The agent ranks unlabeled tickets by predictive entropy.
In-browser machine learning / WebMCP
Train a live classifier with your browser agent.
A multinomial Naive Bayes model learns incrementally from human-confirmed support tickets. The agent finds uncertain samples, requests labels, starts bounded training steps, and reports metrics, explains learned features, and serves predictions through nine page tools.
Checking for native WebMCP...
A person assigns the authoritative billing, bug, or access label.
The browser updates token counts without a server or model API.
Accuracy, macro F1, confidence, and log loss update on a hidden holdout.
Model core
client-side
Online Naive Bayes
Laplace alpha1.00
Vocabulary0
Review threshold72%
Checkpoints1
Accuracy historyseed model
Acquisition queue
12 unlabeled
Uncertain samples
Entropy measures disagreement across the three class probabilities. Higher values are better candidates for a human label.
Model observability
WebMCP readable
What the classifier learned
Top discriminative tokensone-vs-rest log weights
Holdout confusion matrixactual rows / predicted columns
Calibration checkconfidence against accuracy
Available to agents through inspect_model_diagnostics. New text can be queried with predict_ticket.
Shared protocol
0 calls
WebMCP call trace
Every read and write is recorded with its source. Local replay is labeled and never presented as an external agent.
Training ledger
Human and model events
Why WebMCP
The page owns the model. The agent works through contracts.
inspect_uncertain_samples returns ranked candidates with probabilities and entropy. The agent does not scrape cards or infer control meaning from the layout.
queue_label_review can request labels but cannot assign them. train_confirmed_batch rejects unreviewed predictions, and configuration changes remain pending until a person accepts them.