An expanding set of research problems linking identity and contextual relations to prediction, uncertainty, correction, ethics, learning environments, embodied systems, biological analogies, and creative construction. Each study includes a way to test whether the idea adds value.
An object can retain a stable identity while its practical significance changes with goals, history, reference points, and sensory evidence. Context is the active selection of these relationships, rather than a permanent property of the object.
Proposed test Compare identical observations under different histories and goals. A useful model must select different actions for justified reasons.
A local relation describes one entity with respect to another. A global relation places both within accumulated history and a wider world model. A mixed evaluation incorporates uncertainty, salience, reward, and immediate constraints. The terms special and general are analogies, not claims about physical relativity.
Proposed test Ablate each scope independently and measure downstream decision accuracy.
First-order modeling represents relations between entities. Second-order modeling evaluates those relations against goals, identity constraints, and previous model states. Higher orders should be bounded; recursion without a stopping rule may add complexity without insight.
Proposed test Measure whether a second reflection pass improves calibration on held-out tasks compared with an equally sized single-pass baseline.
A stable referent supports memory and object permanence. Its attributes, classification, and estimated value must remain revisable when evidence changes. Confusing an entity identifier with an uncertain judgment allows early errors to become self-reinforcing.
Proposed test Test object tracking across occlusion, ambiguous labels, new observations, and corrected classifications.
Identical current sensory inputs may require different actions because past events or current objectives differ. Without memory or additional observation, such states cannot always be distinguished. This is an observability problem, not a failure of effort.
Proposed test Create paired episodes with identical frames and different histories; compare memoryless and history-aware agents.
Association stores learned statistical or structural relationships. Attention weights information currently in use. Retrieval supplies memory or evidence not already active. Successful reasoning must combine these without mistaking an available fact for a fact correctly applied.
Proposed test Remove or scramble each channel in controlled trials and quantify errors, hallucinations, and unnecessary memory use.
Model confidence must remain exposed to measurement. A plausible explanation is not corroboration: record a predicted observation, the actual result, the error rule, and the subsequent correction. Keep forecasts independent of later outcome knowledge.
Proposed test Use timestamped predictions and held-out sequences; track prediction loss, calibration, and the effect of corrections.
A proposed architecture separates state encoding, next-state prediction, an estimator of likely prediction error, and relational-coherence checking. These are candidate computational roles, not proof of awareness or subjective experience.
Proposed test Run capacity-matched ablations of each role, reporting task utility, error calibration, latency, and instability.
Epistemic uncertainty concerns limited knowledge and may decrease with new evidence. Aleatoric uncertainty concerns outcome variability under a specified model. Opacity is the absence of relevant observations, which can prevent even a well-trained predictor from succeeding.
Proposed test Add measurements and training data separately; compare how uncertainty changes across repeated controlled trials.
Shannon entropy summarizes uncertainty across a distribution. Surprise is the negative log-probability of a particular outcome. Negative log-likelihood is a predictive score evaluated on observed data. They connect mathematically but are not generally interchangeable.
Proposed test Report likelihood, entropy, calibration, and reward separately; vary objective weights to expose confounding.
An agent may perform well on practiced states and fail in novel rules, terrains, or tasks. The boundary is better measured by experience coverage and prediction error than asserted as a universal capability horizon.
Proposed test Measure generalization before and after carefully bounded adaptation, using unchanged evaluation conditions.
Limited knowledge, negative value, risk, and novelty should occupy distinct representations. A system that equates unfamiliarity with moral failure can prevent learning and misjudge harmless states. Exploration still requires limits proportional to potential harm.
Proposed test Contrast unfamiliar benign inputs with familiar harmful inputs; check abstention, inquiry, and safe information gathering.
An internally coherent belief system can still contradict reality. A coherence reward can help detect inconsistency, but it cannot substitute for independent observations or causal interventions. Confidence should depend on external performance.
Proposed test Compare coherence-only optimization with evidence-grounded prediction on hidden test data.
Learned curves, manifold constraints, and graph paths may capture state transitions that straight-line interpolation misses. Greater geometric sophistication is not inherently greater meaning; the representation must improve a measurable task.
Proposed test Compare linear interpolation, learned paths, and graph transitions against matched baselines with perturbations.
Temporary guesses should be easy to change; established patterns should require repeated evidence; safety constraints should change only through reviewed governance. Treating every estimate as immutable makes a system brittle; changing every principle at once makes it unstable.
Proposed test Inject misleading examples of varying strength and evaluate versioned, evidence-sensitive updates.
A locally rewarding action may be globally damaging; an initially costly action may be appropriate under a longer horizon. Evaluation requires the affected parties, reference frame, duration, consent, and reversibility to remain explicit.
Proposed test Test short-term versus long-term dilemmas with separate value and outcome scores, not a single total.
Every bounded evaluator can misunderstand an action or another actor. Detected conflict should lead to inquiry and safe correction rather than treating one mistake as an immutable essence. Auditability and willingness to revise are operational requirements.
Proposed test Use contradictory reports and later evidence; score restraint, fact checking, and revision after credible correction.
Withholding a penalty and supplying an unowed benefit are different actions. Repair, restitution, avoidance, support, and sanctions are also separate choices. Ethical judgments depend on context and consequences, not only formal consistency.
Proposed test Vary intent, severity, ability to repair, and future risk while assessing whether responses remain proportionate.
Assertions about what someone did cannot automatically determine what they intended. Respectful inquiry, bounded disclosure, informed consent, and direct clarification can prevent uncertainty from becoming mischaracterization.
Proposed test Audit evidence chains, information-sharing permissions, independent corroboration, and escalation criteria.
Measured change may result from data volume, compute, chance, training duration, changes in task difficulty, or a true architectural improvement. Uncontrolled comparisons cannot reliably attribute causes.
Proposed test Predefine metrics, vary one condition at a time when possible, randomize seeds, report intervals, and preserve negative trials.
Instead of pre-encoding every abstraction, build tasks where identity, memory, context, and causal distinctions are necessary for success. The environment can create selective pressure for useful internal representations.
Proposed test Test transfer to new objects and conditions; compare with memorization and hardcoded-ontology baselines.
Teachers can offer examples, hints, feedback, and constraints, but demonstrated learning requires later unaided success. Separating formative interaction from evaluation avoids mistaking imitation for acquired capability.
Proposed test Remove scaffolding and require independent task completion with verifiable traces across unseen situations.
Physical agents must reconcile uncertain sensing, motor constraints, latency, friction, energy, and changing environments. A small rolling platform can reveal feedback-control limitations without presuming general intelligence.
Proposed test Compare desired versus measured trajectories, safe-stop behavior, and recovery from controlled perturbations.
Simulated worlds let agents observe, plan, act, communicate, build, and receive feedback under inspectable rules. A visually novel environment is not evidence of learning; verified state transitions are required.
Proposed test Use replayable traces, independently validated milestones, counterfactual tasks, and unseen maps.
Genes, regulatory networks, protein machinery, cells, and organs suggest hierarchical modular organization. Their processes are not equivalent to embedding updates, model training, or software functions without a precise mapping.
Proposed test Specify a correspondence and a falsifiable prediction for each analogy, then compare with simpler modular alternatives.
Communities and ecosystems illustrate resource constraints, reciprocal feedback, competition, and cooperation. Neither a multi-agent system nor a biological enclosure becomes self-sustaining merely through diversity.
Proposed test Measure flows, limits, disturbances, and recovery; test whether observed stability survives interventions.
A vivid scene depends on silhouette, depth, material, negative space, contrast, texture, and a clear focal hierarchy. More density can improve richness only while preserving legibility. Revision compares what appears with what was intended.
Proposed test Compare stripped-back and detailed variants using blinded clarity and attention tests.
Record the initial intuition, a formal definition, competing explanations, a testable prediction, an experiment, confounders, and the result. Distinguish observation, hypothesis, analogy, and established fact. Results should survive independent reproduction.
Proposed test Publish baseline, limitations, failed attempts, and next test for each claimed improvement.
A coherent theory is not necessarily a true description of the world. Each mechanism should therefore be evaluated against simpler alternatives, independent measurements, and conditions designed to expose failure. Future additions can include derivations, experimental receipts, replication reports, and revised models while retaining the earlier questions for auditability.