A working local preview for research, decision modeling, and task classification. This page describes what it does, how it works, and where it stops.
Two bounded research workflows
Research
Add plain-text sources and enter a question. The workflow indexes passages, runs keyword retrieval, removes duplicate excerpts, and checks the exact source location of each finding. Select findings, add your own interpretation, and export a Markdown brief.
Reference review
Every returned quote is compared with its original source text at the recorded character offsets. Changing or deleting a source invalidates the current research results. The review checks reference integrity; it does not check whether the source is true or whether a claim follows from it.
These are deterministic workflows. They do not plan autonomously, generate answers, search the web, send messages, or execute instructions in a document. There is no large language model in this release.
Decisions
Create 2–6 options and 1–8 criteria. Assign nonnegative weights and scores from 0 to 10, where higher means better. Each weighted result is the sum of criterion scores multiplied by normalized weights.
option score = Σ(score × weight) / Σ(weights)
The sensitivity check changes one weight at a time by −20% and +20%, renormalizes the weights, and reports changes in the leading option or tie. It does not test every combination of weights or uncertainty in the scores. Rankings express the user's judgments; they are not predictions or independently validated recommendations.
Model cards
SKAR ROUTE 0.1
A small, trainable task classifier
Intended use
Suggest Research, Decision, or Evaluation for short English workspace requests. Custom labeled examples can train a replacement classifier in the browser.
Algorithm
Multinomial Naive Bayes with word counts, learned class priors, and Laplace smoothing (alpha = 1). Implemented in plain JavaScript.
Base training set
48 synthetic English requests authored for this release, evenly split across three labels. No user documents were used. Texts are available in the downloadable dataset and source code.
Features
Unicode word tokenization, lowercasing, and a fixed English stop-word list. No embeddings, stemming, pretrained language model, or external inference.
Review rule
Hold a request for review when it contains fewer than two known word occurrences, has under 25% known vocabulary coverage, has a leading model share below 55%, or has a top-two margin below 12 percentage points.
Limits
Uncalibrated model shares can be overconfident. Word overlap can cause incorrect routing; unfamiliar phrasing can be rejected. The model does not understand intent or reason about a task. It is not evaluated for non-English use or consequential automated decisions.
EVIDENCE RANK 0.1
Exact excerpts, ranked locally
Intended use
Find source passages containing the terms in a question, retaining exact source references for review.
Algorithm
BM25 with k1 = 1.2 and b = 0.75. The corpus is rebuilt from the current sources for each run; this model requires no training.
Text handling
Paragraphs and sentences become passages capped at 900 characters. Results include the source ID, title, original quote, and zero-based UTF-16 character offsets with an exclusive end.
Output
Up to eight distinct excerpts. Scores indicate lexical relevance, not factual accuracy, confidence, or completeness.
Limits
No semantic search, synonym handling, web access, PDF extraction, fact checking, or claim entailment. A source using different wording may be missed. The engine may retrieve irrelevant passages that share keywords.
Reproduce this release with node workspace/scripts/train-route.mjs --check in the source repository. The report includes dataset fingerprints so changes can be traced.
The base task router returns the expected label on 20 of 24 synthetic held-out cases (83.3%). It holds six requests for review. The set includes six out-of-scope requests, and two of those are incorrectly routed to Research. Two legitimate requests using unfamiliar wording are also held for review.
Of the 18 requests given a route, 16 are correct. Of the six out-of-scope requests, four are correctly held for review. These counts separate correct routing from successful rejection.
Training and test texts are separate, but they were authored for the same small domain. This diagnostic is not an independent benchmark, a measure of general intelligence, or a comparison with OpenAI or xAI. Independent task data and broader testing are needed before expanding the model's role.
Failure example
Expected
Actual
Look through the notes for unanswered questions
Research
Needs review
How often does the routing system get things wrong
Custom evaluations reject repeated requests and training overlap after normalizing Unicode, case, and whitespace. This does not detect paraphrases or other semantic leakage.
Your data stays on your device
The workspace processes text, training data, and scoring inputs in your browser. It makes no AI API calls and includes no analytics. The website host serves the application files and may receive normal page requests; document contents are not transmitted.
Work is held in memory unless you choose Save on device. That saves sources, the question, notes, and decision inputs in this browser's local storage. The saved copy is not encrypted by the application. Export a workspace JSON file for a portable backup; browser storage can be cleared or unavailable.
Search results are regenerated after reopening a project. Models stay in memory unless you choose Save model on device. This saves the model and available training examples in unencrypted browser storage. Remove saved model deletes that stored copy. Export reusable model creates a portable bundle with the same information; Open model restores it. Raw 0.1 model exports can also be opened, but without training examples their evaluation cannot check training overlap. Model imports validate counts, vocabulary, and totals; bundles with training examples must reproduce the supplied model.
Limits: 12 source files and 200,000 source characters per workspace, plain .txt or .md text, 500 training examples and 500 custom evaluation cases. Use a current browser with JavaScript enabled.
Train, reuse, and classify
Choose Build or edit a model to label requests directly. Use base examples or start blank. Repeated requests increase their influence; conflicting labels for the same normalized request are rejected. Training and batch classification run in a local worker to keep the interface responsive.
Batch classification accepts up to 500 requests, one per line. Results identify suggestions and cases held for review, with CSV export. Editing inputs or changing models clears previous batch results. Category suggestions do not execute tasks. For the built-in Research, Decision, and Evaluation categories, an explicit button opens the relevant workspace; a research request is carried into the question field.
Workbench 0.2 improves model management and workflows. The base Skar Route model and its measured 0.1 diagnostic results are unchanged.
Start with a small project
Open Workbench and choose Try a sample project, or add your own text files.
Ask a specific question and select Find evidence.
Open the references, select relevant excerpts, and write your interpretation.
Export the brief, then compare alternatives in Decisions.
Use the Models section of Workbench to inspect, retrain, and evaluate the task classifier.
The included support-pilot documents and scores are fictional examples, not a client case study.
