What we're testing right now
Research questions, evaluation boundaries, and protocols under review. Results are published only when the run data and scoring method are ready to inspect.
Protocols under review
Multi-Agent Task Routing
Define how specialized agents should receive subtasks, expose their source context, and stop for approval. The protocol compares routing quality, latency, and operating cost.
Distillation Efficiency Benchmarks
Specify a repeatable comparison between a compact task model and its larger reference model across classification, summarization, cost, and failure cases.
DomeAI Retrieval Latency
Map retrieval quality and response time as a private context store grows. Results will be reported only with the dataset, index settings, and hardware attached.
Long-Context Faithfulness
Test whether instruction placement changes compliance in long contexts, using the same tasks and scoring rubric across each model under review.
Synthetic Data Quality Scoring
Design a scoring rubric for generated training examples before any sample is admitted to a task-specific training set.
Agent Self-Correction Loops
Compare a baseline agent with a review-step agent while keeping tasks, tool access, and scoring constant. Human approval remains part of the test boundary.
How a completed experiment should run
A completed experiment should follow a structured process so its evidence can be reviewed and repeated.
Hypothesis
Start with a clear, testable question. What are we trying to prove or disprove?
Design
Define metrics, control variables, dataset size, and success criteria before running anything.
Execute
Run the experiment with logging at every step. Capture raw data, not just conclusions.
Analyze & Document
Document positive and negative results under the same reporting standard when the evidence is reviewable.
Questions the protocols must answer
How small can a task model be before error cost outweighs inference savings?
The protocol records accuracy, abstention behavior, review load, latency, and cost together so a cheaper model is not treated as a win by default.
Which context placement produces the most faithful output?
Each placement uses the same instructions and tasks. Any reported difference must include the model, context length, run count, and scoring method.
When does an agent review step reduce errors enough to justify the latency?
The evaluation separates task completion from safe completion and records how often a human still needs to correct or reject the prepared action.
Common questions about our experiments
Have an experiment idea?
Bring a specific hypothesis, dataset, or workflow risk. We can define the evidence and review gates a scoped evaluation would require.