diff --git a/app/jobs/ai_log_job.py b/app/jobs/ai_log_job.py
index 2c0277e..9b5683e 100644
--- a/app/jobs/ai_log_job.py
+++ b/app/jobs/ai_log_job.py
@@ -40,9 +40,9 @@ async def translate_log_for_active_languages(session, log_id: int) -> None:
     Reads ``users.lang`` (deduplicated, restricted to ACTIVE_LANGUAGES
     minus English), one translation call per language in parallel via
     ``asyncio.gather``, persists each successful result as a
-    ``StrategicLogTranslation`` row. Per-language failures are logged
-    but never raise — the strategic log itself is already committed at
-    this point and translation is a best-effort enhancement.
+    ``StrategicLogTranslation`` row. Each row is committed in its own
+    savepoint so a per-language LLM error or DB error doesn't roll back
+    the languages that already succeeded.
 
     The job orchestrator calls this AFTER the English ``StrategicLog``
     row is committed; pass the row's ``id`` in.
@@ -68,22 +68,39 @@ async def translate_log_for_active_languages(session, log_id: int) -> None:
             for lang in active_langs
         ], return_exceptions=True)
 
+    succeeded = 0
+    failed = 0
     for lang, result in zip(active_langs, results):
         if isinstance(result, Exception):
             log.warning("log.translate.failed", lang=lang, log_id=log_id,
                         error=str(result)[:200])
+            failed += 1
             continue
         translated_md, llm_result = result
-        session.add(StrategicLogTranslation(
-            log_id=log_id, lang=lang,
-            content=translated_md,
-            generated_at=utcnow(),
-            model=llm_result.model,
-            prompt_tokens=llm_result.prompt_tokens,
-            completion_tokens=llm_result.completion_tokens,
-            cost_usd=llm_result.cost_usd,
-        ))
-    await session.commit()
+        try:
+            async with session.begin_nested():
+                session.add(StrategicLogTranslation(
+                    log_id=log_id, lang=lang,
+                    content=translated_md,
+                    generated_at=utcnow(),
+                    model=llm_result.model,
+                    prompt_tokens=llm_result.prompt_tokens,
+                    completion_tokens=llm_result.completion_tokens,
+                    cost_usd=llm_result.cost_usd,
+                ))
+            await session.commit()
+            succeeded += 1
+        except Exception as exc:
+            log.warning("log.translate.persist_failed",
+                        lang=lang, log_id=log_id, error=str(exc)[:200])
+            failed += 1
+
+    if failed and succeeded == 0:
+        log.error("log.translate.all_failed",
+                  log_id=log_id, attempted=len(active_langs))
+    else:
+        log.info("log.translate.done",
+                 log_id=log_id, succeeded=succeeded, failed=failed)
 
 
 async def run() -> None:
diff --git a/app/jobs/indicator_summary_job.py b/app/jobs/indicator_summary_job.py
index fb21f24..97c5f80 100644
--- a/app/jobs/indicator_summary_job.py
+++ b/app/jobs/indicator_summary_job.py
@@ -47,8 +47,8 @@ async def translate_summary_for_active_languages(session, summary_id: int) -> No
     Mirrors ``ai_log_job.translate_log_for_active_languages``: reads the
     distinct non-en ``users.lang`` set, translates the English content
     once per active language in parallel via ``asyncio.gather``, and
-    persists each result as an ``IndicatorSummaryTranslation`` row.
-    Per-language failures are logged but never raise.
+    persists each result as an ``IndicatorSummaryTranslation`` row in
+    its own savepoint so one bad row doesn't lose the rest.
     """
     target_langs = sorted({l for l in ACTIVE_LANGUAGES if l != "en"})
     if not target_langs:
@@ -70,23 +70,40 @@ async def translate_summary_for_active_languages(session, summary_id: int) -> No
             for lang in active_langs
         ], return_exceptions=True)
 
+    succeeded = 0
+    failed = 0
     for lang, result in zip(active_langs, results):
         if isinstance(result, Exception):
             log.warning("ind_summary.translate.failed",
                         lang=lang, summary_id=summary_id,
                         error=str(result)[:200])
+            failed += 1
             continue
         translated_md, llm_result = result
-        session.add(IndicatorSummaryTranslation(
-            summary_id=summary_id, lang=lang,
-            content=translated_md,
-            generated_at=utcnow(),
-            model=llm_result.model,
-            prompt_tokens=llm_result.prompt_tokens,
-            completion_tokens=llm_result.completion_tokens,
-            cost_usd=llm_result.cost_usd,
-        ))
-    await session.commit()
+        try:
+            async with session.begin_nested():
+                session.add(IndicatorSummaryTranslation(
+                    summary_id=summary_id, lang=lang,
+                    content=translated_md,
+                    generated_at=utcnow(),
+                    model=llm_result.model,
+                    prompt_tokens=llm_result.prompt_tokens,
+                    completion_tokens=llm_result.completion_tokens,
+                    cost_usd=llm_result.cost_usd,
+                ))
+            await session.commit()
+            succeeded += 1
+        except Exception as exc:
+            log.warning("ind_summary.translate.persist_failed",
+                        lang=lang, summary_id=summary_id, error=str(exc)[:200])
+            failed += 1
+
+    if failed and succeeded == 0:
+        log.error("ind_summary.translate.all_failed",
+                  summary_id=summary_id, attempted=len(active_langs))
+    else:
+        log.info("ind_summary.translate.done",
+                 summary_id=summary_id, succeeded=succeeded, failed=failed)
 
 
 # Strip known meta-commentary openers the model sometimes leaks despite the