Research goals
What we need to understand next
We are preparing a study of learning from verified experience and retaining earlier abilities across repeated updates.
Learning beyond stored examples
The proposed follow-up asks whether training on verified experience adds a measurable benefit beyond access to stored examples, while preserving earlier performance after further updates and a restart.
A separate pilot of 12 task sequences will prepare a planned main study of 100 independently generated sequences. Each sequence will be tested under six conditions across three learning rounds. The proposed study runs for six months with a US$25,000 budget, including fees.
What the comparison separates
The conditions compare a frozen reference, memory alone, training with replay, training with replay and retrieval together, training with mismatched new answers, and training without replay.
Tasks cover modular arithmetic rules, finite-state systems, symbol substitutions and binary cellular automata. Fixed rules let independent reference implementations calculate exact answers.
What would count as progress
The combined procedure must improve new-input accuracy beyond memory alone and preserve initial-task performance within the study’s proposed five-percentage-point margin. Both criteria must be met, with enough initial accuracy for retention to be meaningful.
After the third learning round and a process restart, earlier abilities will be tested with memory disabled and compared with the checkpoint saved after initial training. Update-selection queries and final tests will be kept separate.
Pilot before the main study
The pilot will check task generators, reference answers, model loading, training behaviour and resource costs. It will also verify what happens to weights and optimizer state when an update is rejected. Settings will be fixed before main collection.
Harder cellular-automaton tasks may later compare final-answer prediction with working through intermediate states. Training measurements will help separate task difficulty from unstable updates; neither diagnosis should be assumed from a rule’s appearance.
Human benefit and control
Useful applications may include checking scientific calculations, finding software errors and proposing methods that people can test. Any application will need comparisons of practical value, mistakes and resource costs.
People should be able to inspect evidence, correct errors, revise an authorised task, restrict tool access and stop operation. Future studies will assess whether those interventions still work after repeated learning. These are intended outcomes, not established guarantees.
Evidence, replication and review
Follow-up results will inform a second paper and external review. We intend to release the protocol, task generators, outputs and analysis code, including unsuccessful tests and deviations.
If gains disappear under matched controls, fail on unfamiliar tasks or come with excessive losses, the learning procedure will need revision. Results that remain uncertain will be reported as uncertain.
Long-term goal
Secure and dependable infrastructure
Our long-term goal is to provide the research behind AI tools that help critical infrastructure resist attacks and remain dependable under pressure. This includes infrastructure used by governments, security agencies, organisations, data centers and space missions.
We also want to test whether computational irreducibility can strengthen cyber defenses. Some systems cannot be predicted much faster than running them step by step. Could this make certain attacks harder? Any security benefit would need to be demonstrated in separate experiments.
Loopseed
A research programme at Dilate Technologies.
© 2026 Dilate Technologies