Kush v7 Release Notes
Kush v7 is now the production release.
This version is not a cosmetic bump. It is a cleanup of the entire release path:
- the runtime prompt now comes from one canonical contract
- the training payload rejects duplicate rows, prompt bloat, and mixed-version leakage
- the eval gate runs pairwise comparisons when a baseline is available
- the load test checks the live model instead of assuming the compare model is loaded
- the deployment path is aligned so the Mac mini does not keep multiple Kush variants warm at once
The goal is simple: make the model more consistent, easier to maintain, and harder to regress.
What changed
The biggest change is structural. Kush used to rely on prompt text scattered across runtime, persona fallback, export scripts, and training builders. That created drift. Different parts of the system were teaching slightly different versions of the same assistant.
Kush v7 replaces that with one short prompt contract. The runtime, the persona fallback, and the export path now all point at the same source of truth. That matters because prompt quality is not just about wording. It is about consistency.
We also tightened the corpus pipeline:
- long system prompts are blocked
- repeated system prompts in a single row are blocked
- near-duplicates are removed
- mixed-version leakage is blocked
- low-signal rows are filtered out before export
This keeps the model from learning noise at scale.
Why it matters
Models drift when the data and the prompt stop agreeing with each other. The result is familiar:
- inconsistent address forms
- unnecessary repetition
- weaker persona control
- brittle behavior under load
Kush v7 addresses those problems by making the prompt contract short, structured, and testable. It also moves release decisions toward evidence:
- corpus lint
- paired evaluation
- judge sampling
- load testing
That is the difference between a model that looks good in a demo and a model that survives production traffic.
What to expect
You should see a cleaner voice, fewer duplicate phrases, and more predictable behavior under real use. The model should be easier to retrain because the release pipeline now gives us a clearer signal about what is working and what is not.
What comes next
The next step is not more prompt inflation. It is more signal:
- higher-quality feedback from live users
- better baseline coverage in evals
- a tighter loop between production traces and retraining candidates
Kush v7 is the current release. The release process around it is now the product.