Independent evidence
Choose a focused set of policies, freeze the test protocol and assess harmful messages alongside difficult benign conversations.
Research & partnerships
Help make private conversations safer.
E2EM is developing moderation that checks messages on the device. We’re seeking partners to validate what works, reduce false alarms and build a phone-native prototype.
A practical question, with testable answers
Can useful moderation happen locally, without sending a conversation to a separate scoring service? An app could use a local score to offer a warning, blur content or ask for review. The app chooses its rules and what happens next.
Gandalf 0.0.1 is our first versioned research release: downloadable weights, exact policy definitions, frozen thresholds and recorded evaluation evidence. The original model package is 291.72 MB, with 70.8 million parameters including embeddings. It runs locally on CPU; physical-phone performance remains to be established.
No category currently passes the formal acceptance gate. The model card publishes the measured results and limitations so that partners can assess the work and reproduce it.
What the next phase would deliver
A phone-native prototype
Test the same frozen model on physical iPhone 14 and Pixel 6 devices. Measure latency, memory and energy use with local processing.
A controlled pilot
Evaluate user-facing warnings and operator feedback with a deployment partner. Publish what succeeds and what still needs work.
Jacob Cattrall Ltd’s immediate funding priority is training compute: compare compact model candidates, reduce false alarms and publish reproducible results. Independent evaluation and device engineering form the next phase, with deployment limited to the policies and conditions that meet agreed criteria.
Who we want to work with
Research teams with expertise in evaluation and online harms; privacy-sensitive messaging and community products; and funders interested in transparent, measurable safety technology.
The initial focus is English text. Independent human assessment, release licensing and a mobile runtime are part of the development work ahead. Local inference alone does not establish an app’s privacy, security or regulatory compliance.
Start a research or integration discussion on Hugging Face. Keep public discussions free of private conversations and identifying information.