diff --git a/backend/app/pipeline/agents/agent_1_validator.py b/backend/app/pipeline/agents/agent_1_validator.py
index d7c1cca..433f5d7 100644
--- a/backend/app/pipeline/agents/agent_1_validator.py
+++ b/backend/app/pipeline/agents/agent_1_validator.py
@@ -85,6 +85,7 @@ class Agent1Validator(BaseAgent):
                 "locale_considerations_file"
             ),
             date_pct_formats_file=self.file_manifest_raw.get("date_pct_formats_file"),
+            supplementary_files=self.file_manifest_raw.get("supplementary_files", []),
         )
 
         # Load reference files (for validation)
diff --git a/backend/app/pipeline/agents/agent_single.py b/backend/app/pipeline/agents/agent_single.py
index 39dde3b..4f9abd3 100644
--- a/backend/app/pipeline/agents/agent_single.py
+++ b/backend/app/pipeline/agents/agent_single.py
@@ -57,15 +57,27 @@ def _resolve_all_tm_paths(
     """Resolve TM file paths for multiple channels.
 
     Tries the legacy hardcoded CHANNEL_FILE_MAP first (which has the
-    canonical lowercase patterns for the original channels). For any
+    canonical lowercase keys for the original channels). For any
     channel not in the map (e.g. newly registered like PrimeCBM), falls
     back to the generic `flat_<Channel>_<lc>.json` pattern.
-    Missing files at the resolved path are silently skipped at load time.
+
+    If the canonical path doesn't exist on disk, scan the locale's TM
+    directory for any filename that matches case-insensitively and use
+    that path instead. Protects against drift between the registry's
+    stored channel casing and the on-disk filename casing — both code
+    paths historically have produced mixed-case outputs.
     """
     lc = locale_code.lower()
     tm_dir = os.path.join(settings.STORAGE_ROOT, "amazon", "tm", locale_code)
     paths: list[str] = []
 
+    # Build a one-time case-insensitive index of the locale's TM directory.
+    # Empty/missing dir is fine — the resolved paths just won't exist.
+    dir_index: dict[str, str] = {}
+    if os.path.isdir(tm_dir):
+        for entry in os.listdir(tm_dir):
+            dir_index[entry.lower()] = entry
+
     for ch in tm_channels:
         pattern = CHANNEL_FILE_MAP.get(ch.lower())
         if pattern:
@@ -74,7 +86,14 @@ def _resolve_all_tm_paths(
             # Auto-discovered channel — use the generic pattern with the
             # channel name as-cased in the registry.
             filename = TM_FILENAME_PATTERN.format(channel=ch, lc=lc)
-        path = os.path.join(tm_dir, filename)
+
+        # Case-insensitive lookup: if the canonical filename doesn't
+        # exist exactly, resolve against the dir index.
+        actual = dir_index.get(filename.lower())
+        if actual is not None:
+            path = os.path.join(tm_dir, actual)
+        else:
+            path = os.path.join(tm_dir, filename)  # leave as-is; will be skipped at load
         if path not in paths:
             paths.append(path)
 
@@ -461,7 +480,16 @@ class AgentSingle(BaseAgent):
         channel = context.job_params.channel
         programme = context.job_params.programme
         campaign_name = context.job_params.campaign_name
-        tm_channels = context.job_params.tm_channels or [channel]
+        # Distinguish "user explicitly sent empty list" (load no TMs) from
+        # "tm_channels was None" (legacy job → fall back to campaign channel).
+        # Bare `or` treats [] as falsy and would re-default to [channel] —
+        # which silently reintroduces the campaign TM after the user asked
+        # for none.
+        tm_channels = (
+            context.job_params.tm_channels
+            if context.job_params.tm_channels is not None
+            else [channel]
+        )
 
         logger.info(
             "AgentSingle starting: locale=%s channel=%s tm_channels=%s lines=%d",
diff --git a/backend/app/tasks/job_tasks.py b/backend/app/tasks/job_tasks.py
index 283987f..6937efd 100644
--- a/backend/app/tasks/job_tasks.py
+++ b/backend/app/tasks/job_tasks.py
@@ -282,7 +282,14 @@ def process_locale_instance(self, job_id: str, locale_code: str) -> dict:
                     "programme": job.programme.value,
                     "campaign_name": job.campaign_name,
                     "context_prompt": job.context_prompt,
-                    "tm_channels": job.tm_channels or [job.channel],
+                    # Preserve an explicit empty list ("no TMs") from
+                    # the user; only fall back to the campaign channel
+                    # when the field was never set (legacy jobs).
+                    "tm_channels": (
+                        job.tm_channels
+                        if job.tm_channels is not None
+                        else [job.channel]
+                    ),
                     "llm_model": job.llm_model,
                 }
 
diff --git a/frontend/src/components/jobs/JobWizard/StepReview.tsx b/frontend/src/components/jobs/JobWizard/StepReview.tsx
index 6ac9d37..b86c585 100644
--- a/frontend/src/components/jobs/JobWizard/StepReview.tsx
+++ b/frontend/src/components/jobs/JobWizard/StepReview.tsx
@@ -51,8 +51,11 @@ export function StepReview({ data, onBack }: StepReviewProps) {
         programme: data.programme.toLowerCase(),
         channel: data.channel.toLowerCase(),
         sub_channel: data.sub_channel ? data.sub_channel.toLowerCase() : null,
-        // Always send an array (even empty) so the backend respects "no TMs" vs "default to channel"
-        tm_channels: data.tm_channels.map(c => c.toLowerCase()),
+        // Always send an array (even empty) so the backend respects "no TMs" vs "default to channel".
+        // Do NOT lowercase the channel names — the pipeline resolves TM filenames using the
+        // exact case from the registry (e.g. flat_PrimeCBMT_nl-be.json). Lowercasing breaks
+        // case-sensitive filesystem lookups for any channel not in the legacy CHANNEL_FILE_MAP.
+        tm_channels: data.tm_channels,
         llm_model: data.llm_model || undefined,
         locale_codes: data.locales,
         context_prompt: data.context_override || undefined,