diff --git a/ai_pipeline/autoria_ai/generator.py b/ai_pipeline/autoria_ai/generator.py
index a697896..3448474 100644
--- a/ai_pipeline/autoria_ai/generator.py
+++ b/ai_pipeline/autoria_ai/generator.py
@@ -42,23 +42,27 @@
 # an unknown key is dropped in silence, not rejected. ``max_new_tokens`` is the
 # *text-generation* spelling and has no effect on ``ModelInference.chat()``,
 # which is what we call — see backend/tests/test_generation_params.py.
-# 320, not the 512 originally written here, because 512 does not fit the
-# latency budget. Measured against real Watsonx (eu-de, llama-3-3-70b, n=5 per
-# cell, hard timeout lifted so the true duration shows):
+# The cap stays at 512 because nothing is wasted there: across ten recorded
+# generations (bob/sessions/Sprint_1/baseline_eval.md) the longest output was
+# 415 tokens and all ten ended on a complete sentence. Lowering the cap would
+# cut real prose, not slack.
+#
+# Measured against real Watsonx (eu-de, llama-3-3-70b, n=5 per cell, hard
+# timeout lifted so the true duration shows):
 #
 #     cap=512  vanilla  median 4.5s  worst 5.9s (~196 words)
-#              autoria  median 7.2s  worst 8.4s (~313 words)   <- over the 8s cap
+#              autoria  median 7.2s  worst 8.4s (~313 words)
 #     cap=320  vanilla  median 5.0s  worst 5.3s (~242 words)
 #              autoria  median 5.4s  worst 5.6s (~223 words)
 #
-# The conditioned branch straddles HARD_TIMEOUT_SECONDS at 512: it sometimes
-# returns and sometimes exhausts all four attempts, and because a failed
-# AutorIA branch cannot be degraded away (the passport needs it), that surfaces
-# as an intermittent 503 from POST /api/generate. 320 keeps a comfortable
-# margin under both the 8s timeout and the 10s client-side abort in
-# frontend/src/lib/api.ts, and still yields ~220 words per column.
+# At 512 the conditioned branch runs past the *old* 8s HARD_TIMEOUT_SECONDS
+# often enough to matter, and a failed AutorIA branch cannot be degraded away
+# because the passport needs it — that surfaced as intermittent 503s. The fix
+# is on the timeout side, not the cap side: see HARD_TIMEOUT_SECONDS in
+# backend/app/services/watsonx_client.py. A generation that is merely slow
+# should arrive late, not die.
 _GENERATION_PARAMS: dict[str, Any] = {
-    "max_tokens": 320,
+    "max_tokens": 512,
     "temperature": 0.7,
     "top_p": 0.9,
 }
diff --git a/backend/app/services/watsonx_client.py b/backend/app/services/watsonx_client.py
index af4c790..c61bc0f 100644
--- a/backend/app/services/watsonx_client.py
+++ b/backend/app/services/watsonx_client.py
@@ -20,7 +20,16 @@
 
 logger = logging.getLogger(__name__)
 
-HARD_TIMEOUT_SECONDS = 8.0
+# 15s, not the 8s this used to be. The old value was the MVP's *SLA* number
+# ("side-by-side <8s P95", docs/MVP.md) used as a *kill* threshold, which left
+# no band in which a generation is late but still working: a conditioned call
+# measured at 8.4s worst case was retried and ultimately failed rather than
+# returned, surfacing as an intermittent 503 (#106).
+#
+# 15s is ~1.8x the worst conditioned generation observed at max_tokens=512
+# (8.4s, eu-de, llama-3-3-70b). The SLA is unchanged and still tracked; it is
+# now measured on results instead of enforced by killing them.
+HARD_TIMEOUT_SECONDS = 15.0
 # Three retries after the first failure → four attempts total.
 _RETRY_DELAYS_SECONDS: tuple[float, ...] = (1.0, 2.0, 4.0)
 _DEFAULT_URL = "https://us-south.ml.cloud.ibm.com"
diff --git a/backend/tests/test_generation_params.py b/backend/tests/test_generation_params.py
index 171dbae..28beebc 100644
--- a/backend/tests/test_generation_params.py
+++ b/backend/tests/test_generation_params.py
@@ -64,7 +64,7 @@ def test_generation_params_match_chat_schema():
 
 def test_output_cap_is_declared():
     """The output cap must be present — an absent key falls back to 1024."""
-    assert _GENERATION_PARAMS.get("max_tokens") == 320
+    assert _GENERATION_PARAMS.get("max_tokens") == 512
     assert "max_new_tokens" not in _GENERATION_PARAMS
 
 
diff --git a/frontend/src/lib/api.ts b/frontend/src/lib/api.ts
index f35fcd3..7fe3cd7 100644
--- a/frontend/src/lib/api.ts
+++ b/frontend/src/lib/api.ts
@@ -179,9 +179,15 @@ export function slugifyAuthorName(name: string): string {
 /**
  * POST /api/generate — run vanilla + AutorIA generation in parallel on the backend.
  *
- * Uses AbortController with a 10 s timeout so a hanging Watsonx call resolves
+ * Uses AbortController with a 20 s timeout so a hanging Watsonx call resolves
  * to a timeout state rather than an infinite spinner (design-system §1 / §7.4).
  *
+ * 20 s, not the 10 s this used to be, so the client outlives the server: the
+ * backend allows a single Watsonx attempt 15 s (HARD_TIMEOUT_SECONDS) and the
+ * worst conditioned generation measured at max_tokens=512 took 8.4 s. A 10 s
+ * abort here would cancel requests the backend was about to answer, and the
+ * user would see a timeout for a generation that actually succeeded.
+ *
  * Return contract:
  *   200       → resolves with GenerateResponse
  *   AbortError → throws with name "AbortError" (timeout path)
@@ -192,7 +198,7 @@ export async function generateText(
   prompt: string,
 ): Promise<GenerateResponse> {
   const controller = new AbortController();
-  const timerId = setTimeout(() => controller.abort(), 10_000);
+  const timerId = setTimeout(() => controller.abort(), 20_000);
 
   try {
     const body: GenerateRequest = { author_id: authorId, prompt };