EVIDENCE, WITH A TRAIL
Make the sources useful.
Ask a question. Find the passages. Build a brief you can inspect.
Sources 0
Use text exports for PDFs and Word documents. Source contents are treated as data, never as commands.
Start with the evidence.
Add your source material or open the sample project. The research workflow will find relevant passages and retain their exact references.
Your interpretation
Written by youJUDGMENT, MADE EXPLICIT
Understand the tradeoffs.
Compare alternatives using your criteria. See how changes affect the result.
These results reflect your inputs, not a prediction. The sensitivity check changes one weight at a time by ±20%; it is not an exhaustive uncertainty analysis.
SMALL MODELS. VISIBLE METHODS.
Test what the model does.
Build a classifier, test real requests, and reuse it in your work. Everything runs on this device.
Skar Route 0.1
Suggests Research, Decision, or Evaluation using a small learned word-count model. Review the category, then open a workspace to continue.
- Method
- Multinomial Naive Bayes
- Training data
- 48 synthetic examples
- Execution
- Your browser · no API
Evidence Rank 0.1
Ranks source passages against a question using keyword relevance. Returns exact text and character offsets so you can verify every excerpt.
- Method
- BM25, k1 1.2 / b 0.75
- Training required
- None · indexed per search
- Limits
- Lexical matches, not semantic reasoning
Try the task router
Base modelTrain with your own examples
Import a JSON array of objects with text and label. Use 6–500 examples across 2–12 labels. Training stays on this device. Save your model or export a reusable model to keep it after closing this page.
Reproducible evaluation
Base model readyRun 24 held-out synthetic requests, including unfamiliar phrasing and out-of-scope cases. This is a small diagnostic set, not evidence of general capability or a comparison with frontier models.
Test JSON uses the same text / label format as training data. It evaluates the currently loaded classifier; keep test examples separate from training.
PUT YOUR MODEL TO WORK
Classify a batch of requests
Paste one request per line. Review the categories and export a CSV for your workflow.