Recognize the request before generating the answer.
Test a compact local model that predicts prompt task and complexity, then converts the visible input into a transparent cost comparison. The prompt stays in your browser.
A visible result, with its uncertainty left intact.
The test exposes task probabilities, complexity confidence and cost assumptions separately. Classification is measured locally; routing remains a future decision layer.
1,544 local features
Word and bigram hashes, character trigrams and eight structural signals.
Two predictions
Nine task probabilities and three independent complexity probabilities.
No hidden certainty
Low-confidence results remain visibly uncertain instead of becoming definitive labels.
Browser-native
The ONNX model and feature extraction run on the client through WebAssembly.
Inspect the model, training process and evaluation.
The complete ONNX release includes the model card, reproducible training scripts, configuration, dataset manifest, evaluation results, license and browser integration examples.
The classifier identifies. The commercial layer decides.
This page intentionally stops before production routing. It demonstrates recognition and comparison, while preserving the monetization boundary around execution policy and optimization.
Recognition
- Task and complexity probabilities
- Confidence and uncertainty
- Local browser inference
- Token and cost comparison preview
- Open ONNX release
Execution decisions
- Provider and model selection
- Context and tool policies
- Budgets, fallback and safeguards
- Organization-specific rules
- Measured quality and routing improvement
Classification is the first signal, not the final decision.
A future Veritiana routing layer can combine task, complexity and confidence with model capability, provider pricing, context requirements, policy and observed execution quality.
What this test deliberately does not claim.
No real-world accuracy claim
Internal weak-label evaluation is not an independent production benchmark.
No automatic best-model claim
The processing profile is illustrative and does not select or execute a provider route.
No billing replacement
Cost values are versioned API-equivalent assumptions, not provider invoices.