
+un CV anterior había rechazado no podían reconsiderarse nunca; ahora son 0.
+
+**`assess` dejó de informar de un trabajo que no hacía**
+Filtraba por la columna heredada antes de consultar los resultados de coincidencia y,
+como cada fila llevaba una puntuación heredada, anunciaba "todo evaluado" mientras 205
+empleos carecían de puntuación para el CV actual. Empleos alcanzables: 0 → 205.
+
+**Una sola escala de puntuación en lugar de dos**
+Una puntuación heredada de 0–10 y una de coincidencia de 0–100 se devolvían por la
+misma propiedad y se ordenaban entre sí. Escalas en `effective_score`: 2 → 1.
+
+**Nuevo: `jobradar cache prune-scores`**
+Elimina resultados de coincidencia de una versión de prompt obsoleta y, opcionalmente,
+de un CV obsoleto, para que la siguiente búsqueda los recalcule.
+
+**Nota sobre el coste**
+Esta versión añade trabajo de evaluación en lugar de quitarlo. La primera búsqueda tras
+cambiar de CV reevalúa los empleos en caché en vez de reutilizar el veredicto de otro CV;
+las búsquedas posteriores no se ven afectadas. No se afirma ninguna mejora de latencia ni
+de coste.
 
 ## 0.4.0
 
diff --git a/docs/release-highlights.md b/docs/release-highlights.md
index 3a55fd5..6f23dcc 100644
--- a/docs/release-highlights.md
+++ b/docs/release-highlights.md
@@ -6,7 +6,40 @@ User-facing summary of what each release improved. Every figure here must trace
 back to a `### Validation` entry in [CHANGELOG.md](../CHANGELOG.md); this file
 never introduces a number of its own.
 
-From 0.5.0 onward each release is compared against the previous release.
+From 0.5.0 onward, a release appears here when it changes performance or
+user-facing behavior, and is described against the previous release. Which
+kind of change it is determines what gets compared: performance work is
+compared on measured figures, behavior work on what the tool now does
+differently. A release that changes neither is not listed.
+
+## 0.5.0
+
+A behavior release, not a performance one. Scoring became CV-aware, which changes
+what a search returns and what it costs. Figures below come from the
+`### Validation` entry for `f14e369` → `b47dfc0`.
+
+**Changing your CV now re-opens every cached job**
+Until 0.4.0 a score was stored without recording which CV produced it, so a job
+judged under an old CV stayed judged. On a 293-job cache, 95 jobs an earlier CV
+had rejected could never be reconsidered — now 0.
+
+**`assess` stopped reporting work it wasn't doing**
+It selected on the legacy column before consulting match results, so with every
+row carrying a legacy score it announced "all assessed" while 205 jobs had no
+score for the current CV. Reachable jobs: 0 → 205.
+
+**One scoring scale instead of two**
+A legacy 0–10 score and a 0–100 match score were both returned by the same
+property and sorted against each other. Scales feeding `effective_score`: 2 → 1.
+
+**New: `jobradar cache prune-scores`**
+Drops match results from an outdated prompt version, and optionally from an
+outdated CV, so the next search recomputes them.
+
+**Cost note**
+This release adds assessment work rather than removing it. The first search after
+changing CV re-evaluates cached jobs instead of reusing another CV's verdict;
+later searches are unaffected. No latency or cost improvement is claimed.
 
 ## 0.4.0
 
diff --git a/docs/release-highlights.zh.md b/docs/release-highlights.zh.md
index f6fc2d7..d243978 100644
--- a/docs/release-highlights.zh.md
+++ b/docs/release-highlights.zh.md
@@ -5,7 +5,33 @@
 面向使用者的版本优化摘要。这里的每个数字都必须能在 [CHANGELOG.md](../CHANGELOG.md)
 的 `### Validation` 中找到出处；本文档不产生任何自己的数字。
 
-从 0.5.0 起，每个版本对比上一个发布版本。
+从 0.5.0 起，只要版本改变了性能或使用者能感知的行为，就在这里与上一个发布版本
+对比说明。对比什么取决于改动的性质：性能改动比实测数字，行为改动比工具的实际
+表现差异。两者都没有改变的版本不会出现在本文档中。
+
+## 0.5.0
+
+这是行为变更版本，不是性能版本。评分开始区分 CV 归属，因此搜索返回什么、花多少钱
+都随之改变。下列数字出自 `f14e369` → `b47dfc0` 的 `### Validation` 条目。
+
+**换 CV 后所有缓存职位重新参与评估**
+0.4.0 之前，评分落库时不记录是哪份 CV 算的，于是被旧 CV 判定过的职位就一直维持
+那个判定。293 条缓存中，有 95 条被旧 CV 拒绝的职位永远无法被重新考虑——现在为 0。
+
+**`assess` 不再谎报自己已完成**
+它先按 legacy 列筛选再查匹配结果，而每条记录都带着 legacy 评分，于是它宣称
+"所有 JD 已有评估"，实际有 205 条在当前 CV 下毫无评分。可捞出的职位：0 → 205。
+
+**评分只剩一种分制**
+legacy 的 0~10 分与 match 的 0~100 分曾由同一个属性返回，并在同一次排序里比大小。
+参与 `effective_score` 的分制：2 → 1。
+
+**新增 `jobradar cache prune-scores`**
+删除过时 prompt 版本的匹配结果，也可一并删除旧 CV 的，让下次搜索按当前口径重算。
+
+**成本说明**
+本版本增加而非减少评估量。换 CV 后的第一次搜索会重新评估缓存职位，而不是复用另一份
+CV 的判定；之后的搜索不受影响。本版本不声称任何延迟或成本改进。
 
 ## 0.4.0
 
diff --git a/pyproject.toml b/pyproject.toml
index de455b0..8224987 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "jobradar"
-version = "0.4.0"
+version = "0.5.0"
 description = "Radar-scan the job market against your CV — auto-discover, scrape and LLM-assess matching roles"
 license = {text = "BUSL-1.1"}
 requires-python = ">=3.11"
diff --git a/uv.lock b/uv.lock
index 6d4101e..f9a7819 100644
--- a/uv.lock
+++ b/uv.lock
@@ -581,7 +581,7 @@ wheels = [
 
 [[package]]
 name = "jobradar"
-version = "0.4.0"
+version = "0.5.0"
 source = { editable = "." }
 dependencies = [
     { name = "anthropic" },