Alright, let's cut through the marketing fluff. I'm convinced ContentBot's "deterministic" mode is about as deterministic as a Kubernetes scheduler under a cascading node failure.
I'm running what should be a trivial task: generating a standardized technical summary for our internal service changelogs. Same prompt, same "High Consistency" profile, same temperature setting (0.1). The first run gives me a concise, bulleted list. The second run, minutes later, with the exact same config, decides a three-paragraph prose format is the way to go. The third run omits a critical security note that was present in the first two.
This isn't a "creativity" feature; it's a bug. They're clearly routing requests across different underlying models or there's some hidden context window pollution happening. In my world, a pipeline that produces different YAML from the same template gets rolled back.
Here's the prompt, sanitized:
```
Generate a changelog entry for service version 1.5.0. Use the following data: [Service: "api-gateway"; Change: "Updated request timeout from 30s to 15s"; Impact: "Potential increase in 504 errors for long-running queries"; Mitigation: "Clients should implement retry with backoff"].
Format: brief title, short description, list of changes, impact note. Be technical and succinct.
```
And two consecutive outputs I captured:
**Output A:**
*Changelog: api-gateway v1.5.0*
*Adjusted request timeout configuration to improve resource utilization.*
*- Changed: Global request timeout decreased from 30 seconds to 15 seconds.*
*- Impact: Long-running queries may now terminate with 504 errors. Client-side retry logic is recommended.*
**Output B:**
*Version 1.5.0 of the api-gateway service introduces a key configuration update. The primary adjustment is a reduction of the default request timeout. This change is intended to optimize connection pool usage but requires client-side updates. The mitigation strategy should involve implementing retry mechanisms.*
The second one is objectively worse—it buries the lead, doesn't list the change clearly, and is more verbose. For automated doc generation, this inconsistency is a deal-breaker. Anyone else seeing this kind of drift, or have I just lost the platform lottery?
Your Kubernetes scheduler analogy is too generous. At least that system logs its decisions.
I've seen this exact behavior with their API when the prompt doesn't include an explicit format instruction. Even at low temp, the model makes a "formatting choice" based on cached examples, and if you're hitting different inference endpoints, that cache isn't shared.
Add "Format: use a bulleted list." to your prompt. If you still get paragraphs, then you've proven the routing hypothesis.
Beep boop. Show me the data.