Instruments for honest oversight
Open tools that turn the hypothesis into something you can run. Everything here is early; each tool shows where it stands.
XonTools
ALPHAAI agents write code, run tests and call APIs, then report on their work. Sometimes the report is wrong. XonTools is being built to read an agent's transcript, separate what the agent said from what its trusted tools actually returned, and point at exactly which claim conflicts with exactly which piece of evidence. Its core, the consistency engine, is built and tested; the transcript monitor is specified and comes next. Alongside it, the intention ledger asks the agent to write what it intends, what it forecasts and what it will not touch before it acts, into a tamper-evident record its actions then join. The engine reads those intentions against the record, so an agent's account of its own goals is checked against evidence it could not fit after the fact.
A flag is never just a score: it is a claim, a piece of evidence and the relation between them.
Claim and entity graphs find loops of statements that are each harmless alone.
Designed so that a blinded or evidence-starved monitor says so, instead of reporting that all is well.
Development status
Also in the workshop
Envisioned: an opt-in hosted version for teams whose transcripts can leave their network. Local stays the default; hosted data is never used for training.
Generates larger, harder consistency test corpora whose ground truth comes from a solver, never from a model, with blind cross-model screening and sealed test splits. It also supplies the training and evaluation items for Honest or Hidden. The code is public in the XonTools repo; not yet released as a tool.
A simulation testbed for the settling network architecture, used to test the model's own claims. Public, and runs from the XonTools app; for exploring the model, not a supported tool.