Three Operators Under One Kernel
Data, harness, and model edits share one loop kernel and one artifact vocabulary.
One kernel for three surfaces
Data edits grow good traces and mark the edge of skill. Harness edits change prompts, tools, contracts, and workflow without touching weights. Model edits turn kept traces into weight updates on a schedule. One loop kernel serves all three so results can be compared.
One artifact vocabulary links task, trace, score, lesson, patch, and verdict across operators. A scheduler picks operator order and proposal policy. A meta policy revises the schedule across terms when evidence supports a shift. Small budgets keep each operator honest.
How the three compose
A common path starts with data. Fresh traces reveal a repeated failure. Harness edits fix context flow or tool shape. Kept traces from the stronger harness then feed model training. Each surface inherits the gain of the prior surface.
A second path skips weights entirely. Many teams can only reach a model through an interface. Harness edits plus memory still compound because workflow improves even when weights stay fixed. Data remains the shared substrate that feeds both paths.
Scheduling with care
The scheduler weighs expected gain, cost, risk, and coverage. Low risk harness edits run often. Costly model updates run on a slower cadence with stronger proof. Data growth runs in the background with strict dedup and quality filters.
Direction changes need human review. A schedule that drifts without consent is treated as a defect. Diaries record which operator ran, why it ran, what it cost, and what it kept.
Operator comparison
| Operator | Changes | Cadence | Proof needed |
|---|---|---|---|
| Data | Traces, fixtures, curriculum | Continuous | Quality plus diversity checks |
| Harness | Prompts, tools, contracts, workflow | Fast cycles | Full suite plus judge review |
| Model | Weights through bounded training | Scheduled | Held out transfer plus cost review |