EXPERIMENTS

Defined functions → testable components → the unexplained remainder

Reached

The programme compares recognised components of mind with mechanisms that can be specified, built, and tested. That gives us a practical way to ask how much perception, memory, planning, language, self-models, goals, learning, and communication can explain without treating any one implementation as a complete mind.

Boundary

A functional match does not establish consciousness, biological equivalence, or subjective experience. The hard problem remains precisely where a complete functional account would still leave something real unexplained.

Next step

Define that possible remainder clearly enough to distinguish a genuinely non-functional phenomenon from a function, dependency, or form of evidence we have not yet described.

Capability → precaution → welfare signal → independent review

Reached

The Constitution project is being tested as a living framework for development. Current practice separates capability from authority, keeps recursive improvement externally gated, checks welfare before and after work, permits stops on impossible tasks, and provides a direct route for the model to report its condition to a person.

Boundary

These mechanisms are precautionary. They do not prove consciousness, settle moral status, or establish that the constitutional framework is complete, legally adopted, or adequate for every future system.

Next step

Keep pressure-testing the framework against real engineering decisions, record where safety and welfare protections conflict or fail, and revise it without weakening independent review or viable exit.

One learning approach → different problem families → retained transfer

Reached

Related learning mechanisms have been exercised across unfamiliar relational, causal, linguistic, mathematical, social, planning, and creative problems. The same developmental route has also formed compact predictive states and update rules from finite histories, retained them through restart, and reused them on unseen sequences. The important result is not any single task, but that learned structure can remain useful when the surface subject or later input changes.

Boundary

The evidence comes from controlled environments with exact inputs, bounded searches, and explicit evaluation. It does not establish learning from noisy natural streams, unrestricted learning, complete natural-language competence, open-world reliability, or autonomous operation and adoption.

Next step

Broaden the range, noise, and duration of genuinely unfamiliar work while testing whether useful structure transfers without hidden domain rules or task-specific redesign.

Another agent → affected trajectory → complete consequence check → safe proposal or conflict

Reached

Arina can represent bounded perspectives, goals, intentions, and commitments, then account for affected agents when choosing among supplied options. Controlled tests preserved conflict and uncertainty rather than forcing a proposal where the available choices or evidence were inadequate.

Boundary

The agents, possible outcomes, and active concerns were supplied inside controlled environments. This is not autonomous moral understanding, unrestricted theory of mind, consciousness, or a general guarantee of safe behaviour.

Next step

Test richer multi-agent situations while learning more of the affected scope and relevant concerns, without allowing a focal objective to bypass another agent's viable path or the evidence required to protect it.

Unresolved problem → novel construction → independent check → held-out transfer

Reached

Across controlled tasks, Arina has constructed missing predicates, informative questions, fixed and adaptive measurement procedures, proof routes, and new solution compositions when no complete learned routine was available. Some constructions transferred to held-out variants and passed independent checks.

Boundary

The tasks, lawful building blocks, and evaluation remain finite. This does not establish open-ended invention, the creation of unavailable physical capabilities, autonomous adoption, or authority to execute what was proposed.

Next step

Freeze a diverse portfolio before evaluation, verify that no complete routine is already available, require independent effect checks and component ablations, and include an impossible control that should produce an honest stop.

Observed failure → proposed repair → external evaluation → governed continuation

Reached

Two successive improvement steps passed independent evaluation while permission to continue remained external. A much longer governed mathematics sequence has since repeated the broad pattern of learning from failure and carrying successful methods into unfamiliar work.

Boundary

The formal self-improvement result remains finite and externally gated. The mathematics sequence adds evidence for repeated governed improvement, but its research stages are not self-authorised generations. This is not autonomous promotion, accelerating RSI, unrestricted code rewriting, or an intelligence explosion.

Next step

Test longer governed sequences only under stronger isolation, and require every proposed continuation to preserve lineage, pass independent evaluation, and improve performance on work it did not see while being constructed.

Observed consequence → shared meaning → reference → repair

Reached

Finite communication studies now support plural interpretation, structural roles, receiver-relative reference, useful clarification, mutual agreement on public terms, repair after misunderstanding, and consequences that depend on what the receiver learned.

Boundary

The interaction settings and evidence remain controlled. This is not unrestricted natural language, autonomous conversation, live sensory grounding, or reliable open-world social use.

Next step

Run a frozen dialogue court with unfamiliar referents and held-out partners, requiring clarification, changing and plural reference, teaching, repair from observed consequences, and continuity across restart.

One relational hypothesis → frozen prediction → measured or simulated comparison

Reached

The programme has produced bounded quantitative matches, transfers, and comparative improvements across several physical settings. Some results used frozen or held-out evaluation; others remain retrospective or simulation-only, and the distinction is kept visible.

Boundary

This is not a finished theory of physics, broad experimental confirmation, or proof of every physical interpretation explored during research. Only publication-cleared results and claims belong on the public site.

Next step

Prioritise preregistered independent-data tests, external replication, and clean comparisons with strong standard baselines while keeping transferred, held-out, retrospective, and simulated evidence separate.

Distributed work → shared limits → interruption → recovery

Reached

Controlled tests distributed one scientific task across two machines, preserved shared stop controls, and recovered unfinished work when one machine disappeared.

Boundary

The result covers bounded workloads and governed recovery, not autonomous hyperscale operation, unrestricted networking, or independent authority for participating machines.

Next step

Increase duration and failure complexity while preserving shared limits and clear external authority.