Anyone else having consistency problems with Jasper's brand voice?

I've been conducting an extensive evaluation of several AI writing assistants for generating technical documentation and architecture briefs, with a particular focus on maintaining a consistent, professional brand voice across multiple documents and outputs. My primary stack involves AWS, GCP, and Azure, so the content is inherently complex. While I've tested Copy.ai, ShortlyAI, and Writesonic, my team gave Jasper a substantial trial due to its strong marketing around "Brand Voice."

The promise was compelling: ingest your style guide and existing documents, then produce on-brand content. In practice, the inconsistency is problematic, especially for technical material. The tool seems to struggle with the nuanced, precise language required in our field.

Here are my observed pain points:

* **Fluctuating Technical Rigor:** One output will correctly use formal terms like "multi-availability zone deployment" and "service mesh ingress controller." The next, on a similar topic, will lapse into vague marketing speak like "leverage cloud scalability" and "orchestrate your containers," which we explicitly forbid.
* **Inconsistent Formatting Directives:** We provide clear examples preferring bulleted lists for specifications and code blocks for Terraform snippets or Istio configurations. Jasper will sometimes respect this, but often reverts to paragraph form for lists or attempts to write pseudo-code in a sentence, which is unusable.
* **Tone Drift:** Our required tone is academic and authoritative, avoiding hype. Despite training on our whitepapers, Jasper intermittently injects enthusiastic adjectives and "powerful," "seamless," "robust" modifiers that we must manually strip out. It fails to maintain a steady, objective cadence.

For example, I used an identical prompt for two different cloud security overviews:

**Prompt:** "Draft a short paragraph on the principle of zero-trust networking in a hybrid cloud context, referencing network policy enforcement."

**Output A:** "A zero-trust architecture mandates that no entity, internal or external, is implicitly trusted. In hybrid cloud scenarios, this requires consistent network policy enforcement at choke points, which can be implemented via service mesh sidecar proxies (e.g., Istio) and cloud-native firewall rules, creating a unified security perimeter."

**Output B:** "Embrace a zero-trust model to powerfully secure your hybrid cloud! By seamlessly enforcing network policies across your environments, you can robustly protect your workloads whether they're in Azure or on-premises."

Output A is acceptable. Output B is a drastic tonal shift and would require a complete rewrite, negating the efficiency gain. This level of inconsistency forces a full editorial review on every single output, which defeats the purpose of automation for high-volume documentation.

My hypothesis is that the underlying model's propensity for "creative" variation is overriding the brand voice training on specialized vocabulary and tone. It feels like the brand voice feature is a thin layer on top of a general-purpose LLM, rather than a fine-tuned model truly constrained by the provided style guide. Has anyone else in the technical writing or architecture space encountered this? I'm particularly interested if others have found specific "Brand Voice" training configurations or prompt engineering techniques that yield more deterministic results for technical content.