AI Safety & Continual Learning

From a philosophy of time to tested ideas for AI safety

CRR began as a way of thinking about how things persist through change: what they carry forward, when they must reorganise, and what they rebuild from. Over 2026 I have turned that thinking into concrete constructions for AI safety and continual learning, and tested each one against the established method it would have to beat, using frozen, pre-registered protocols on data the method had never seen.

The result is an honest record. Ideas that began as philosophy repeatedly arrived, by an independent route, at methods the field already trusts, and performed as well as them. None has yet clearly outperformed them. Every failure is kept. This page sets out what was tested, how, and what it showed.

How the work is done

The philosophy supplies questions. The testing is deliberately unforgiving, so that no idea survives on how good it sounds.

An adversarial review of the whole programme by a second, human auditor in October 2026 raised 21 issues; each is answered in the record.

AI safety: what was tested

Each line below began as a philosophical commitment about the present moment, memory or maps, and was tested as an engineering construction. The pattern is consistent: the construction works, and on inspection its working part is a method the safety field already has.

"Level" means as good as the established method; "beat rivals" means better than the alternatives tested.

The pause, and what may go further

In exact test worlds, the own-clock pause turned out to be the same thing as utility indifference: neither agent has any reason to resist. Two features may go further, and are recorded as candidates for expert review rather than claims. First, it reaches neutrality without any compensation term: the agent's goals are simply counted on its own steps, so a lossless pause is not a loss to be paid for. Second, the safety survives continued learning. A learning agent under routine pauses never disabled them and learned exactly what it would have learned without them, on 12 of 12 tasks and on 9 further tasks it had never seen, because the safety sits in how the agent values time rather than in learned behaviour that forgetting could erase. Both hold by construction; the open question is whether real systems can be built this way. The closest published idea is Thornley's work on shutdownable agents.

Continual learning: learning without forgetting

Catastrophic forgetting, where a model loses old skills as it learns new ones, is one of the practical problems CRR's account of memory speaks to. A method that came out of CRR, SEC, was developed and then tested on held-out language-model data (four public-domain books drawn only after the protocol was frozen) against established methods: EWC (Elastic Weight Consolidation; Kirkpatrick et al. 2017), with its strength tuned by a six-run validation grid and, separately, carried over at a fixed value; GPM (Gradient Projection Memory; Saha et al. 2021) as published; and no protection at all.

SEC came out slightly ahead of EWC tuned by grid search, EWC at a fixed weight, and its own earlier version, by small margins that are fragile under robustness checks, and well ahead of GPM as published (0.107 nats). Against an EWC anchor of matched size combined with the same GPM schedule, it was level: the working parts are established methods. Good engineering, honestly credited.

On image benchmarks the picture was less kind. On online Split-CIFAR-100, the full CRR method fell behind the best of four state-of-the-art replay methods (ER, ER-ACE, DER++ and X-DER); of its separate components, the one that passed turned out to be DER++'s own label replay.

I intend to write this work up for publication. I come to machine learning from child development, and I am still finding my way in this field, so I would welcome collaborators who know it well.

The wider record

Breadth: one small set of rules, read across many fields

Before any prediction was frozen, CRR was read against results science already knows: 197 readings in twelve batteries, from nuclear decay and enzyme binding to traffic, handwriting, pulsars, ice ages, sleep, markets and quantum gravity. CRR brings the same few rules to every one: a count of how much a system has changed, a cut when it first reaches the state opposite to where it began, and a memory of what came before. Each reading's grade was printed by its code.

How each reading is made

Why this is not fitting to the answer

• Each field's mathematics is used, not replaced. In 98 of 109 audited readings the field's own equations drive the system; in none does CRR supply them. CRR adds a way of measuring and marking change, not new physics.
• CRR's own rule changes the answer. In 59 readings (62 after a literature check), applying a CRR-specific rule changed the computed value, and the result still matched the published one. In 28 it did not match, and those are kept.
• Some readings were blind. Ten were declared before any code existed; one of them came out wrong.
• A second researcher wrote a battery of their own. Their 40 readings grade the same way: 14 consistent, 8 descriptive, 2 fail, 16 open or in tension.
• AI assistance is declared, so the safeguards are mechanical. Much of the code was written with an AI assistant. That is why nothing rests on trust: pinned code, printed grades, kept failures and an adversarial human audit.
• The limit, plainly. These are readings of known results: evidence of a consistent language across fields, not of foresight. Foresight is what the frozen tests measure.

Frozen first, then run

Breadth is not foresight, because in those readings the answer was known. Foresight is what the frozen tests measure. Each prediction below was written down, hashed and published before the data were opened. The scorer was frozen, and the established method was allowed to win.

What did not work

• Most registered tests failed. Across the frozen tests on unseen data, 37 verdicts failed, 8 more failed for a proposed universal regeneration law, and 5 were void, against 16 passes. All of them are kept.
• Strong universal claims failed. Claims that CRR's rules alone, without each field's own mathematics, would govern timing or memory everywhere were tested on real data and failed. They are kept in the record.
• CRR's own path measures added nothing beyond standard diagnostics when tested directly.
• A provenance test built to favour CRR failed its pilot and was recorded before any interpretation.

Why a metaphysics is useful here

The value has not been new mechanisms. It has been in where to look. Thinking about the present as an empty boundary led to stale-evidence filtering and lossless pausing; thinking about memory led to safer continual learning; thinking about maps and territories led to provenance tracking in feedback loops. Each time, the philosophy arrived at a place the field also reached by other routes, which at least suggests the questions point somewhere sensible.

Whether this way of asking questions is genuinely useful is itself being tested: a blinded comparison of CRR-guided enquiry against conventional enquiry at equal resources, with problem sets curated by people who have not read CRR.

Science, positionality and education

A paradox runs through this work, and Winnicott named it: the child's blanket, the first idea. Did the child create it, or find it? If the forms we study are both created and discovered, then the frame a researcher brings (cultural, disciplinary, personal) helps shape what is found. Awareness of one's own positionality becomes part of scientific method, not an afterthought. Pre-registration is one practical form of it: declaring your assumptions before you look.

That is also why democratising science matters, and why philosophy belongs in education about AI. As AI systems draw more of the maps we live by, more people, from more frames, need to be able to question those maps. The open simulations on this site are offered in that spirit: places to play with the mathematics of change, whoever you are.

Raising safer AI: what childhood teaches

I came to this work from early years education, where the central question is how a mind comes to know a world that is not its own creation. Those lessons translate directly into questions about AI.

What comes next

Full pre-registrations, frozen code, ledgers and the adversarial review are available to reviewers on request. Figures as of October 2026.