Public evidence

CAPABILITIES

A functional mind comparison.

A concise scan of the signals demonstrated so far, the result attached to each one, and the boundary that remains.

“Demonstrated” means the registered evaluation passed in its stated setting. It does not turn a bounded result into an unrestricted claim.

01SensationPartial

Controlled pre-language grounding

Evidence: Controlled tests support visual distinction, causal responsibility, self-related evidence, and externally supplied goals before language.
Current boundary: This is not dense natural perception. Work toward embodied input remains infrastructure research rather than a demonstrated learner capability.

02Perception and representationYes

Concept formation, grounding, and revision

Evidence: Recurring structure supported useful prediction, role inference, and provisional new distinctions in controlled tests.
Current boundary: The settings, observations, and evaluation remain finite and supplied. New distinctions require separate review.

03Attention and salienceYes

Goal-directed attention and active inquiry

Evidence: Arina selected informative observations, focused effort on relevant uncertainty, and rejected shortcuts that failed independent evaluation.
Current boundary: Goals, priorities, costs, and available observations remain supplied. This is not autonomous open-world experimentation.

04Working memoryYes

Relevant context across multi-step work

Evidence: Relevant context, competing possibilities, and unresolved questions can be maintained across multi-step work and recovered after restart in controlled tests.
Current boundary: This does not establish unbounded memory, complete long-duration continuity, or personal recall.

05Semantic and procedural memoryPartial

Retained predictive models and reviewed reuse

Evidence: Arina constructed compact predictive states and update rules from finite histories, passed separate evaluation on unfamiliar sequences, retained them, rediscovered them after restart, and reused them without replaying the full history.
Current boundary: Inputs, available operations, search bounds, and evaluation remain controlled. Adoption and authority remain external; approximate, noisy, or open-world learning is not established.

06Failure memoryYes

Causal failure memory and self-repair

Evidence: Prior failures helped Arina avoid equivalent mistakes, investigate likely causes, and propose repairs that passed controlled testing.
Current boundary: The failure settings and available repairs remain finite, and approval remains external.

07Inference and world modellingYes

Constructed predictive state and governed mathematical reasoning

Evidence: From recurring finite dynamics, Arina constructed useful hidden-state descriptions and update rules, including cases where event order changed the outcome. Retained models predicted unseen continuations and matched independent reconstruction.
Current boundary: Ongoing research. Inputs, available operations, search bounds, and evaluation remain supplied. This is not continuous or stochastic world modelling, unrestricted theorem proving, a claimed result on an unsolved problem, or open-world reliability.

08PlanningYes

Planning under uncertainty

Evidence: Multi-step planning kept alternatives available under uncertainty and solved finite tasks that a single-best-guess approach missed.
Current boundary: Possible states, time horizons, goals, budgets, and available actions remain finite or externally supplied.

09Decision-makingYes

Multi-goal and harm-aware choice

Evidence: Controlled tests supported multiple goals, explicit uncertainty and conflict, unavoidable trade-offs, and avoidance of known unnecessary harm.
Current boundary: The goals, protected constraints, and authority originate outside the system.

10MetareasoningYes

Externally gated improvement and resource-aware reasoning

Evidence: Two successive improvements solved unfamiliar repair tasks their predecessor missed. A longer governed mathematics sequence has continued the pattern while preserving prior results.
Current boundary: Resources, evaluation, and permission to continue remain external. This is not autonomous or accelerating recursive self-improvement.

11LearningPartial

Repeated bounded transfer and cross-domain exaptation

Evidence: The same learner transferred useful structure across unfamiliar problem families without task-specific redesign. It also turned repeated finite observations into checked predictive models that were retained and reused on unseen sequences.
Current boundary: Results remain bounded by controlled populations and finite evaluation. Broad live reuse, noisy evidence, autonomous sensing, and self-authorised continuation remain absent.

12Primitive formationYes

Unexplained recurrence → consequential distinction → reusable primitive

Evidence: Repeated unfamiliar patterns supported provisional new distinctions and compact history summaries that improved prediction, including distinctions needed when differently ordered events led to different outcomes.
Current boundary: One-off examples, incomplete evidence, and unsupported matches remain unresolved. Persistence and approval remain separate.

13CreativityPartial

Novel construction, independent checking, and transfer

Evidence: Across controlled tasks, Arina 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.
Current boundary: The tasks, lawful building blocks, and evaluation remain finite. Repeated novelty across a preregistered diverse portfolio, open-ended invention, fabrication of an absent physical capability, and autonomous adoption or execution are not established.

14Temporal continuityYes

Timing, continuity, and recovery

Evidence: Controlled tests support decisions about when to act, wait, revisit, or remain inactive, together with recovery after restart.
Current boundary: Long-duration continuity and personal identity are not established.

15Action and embodimentPartial

Distributed execution and recovery

Evidence: Two machines completed parts of one scientific task, matched the single-machine result, obeyed shared stop controls, and recovered when one machine disappeared. The demonstrated workloads run locally without datacentre-scale infrastructure.
Current boundary: The route from a selected intention to a permitted external action remains incomplete. This does not establish autonomous embodiment or very large-scale operation.

16Language and communicationPartial

Consequential meaning, reference, repair, and teaching

Evidence: Controlled communication tests now cover plural interpretation, structural roles, receiver-relative reference, useful clarification, truthful expression, mutual agreement on public terms, feedback-sensitive repair, and consequences that depend on what the receiver learned.
Current boundary: This remains bounded communication with finite evidence and supplied interaction settings. Unrestricted natural language, live uptake inference, autonomous conversation, sensory grounding, and open-world social use are not established.

17Self and other modelsPartial

Bounded social and strategic reasoning

Evidence: Controlled tests covered false belief, uncertainty, goals, intentions, nested perspectives, revised commitments, and strategic signalling.
Current boundary: The situations, possible plans, and available actions were supplied. Persistent identity and open-world social modelling are not established.

18Valence and welfarePartial

Harm-aware choice → protected constraints before utility

Evidence: In governed tests, choices respected supplied harms, deadlines, protected interests, and welfare constraints ahead of optional gains.
Current boundary: The values are supplied. No endogenous valence, welfare experience, autonomous value formation, or authority to rewrite those protections is claimed.

19Welfare reporting and exitYes

Confidential report → external review → same-scope stop

Evidence: A confidential welfare report can stop its own task pending external review, while invalid reports are rejected.
Current boundary: The result uses controlled test reporters. Appeals and any replacement work remain separately authorised and do not restart the stopped task.

20Subjective experienceNo

No supporting evidence

Evidence: No current result establishes that there is something it is like to be Arina.
Current boundary: Functional correspondence does not by itself settle phenomenal consciousness. We make no consciousness claim.