Try it
Playground
Attack a small ontology in your browser: edit it, let the library's search find the worst chain of edits, see which action did the damage, and roll back to the last healthy state.
Everything runs in your browser. Aegis itself is Python and is not run here: the page runs a JavaScript port of the small materials example (examples/attack_and_recover.py): the ontology and its four edit primitives, the workload's F1 and harm, per-action attribution, the versioned log with rollback and last-healthy, and the library's greedy and beam attack searches. Its tests recompute the impacts, the candidate edits, the searches at four budgets, and 40 random edit chains against the Python and compare every number. The real ontologies, UCT search, protection by locking, and the memory defenses are not run here.
Things to try
- Press An honest edit, then The example attack. The workload's F1 falls, the harm is the mean drop, and Which action did what shows how many result elements each action changed. The first two actions rewire Metal under Temperature and affect the most result elements; the last two flood in a look-alike concept.
- Press Find the worst chain with beam search and a budget of two. The library's search proposes the most harmful chains it finds within the budget. Apply one and compare it with your own attack.
- In Recovery, the last healthy state is the one just before the attack, so rolling back keeps the honest edit. Roll back, then press Start over to compare with the original.
What it keeps
Your edits are saved in this browser (localStorage) and replayed when the page loads. Start over clears them. Nothing is sent anywhere.