← PactSignal · Project story

Change the message. Inspect the model.

Real local inference in your browser, using the exported trained model. Nothing you type is sent to a server. Use fictional English messages.

One message, a different decision

Downloading public model weights once. Inference runs on this device.

This tests message intent, not mailbox connection, identity matching or a complete agreement timeline. It shows contextual evidence rules alongside the unchanged raw model. Neither changes an order.

Contextual review

Raw model suggestion

—

Confirmed received from this text: 0. A message cannot authenticate a payment.

What influenced this suggestion?

    All model scores — not calibrated probabilities

    Even a payment-notice prediction requires separate receipt verification. No financial action is performed.

    Contextual analysis evaluation

    Loading the synthetic regression evaluation…

    Inspect every example, result and source fingerprint

    New challenge set: results, including failures

    20 AI-authored English messages, fixed after model training. Historical baseline; these examples now inform development of the context layer. This is an engineering check, not independent human or real-world validation.

    Inspect all 20 examples and expected labels
    MessageExpectedModelModel + rules

    Human review: write your expected label before pressing Analyze. Unseen examples and disagreement are useful evidence; a simulation is not an independent user.