Qwen3-8B
Existing open-weight base in the current grounding service.
Available today: grounding + recordsA different question
for the foundation.
Can a base be designed with less unmanaged factual knowledge, while managed layers provide the knowledge an application needs?
Craton
Change the foundation question.
Craton and Craton Pro investigate a base with less unmanaged factual knowledge.
Managed knowledge
Trained layer application remains in development
Prior knowledge in the existing base is not fully attributable to managed sources.
A conceptual comparison, not a performance simulation. Both programmes remain in development and pre-acceptance. Reported weight-path acceptance: 0 / 15.
Reported evidence
Pre-acceptance means the negative result remains visible.
| Question | Reported evidence | Interpretation |
|---|---|---|
| Does the tested weight path transfer usable answers? | 0 / 15 in the reported test. | Acceptance has not been met. |
| Does the tested base avoid the probed facts? | 10 / 10 on a limited sterility test. | A limited probe, not proof of universal factual sterility. |
The weight-path failure is separate from the context-packing defect. Fixing the latter does not establish that Craton passes its weight-path acceptance test.
What progress must establish
A usable path, explicit acceptance and separated contributions.
Craton is a from-scratch base-model programme. Craton Pro is its larger class. Both remain in research and development; a zero attribution gap has not been demonstrated.
Progress needs tests that separate base prior knowledge, retrieved context and trained-layer contribution. The programme is not presented as a model available for purchase.
Continue the research trail
Keep independent failures independent.
Reported negative results remain part of the evaluation history.