diff --git a/CSCS-PROPOSAL.md b/CSCS-PROPOSAL.md
index e71ab6a..972566e 100644
--- a/CSCS-PROPOSAL.md
+++ b/CSCS-PROPOSAL.md
@@ -25,7 +25,7 @@ One deployment detail, and it's the one we think matters most to you: **inferenc
 This means:
 
 - No new infrastructure for CSCS to stand up — the proposal consumes a service you've already launched.
-- A governance story consistent with what Apertus is for: EU AI Act-aligned, fully open, auditable — rather than routing researcher queries through an opaque third-party model.
+- A governance story consistent with what Apertus is for: fully open, auditable, and described by CSCS as designed with the EU AI Act in mind — rather than routing researcher queries through an opaque third-party model.
 - Everything we demo live is architecturally identical to what we'd deploy — not a "trust us, production is different" hand-wave.
 
 ## Why this fits the hackathon's ask specifically
diff --git a/docs/INFERENCE.md b/docs/INFERENCE.md
index 543a066..8d774d2 100644
--- a/docs/INFERENCE.md
+++ b/docs/INFERENCE.md
@@ -68,7 +68,7 @@ config line is the claim that fails in front of the person who built the endpoin
 
 ## Why this matters for the pitch, specifically
 
-- **Sovereignty / EU AI Act alignment**: Apertus is fully open (weights, training data, alignment documented) and explicitly positioned by CSCS as compliant with the EU AI Act. That's a stronger governance story than "we called OpenAI's API."
+- **Sovereignty / EU AI Act alignment**: Apertus is fully open (weights, training data, alignment documented) and described by CSCS as designed with the EU AI Act in mind; we make no compliance claim of our own (see [trust-and-compliance.md](trust-and-compliance.md)). That's a stronger governance story than "we called OpenAI's API."
 - **No inference infrastructure to operate**: the proposal doesn't ask CSCS to stand up anything new — it consumes a service they've already launched.
 - **Same tokens, no re-training the agent**: because DocMind and Hermes both speak OpenAI-compatible chat completions, nothing about the MCP tools, the Multica orchestration, or the Telegram channel changes when we point at Apertus instead of Ollama. The demo we show live is architecturally identical to the deployment we're proposing — we are not hand-waving a "in production this would be different."
 
diff --git a/docs/MULTICA.md b/docs/MULTICA.md
index 4efa9ad..9ea8299 100644
--- a/docs/MULTICA.md
+++ b/docs/MULTICA.md
@@ -157,8 +157,9 @@ Why this is the strongest single argument in the project:
   retrieval, open model, all auditable. The alternative is routing researcher queries
   through an opaque third-party model, which is the thing CSCS's inference service exists to
   avoid.
-- **EU AI Act alignment stops being a slogan** and becomes a deployment fact, because the
-  model is the one CSCS positions for exactly that.
+- **The EU AI Act story rests on the model, not on a slogan**: Apertus is described by CSCS
+  as designed with the AI Act in mind. We make no compliance claim of our own (see
+  [trust-and-compliance.md](trust-and-compliance.md)).
 - **It fixes a measured problem.** On CPU, one answered question took 45 seconds and
   ingesting 170 KB took thirteen minutes — we measured both. The same stack against an
   inference endpoint turns that into seconds, which is the difference between a system that