Cosmogonic Quantum Mechalogodrom · canonical project standard

CQM-25: the project capability standard.

A research-informed profile for artificial life, embodied synthetic cognition and procedural morphogenesis, spanning the organization of one being, adaptive behavior, consciousness-related mechanisms and the engineering of inspectable worlds.

Version 1.0.0 · adopted 2026-10-22. These 25 operational dimensions govern project capability plans and claims across individual beings, interacting populations and constructed worlds.

Canonical project authority; explicit evidence limits

CQM-25 is the binding project standard, not a consensus certification or an executed 25-task benchmark. Acceptance experiments remain proposed and unexecuted as a unified benchmark. Review states describe available repository evidence. Numerical capability scores remain uncalibrated.

Feature plans and capability claims must identify the affected stable axis IDs and their evidence requirements. Taxonomy changes require an explicit framework version and contract review.

Engineering breadth and visual expression do not count as evidence of consciousness, sentience, AGI or superintelligence. No aggregate verdict is calculated.

The capability law

Adopted 2026-10-22 as the binding capability framework of this repository (ADR 0018). It governs claims made here; it is not an external consensus standard.

CQM-25 is the canonical capability framework of this repository. Every claim about life-like organization, cognition, consciousness-linked mechanisms, valence-related control or engineering scale, in documentation, user interface text, commits and agent output, is made against it.
A capability claim names its CQM-25 axis and that axis's current evidence state, and claims no more than that state supports.
An evidence state changes only with repository evidence that exists at the pinned commit and a passing test. Raising a state follows the axis's acceptance experiment and survives its falsifier.
Numeric capability scores stay null until a calibrated successor edition executes the axis protocols with controls, independent experimental units and reported rater agreement.
The nine-axis, 155-system peer survey is a legacy comparison view. It keeps its own rubric and scores, is crosswalked to CQM-25 at the level of constructs only, and never supplies or receives CQM-25 scores. Construct correspondence does not imply that the legacy survey measured or validated CQM-25 acceptance protocols.
No aggregate life, consciousness, sentience, AGI or superintelligence score is derived from the axes.
Separately labeled provisional author estimates may use the declared uncalibrated heuristic rubric in cqm-capability-estimates.json. They remain pending-review proposals, never populate canonical assessment.score and supply no aggregate or domain mean.

Scope of the standard

Constructed physics is evaluated against declared rules and constraints. Ecology is one domain among several. An entity-level study can declare reproduction or population ecology not applicable, with a reason; a missing measurement remains unknown.

Historical nine-axis comparison and visual review

The existing visual-scale rating remains 4.0 (rich visuals). A single complex being qualifies: Xylos exposes neural activity through deforming membranes and a visible graph, while apex forms and populations add other scales. Population capacity, delivered frame time and individual visual expressiveness are distinct measurements. The rubric's 5.0 anchor, iconic spectacle, has no independent calibration that justifies a numerical uplift in this review.

The original nine-axis, 155-system comparison is historical secondary context. It retains its own rubric and scores; its rankings and averages are not transferred into the canonical CQM-25 standard.

Historical nine-axis comparison (SVG)

Historical construct crosswalk: 19 of the 25 axes

From the nine-axis survey to CQM-25

The legacy survey grades nine broad constructs. This table gives conceptual correspondence to 19 of the 25 CQM-25 axes; it does not mean those protocols were measured or validated. No score or validation transfers in either direction.

Legacy axis	CQM-25 axes	Relation
Reproduction	01	Same broad construct. The legacy rubric grades heredity, mutation and selection; CQM-25 adds pedigree tracking and a randomized-parentage control.
Open-endedness	02	Same broad construct. CQM-25 adds neutral-drift and frozen-heredity controls, retention of new functions and finite-horizon novelty trajectories.
Ecology	03	The legacy rubric grades multi-species interaction and food webs. CQM-25 generalizes it to bidirectional entity-environment coupling, for which a single being is eligible.
Morphology/Physics	04–05	The legacy axis merges body and physics. CQM-25 separates developmental morphology from consistency with declared dynamics and does not require Earth-like realism.
Cognition/Learning	06–12	The legacy rubric is one ladder from reactive control to deep reinforcement learning. CQM-25 defines seven cognitive functions separately.
Substrate pluralism	23	Same construct, reported in CQM-25 as engineering context rather than capability.
Instrumentation	24	Same broad construct. CQM-25 adds preregistration, independent units, archived inputs and replication.
Consciousness-theory	13, 19–21	The legacy rubric grades whether a computational theory of consciousness is implemented. CQM-25 separates theory-linked mechanisms and treats none of them as evidence of experience.
Visual scale	25	Same broad construct. CQM-25 measures single-being expression, population capacity and delivered frame time separately.

Not separately represented in the nine-axis survey: 14 Social cognition and communication · 15 Individuality and maintained boundary · 16 Resource-dependent self-maintenance · 17 Homeostatic regulation · 18 Repair and functional resilience · 22 Valence-related control and motivational trade-offs.

Showing all 25 axes. Scope tags indicate where a protocol can apply, not proof that it has passed.

01life · Life, evolution and embodiment

Reproduction and heredity

Production of descendants with causally inherited, variable organization.

reproduction_heredityentity / populationBounded test evidenceLegacy construct correspondence: Reproduction

Genome crossover and mutation are implemented, and the shipped entity fission test compares inherited-trait selection with a trait-blind control. This supports bounded heredity and fecundity selection; autonomous open-ended reproduction and the new pedigree protocol remain unassessed.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

viable offspring per parent
parent-offspring trait transmission
heritable variation under matched conditions

Acceptance experiment

Track pedigrees across generations; intervene on inherited state, randomize parent-offspring assignments as a control, and distinguish autonomous reproduction from evaluator cloning.

What would weaken the claim?

Offspring are injected by the harness, inherited variables do not affect descendants, or transmission vanishes after environment matching.

Repository evidence

Genome representation, crossover and mutation src/sim/genome.ts
Deterministic inheritance and variation tests tests/genome.test.ts
Production fission versus trait-blind selection control tests/entity-repro-select.test.ts

Scientific basis: avida2003 · bedau2000

02life · Life, evolution and embodiment

Evolutionary innovation and open-endedness

Selection produces retained adaptive novelty across generations; open-endedness remains a finite-horizon research claim.

evolutionary_innovationpopulationMixed or restricted evidenceLegacy construct correspondence: Open-endedness

Existing entity and Petri tests exercise differential selection and strain-specific fitness effects. They do not establish retained new functions, superiority to drift across novel tasks, or sustained open-ended innovation; finite forms and scripted progression must not be counted as such evidence.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

fitness gain over drift control
functional novelty and retention
MODES trajectories with horizon

Acceptance experiment

Run mutation/selection, neutral-drift and frozen-heredity populations under matched budgets; assess new functions on held-out challenges and plot novelty saturation over increasing horizons.

What would weaken the claim?

Novelty is only random visual variation, improvement equals drift, or a fixed catalogue is repeatedly rediscovered.

Repository evidence

Bounded fecundity-selection contrast tests/entity-repro-select.test.ts
Strain-dependent selection and ablation tests/petri-organism-intelligence-selection.test.ts
Current open-endedness and reproduction limits docs/VERIFICATION-ANALYTICAL-DATA.md

Scientific basis: modes2019 · avida2003

03life · Life, evolution and embodiment

Environmental and ecological coupling

Bidirectional functional dependence between an entity and its surroundings; multi-species ecology is an optional extension.

environmental_couplingentity / population / worldBounded test evidenceLegacy construct correspondence: Ecology

Xylos senses local Gray-Scott chemistry and changes that field through gated actions; tests check subsequent perception and compare a closed loop with uniform inputs. This supports a single-entity causal loop, not the full new perturbation protocol or long-horizon ecological persistence.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

effect of world perturbation on behavior
effect of entity intervention on world state
niche or interaction persistence when applicable

Acceptance experiment

Perturb resource locations, environmental rules and entity actions independently; compare a live world with a recorded or shuffled environment. A single entity in a responsive environment is eligible.

What would weaken the claim?

The environment is decorative, entity actions cannot change relevant state, or coupling disappears when shared scripted timing is removed.

Repository evidence

Local chemical sensing and field writes src/sim/xylos-habitat.ts
Action-to-environment-to-perception controls tests/xylos-habitat.test.ts

Scientific basis: modes2019 · animalai2019 · meltingpot2021

04life · Life, evolution and embodiment

Morphology and development

Structured embodiment develops or changes through local rules and interactions, with measurable functional consequences.

morphology_developmententityMixed or restricted evidenceLegacy construct correspondence: Morphology/Physics

Procedural bodies change with internal state, evolution parameters and deterministic morph transitions. Existing body tests demonstrate state-dependent geometry and rendering behavior; they do not establish developmental growth with independently measured functional benefits under body lesions or changed boundary conditions.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

developmental reproducibility
structural organization
body perturbation effect on function

Acceptance experiment

Grow from altered seeds and boundary conditions, quantify topology and task performance, and compare live development with a fixed or replayed shape.

What would weaken the claim?

Apparent growth is a fixed animation or morphology changes have no tested functional consequence.

Repository evidence

Neural-state-dependent procedural embodiment src/sim/xylos-body.ts
Geometry and graph response tests tests/xylos-body.test.ts
Mind and evolution parameters drive body presentation tests/super-body-control.test.ts

Scientific basis: nca2020 · lenia2019 · evogym2021

05engineering · Life, evolution and embodiment

Embodied dynamics and internal consistency

Action, body and environment follow declared, consistent update laws. Earth-like physical realism is not required.

embodied_dynamicsentity / worldBounded test evidenceLegacy construct correspondence: Morphology/Physics

Xylos uses fixed-step cognitive updates and bounded motion, with tests for timestep partitioning, input validation and finite state. These are local consistency checks, not a convergence or invariant audit of every world subsystem; Earth-like physical realism is not required by this axis.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

timestep or resolution convergence
declared invariant violations
action-body-environment causal response

Acceptance experiment

State the simulated laws, units and tolerances; vary timestep/resolution and perturb actuators, body parameters and boundary conditions. Test only invariants actually claimed by the model.

What would weaken the claim?

Behavior depends on rendering framerate, undeclared teleports or impossible state transitions; claimed invariants fail outside tolerance.

Repository evidence

Fixed-step integration and bounded catch-up src/core/xylos-session.ts
Partition invariance and invalid-timestep checks tests/xylos-session.test.ts
Long-run state and topology bounds tests/xylos-mind.test.ts

Scientific basis: evogym2021 · lenia2019 · reproducibility2021

06cognition · Adaptive and general cognition

Perception and perceptual integration

Task-relevant state is inferred from available sensory input across noise, ambiguity and transformations.

perceptionentityBounded test evidenceLegacy construct correspondence: Cognition/Learning

Xylos integrates numeric chemical, hazard and gradient channels with learned reliability; existing tests compare predictable and noisy observations. These channels expose simulation measurements directly, so they do not establish visual recognition, label-free perception or robustness on the proposed held-out sensory tasks.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

held-out discrimination accuracy
noise/occlusion robustness
cross-channel integration benefit

Acceptance experiment

Use unseen stimuli, cue-conflict and sensory-ablation tests; prevent access to ground-truth world labels; compare single-channel, shuffled-input and intact conditions.

What would weaken the claim?

Success depends on privileged hidden-state access, fixed labels or nuisance cues rather than the claimed perceptual information.

Repository evidence

Numeric percepts and reliability-weighted updates src/sim/xylos-mind.ts
Direct simulation measurements supplied to the organism src/sim/xylos-habitat.ts
Reliability and continuous-input response tests tests/xylos-mind.test.ts

Scientific basis: animalai2019 · brainscore2018

07cognition · Adaptive and general cognition

Learning and plasticity

Experience causes durable, useful changes in behavior or internal mappings.

learningentityBounded test evidenceLegacy construct correspondence: Cognition/Learning

Xylos has plastic connections and online sensory/action predictors; existing tests compare learning with frozen controls and check forecast improvement in a bounded feedback environment. This establishes narrow mechanisms, not a broad learning benchmark, reversal generalization or durable improvement across task families.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

learning curve and sample efficiency
retained improvement
contingency-reversal adaptation

Acceptance experiment

Train on controlled experience and test held-out trials with plasticity enabled versus frozen; reverse contingencies and use yoked-exposure controls.

What would weaken the claim?

Performance changes are scripted or transient state drift, or remain unchanged when all learning updates are disabled.

Repository evidence

Plasticity and online predictive updates src/sim/xylos-mind.ts
Frozen-learning and action-feedback controls tests/xylos-mind.test.ts
Associative imprint, recall and decay tests tests/plastic-weights.test.ts

Scientific basis: animalai2019 · homeostasis2014 · continual2017

08cognition · Adaptive and general cognition

Memory and continual retention

Information from earlier experience affects later behavior despite delay, interference and new learning.

memoryentityBounded test evidenceLegacy construct correspondence: Cognition/Learning

Associative fast weights and holographic context-action traces store and retrieve information, with tests for recall, interference-like superposition and decay. These bounded memory mechanisms do not establish continual retention across independently held-out task sequences or the proposed delayed behavioral memory protocol.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

delay-dependent recall
interference resistance
forward/backward transfer and forgetting

Acceptance experiment

Use delayed nonmatching or sequence tasks, remove external traces, interleave new tasks, and selectively erase or scramble the proposed memory store.

What would weaken the claim?

Later performance can be explained by current inputs, leaked labels or environmental traces; memory ablation has no predicted effect.

Repository evidence

Decaying context-action binding and recall src/sim/holographic-memory.ts
Binding, noisy cleanup and context recall tests tests/holographic-memory.test.ts
Associative superposition and forgetting tests tests/plastic-weights.test.ts

Scientific basis: continual2017 · animalai2019

09cognition · Adaptive and general cognition

Predictive and causal world models

Internal models predict outcomes and discriminate the consequences of interventions, rather than only correlations.

predictive_causal_modelsentityMixed or restricted evidenceLegacy construct correspondence: Cognition/Learning

Xylos has bounded action-conditioned prediction tests. Broader evidence remains negative: the V4 adaptive ecology predictor failed both controls, and the Phase-B temporal predictor failed its advancement gate. These branch-specific results neither validate general causal modeling nor show that every predictor lacks useful ability.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

held-out prediction error
intervention prediction accuracy
counterfactual control benefit

Acceptance experiment

Change action-outcome rules and hidden causes; test intervention forecasts on unseen combinations; ablate the model while matching compute and immediate observations.

What would weaken the claim?

Only observational correlations are reproduced, intervention forecasts fail, or the claimed model never influences decisions.

Repository evidence

Bounded action-model and feedback prediction tests tests/xylos-mind.test.ts
Frozen V4 predictor failure and claim boundaries docs/VERIFICATION-ANALYTICAL-DATA.md
Sealed temporal predictor failure with eight failed criteria tests/organism-intelligence-phase-b-predictor-temporal-development.test.ts

Scientific basis: animalai2019 · butlin2023 · cognitive2026

10cognition · Adaptive and general cognition

Autonomous goal regulation

Persistent internal priorities select and revise actions without a human choosing each step.

autonomous_goalsentityImplementation identifiedLegacy construct correspondence: Cognition/Learning

Xylos selects actions without stepwise human commands using energy-dependent hunger, uncertainty-linked curiosity and learned action forecasts. Its priorities are designed control rules; extended goal-persistence, novel goal switching and matched open-loop comparisons for this new axis have not been evaluated.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

goal persistence under distraction
context-sensitive goal switching
intervention rate per successful episode

Acceptance experiment

Run extended unattended episodes, vary internal needs and external opportunities, and intervene on goal states; compare fixed-priority and open-loop policies.

What would weaken the claim?

A script or external operator selects each goal, or declared priorities have no causal effect on action.

Repository evidence

Need-weighted autonomous motor policy src/sim/xylos-mind.ts
Unprompted activity and causal action-model tests tests/xylos-mind.test.ts

Scientific basis: homeostasis2014 · agi2024 · butlin2023

11cognition · Adaptive and general cognition

Reasoning and planning

Multi-step decisions use relations or prospective consequences to achieve new goals.

reasoning_planningentityImplementation identifiedLegacy construct correspondence: Cognition/Learning

NHI computes bounded GOAP next steps and updates its fact model after material outcome acknowledgement; the apex also has a discrete active-inference policy model. These constrained planners do not establish unfamiliar relational reasoning, broad planning generalization or success on the new planning protocol.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

novel-task success
planning-horizon scaling
replanning benefit under changed constraints

Acceptance experiment

Use held-out detours, relational puzzles and changed action costs; compare reactive and planning-disabled controls with matched action/compute budgets.

What would weaken the claim?

A memorized action sequence or local reflex matches performance, or plans fail after a relevant causal change.

Repository evidence

GOAP planning, utility bias and material fact acknowledgement src/sim/nhi.ts
Discrete policy evaluation with stated approximation src/sim/active-inference.ts
Bounded material-action development harness tests/organism-intelligence-phase-b-nhi-closed-loop-development.test.ts

Scientific basis: animalai2019 · agi2024 · cognitive2026

12cognition · Adaptive and general cognition

Transfer and compositional generalization

Previously acquired capacities support unfamiliar tasks, combinations and environments under an explicit adaptation budget.

transfer_generalizationentityMixed or restricted evidenceLegacy construct correspondence: Cognition/Learning

Bounded tasks and controlled development harnesses exist, but V4 passed only the Titan diplomacy family and failed ordinary, Petri and predictor gates. Phase-B remains development-only, including a failed temporal-predictor gate. Those results do not establish broad transfer; the new held-out-family protocol is unrun.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

zero/few-shot transfer
held-out combination accuracy
performance breadth and depth by task family

Acceptance experiment

Hold out task families and combinations before development; log training exposure, adaptation compute and assistance; report disaggregated performance against matched baselines.

What would weaken the claim?

Evaluation overlaps training, extensive hidden retraining is required, or success is confined to familiar surface variants.

Repository evidence

One family pass and three failures under the frozen V4 protocol docs/reports/ORGANISM-INTELLIGENCE-V4-RESULTS-2026-07-11.md
Failed successor development gate tests/organism-intelligence-phase-b-predictor-temporal-development.test.ts
Narrow semantic development evidence and forbidden broader claims tests/organism-intelligence-phase-b-nhi-closed-loop-development.test.ts

Scientific basis: animalai2019 · agi2024 · cognitive2026

13cognition · Adaptive and general cognition

Metacognition and self-modeling

The system monitors uncertainty or its own limitations and uses that information to regulate cognition and action.

metacognition_self_modelentityMixed or restricted evidenceLegacy construct correspondence: Consciousness-theory

Reliability cues and Taylor forecast remainders influence confidence and decision coherence; tests check directional changes and a causal resolver effect. This supports functional uncertainty regulation, not calibrated metacognitive accuracy, a comprehensive self-model or matched first-order-performance evaluation under the proposed protocol.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

confidence calibration
meta-d-prime/d-prime when assumptions hold
adaptive opt-out or information seeking
self-perturbation prediction

Acceptance experiment

Match first-order difficulty, test confidence versus correctness, perturb sensors/actuators and test self-predictions; ablate the monitor and evaluate control consequences.

What would weaken the claim?

Confidence is a decorative label, tracks only task difficulty, or monitoring ablation leaves supposedly dependent control unchanged.

Repository evidence

Reliability-based confidence executive src/sim/metacognition.ts
Bounded confidence and cue-response tests tests/metacognition.test.ts
Forecast confidence causally affects resolver coherence tests/predictive-metacognition.test.ts

Scientific basis: metacognition2014 · butlin2023

14cognition · Adaptive and general cognition

Social cognition and communication

Behavior uses information about other agents; signals causally coordinate or inform interaction.

social_communicationentity / populationMixed or restricted evidenceNot separately represented in the nine-axis survey

Opponent beliefs and a learned rival-proximity head influence social threat/readiness; existing tests examine determinism, learning and isolated control effects. They do not establish communication semantics, unfamiliar-partner generalization or flexible social cognition across the new interaction protocol.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

novel-partner adaptation
coordination benefit
message-intervention effect
partner-state prediction

Acceptance experiment

Test unfamiliar partners and opposing incentives; scramble communication or exchange partner identities; control for environmental cues and fixed coordination scripts.

What would weaken the claim?

Signals are epiphenomenal, partners share privileged hidden state, or coordination cannot generalize beyond a fixed pairing.

Repository evidence

Bounded opponent-model mechanism src/sim/theory-of-mind.ts
Belief and context-response checks tests/theory-of-mind.test.ts
Narrow learned rival anticipation and operational controls tests/super-creature-social.test.ts

Scientific basis: meltingpot2021 · cognitive2026

15life · Individual organization and viability

Individuality and maintained boundary

A temporally identifiable organized unit maintains a functional distinction from its environment; boundaries may be distributed.

individualityentity / populationImplementation identifiedNot separately represented in the nine-axis survey

Xylos maintains a separate recurrent state, topology, position and session history while exchanging numeric inputs and effects with its habitat. This identifies an implemented computational unit; maintained identity through component turnover, selective boundary exchange and functional cohesion under partition have not been assessed.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

identity persistence through component turnover
boundary-selective exchange
functional cohesion under partition

Acceptance experiment

Declare candidate boundaries; exchange or remove components and compare within-unit versus cross-boundary interventions; track identity independently of rendered appearance.

What would weaken the claim?

Identity is only a label or mesh ID, proposed boundaries predict no functional organization, or all coherence comes from an external coordinator.

Repository evidence

Persistent organism state and bounded topology src/sim/xylos-mind.ts
Organism-specific lifecycle and recorded-input boundary src/core/xylos-session.ts
Explicit sensing and action interface src/sim/xylos-habitat.ts

Scientific basis: autopoiesis1974 · bedau2000

16life · Individual organization and viability

Resource-dependent self-maintenance

Acquired resources support ongoing maintenance of the entity's operational organization.

resource_maintenanceentity / worldMixed or restricted evidenceNot separately represented in the nine-axis survey

Entities consume edible resources and receive nourishment; Xylos energy affects hunger and movement. Existing resource tests cover ownership, depletion and timed restoration. These controls do not establish autonomous maintenance of operational organization or deprivation-dependent viability; scheduled resource respawn remains a designed intervention.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

resource-budget accounting
viability versus deprivation
maintenance flux and replenishment recovery

Acceptance experiment

Couple declared resources to real model costs; apply deprivation and replenishment, disable maintenance actions, and compare to a resource-independent control.

What would weaken the claim?

Resource numbers are cosmetic, survival is invariant to resource access, or the evaluator silently restores state.

Repository evidence

Grazing and energy-dependent control src/sim/entities.ts
Resource-linked energy and movement src/sim/xylos-mind.ts
Single-consumption accounting and explicit five-second respawn tests/edible-resource.test.ts

Scientific basis: autopoiesis1974 · homeostasis2014

17life · Individual organization and viability

Homeostatic regulation

Feedback keeps important internal variables within declared viable ranges despite disturbance.

homeostasisentityBounded test evidenceNot separately represented in the nine-axis survey

A criticality controller adjusts activation gain from a leaky branching-ratio estimate; existing tests cover silent-input excitation, bounded trajectories and an active-input regime. This is a narrow neural-gain homeostat, not evidence for whole-organism viable-range regulation or robust criticality across the new disturbance protocol.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

time in viable range
disturbance-recovery time
regulation cost and overshoot

Acceptance experiment

Predefine viable ranges, apply sustained and transient disturbances, compare active regulation with clamped or disabled feedback, and test novel disturbance magnitudes.

What would weaken the claim?

The system is stable only because the world never perturbs it, variables are directly reset, or feedback removal has no effect.

Repository evidence

Bounded gain feedback and simplified branching estimator src/sim/criticality.ts
Local feedback and boundedness tests tests/criticality.test.ts

Scientific basis: homeostasis2014

18life · Individual organization and viability

Repair and functional resilience

Function is restored or preserved after damage through regeneration, reconfiguration or adaptation.

repair_resilienceentityNot establishedNot separately represented in the nine-axis survey

Xylos supports bounded plastic rewiring and stable long-run state, but this review found no matching unseen-lesion recovery comparison for the proposed repair axis. Rewiring, visual morphing and timed resource restoration should not be counted as demonstrated functional regeneration or damage compensation.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

function recovered relative to predamage baseline
recovery time and resource cost
performance under unseen lesions

Acceptance experiment

Apply untrained lesions, remove damaged components, separate structural regrowth from compensatory control, and compare enabled versus disabled repair with matched resources.

What would weaken the claim?

Recovery is automatic scene reset, reserved replacement injection, or occurs only for memorized damage cases.

Repository evidence

Sparse topology rewiring mechanism src/sim/xylos-mind.ts
Existing stability checks, not an unseen-lesion recovery assay tests/xylos-mind.test.ts
Mechanisms and remaining research interventions docs/XYLOS.md

Scientific basis: nca2020 · repair2015

19theory-linked · Consciousness and sentience research

Recurrent processing

Feedback interactions perform temporally extended processing with a demonstrated functional contribution.

recurrent_processingentityBounded test evidenceLegacy construct correspondence: Consciousness-theory

Xylos has sparse recurrent state, and existing matched-input tests cut recurrence and observe changed activity. This establishes a causal local mechanism; selective task benefits under timed, compute-matched interruption and broader theory-linked interpretation remain unassessed. It is not a consciousness finding.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

recurrent influence over time
task-specific recurrence-ablation effect
stability under input removal

Acceptance experiment

Identify feedback pathways; interrupt recurrence at controlled times while preserving feed-forward computation where possible; compare intact, sham and compute-matched conditions.

What would weaken the claim?

A loop exists only in the scheduler, recurrent states carry no task-relevant content, or cutting them has no predicted selective effect.

Repository evidence

Explicit recurrent pathways and intervention option src/sim/xylos-mind.ts
Recurrence ablation with matched numeric inputs tests/xylos-mind.test.ts

Scientific basis: butlin2023 · butlin2025

20theory-linked · Consciousness and sentience research

Global availability and broadcast

Selected information becomes usable by multiple otherwise specialized processes through a constrained shared route.

global_availabilityentityBounded test evidenceLegacy construct correspondence: Consciousness-theory

Xylos admits two of six specialist groups into a constrained workspace, and existing tests check admission and broadcast effects across groups. This is a bounded broadcast mechanism; flexible use of novel content by independently tested consumers and the full new protocol remain unassessed.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

cross-module access
broadcast bottleneck behavior
downstream dependence on selected content

Acceptance experiment

Introduce novel content into one specialist, test flexible use by independent consumers, lesion the shared route and retain local processing; distinguish broadcasting from a globally readable log.

What would weaken the claim?

All modules simply read privileged world state, no selective bottleneck exists, or shared content has no causal effect on consumers.

Repository evidence

Two-slot workspace and cross-group broadcast src/sim/xylos-mind.ts
Admission bottleneck and broadcast intervention checks tests/xylos-mind.test.ts

Scientific basis: butlin2023 · butlin2025 · cogitate2025

21theory-linked · Consciousness and sentience research

Causal integration and differentiation

Perturbations reveal both interaction between parts and distinguishable system responses beyond shared input or noise.

causal_integrationentityMixed or restricted evidenceLegacy construct correspondence: Consciousness-theory

The indicator lab compares structured oscillator coupling with shuffled and ablated conditions, and tests check resulting index changes. Its synthetic driver and chosen estimator do not establish integrated information or perturbational complexity of the full organism, human PCI validity, or consciousness.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

perturbation spread and duration
response differentiation
partition and shuffled-surrogate effects

Acceptance experiment

Record standardized internal perturbations, partition candidate mechanisms, and compare coupled, disconnected, replayed and noise-matched surrogates. Publish the exact estimator and state-space scale.

What would weaken the claim?

The metric is explained by noise or common input, rises equally in disconnected surrogates, or changes mainly with sampling/estimator choices.

Repository evidence

Synthetic oscillator driver and explicit comparison conditions src/sim/consciousness-lab.ts
Bounded null and ablation checks tests/consciousness-lab.test.ts
Indicator-only and live-coupling limitations docs/VERIFICATION-ANALYTICAL-DATA.md

Scientific basis: pci2013 · cogitate2025 · butlin2025

22theory-linked · Consciousness and sentience research

Valence-related control and motivational trade-offs

Persistent positive/negative functional states modulate learning and flexible trade-offs; this is evidence about control, not felt experience.

valence_tradeoffsentityMixed or restricted evidenceNot separately represented in the nine-axis survey

Hunger, threat and uncertainty influence action utilities, and drive tests compare context-dependent hunting, fleeing and social responses. Signed valence telemetry and these designed trade-offs do not establish a persistent learned valence pathway, the new devaluation protocol or felt positive/negative experience.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

context-sensitive approach/avoidance
devaluation and reversal response
cross-context state persistence
state-ablation effect

Acceptance experiment

Vary competing benefits and costs, learned cues and internal state; test persistence and generalization; disable the proposed valence pathway and compare fixed-reflex/reward-only controls.

What would weaken the claim?

Only a signed reward variable or emotional animation is present, responses are fixed reflexes, or state manipulation has no specific behavioral consequence.

Repository evidence

Energy, hazard, valence telemetry and motor utility src/sim/xylos-mind.ts
Context-dependent drive functions src/sim/cognition.ts
Threat, hunger and social-context contrasts tests/cognition.test.ts

Scientific basis: sentience2021 · tradeoffs2022 · homeostasis2014

23engineering · Engineering and observation context

Substrate pluralism and model coverage

The platform supports meaningfully distinct dynamics or implementation substrates under explicit interfaces and evaluation conditions.

substrate_pluralismentity / population / worldImplementation identifiedLegacy construct correspondence: Substrate pluralism

CQM combines distinct classical mathematical mechanisms, including simulated quantum dynamics, geometry, recurrent controllers and fields, with documented adapters. Selected substrate coupling is implemented; equivalent cross-substrate evaluation and universal semantic fidelity are unassessed. These integrations do not imply physical quantum hardware or complete upstream runtimes.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

independent substrate families supported
cross-substrate protocol coverage
semantic fidelity of adapters

Acceptance experiment

Inventory implemented substrates, run equivalent declared probes on each, document where equivalence fails, and separate distinct dynamics from reskins or parameter presets.

What would weaken the claim?

Different substrates are only renamed copies, adapters leak answers, or comparison assumes unverified equivalence.

Repository evidence

Wave, geometry and structural-causal adapters src/sim/latent-substrates.ts
Bounded adapter behavior checks tests/latent-substrates.test.ts
Per-integration scope, provenance and limits docs/TSOTCHKE-INTEGRATION-MAP-2026-06-26.md

Scientific basis: bedau2000 · lenia2019 · evogym2021

24engineering · Engineering and observation context

Instrumentation and reproducibility

Experiments expose traceable state, interventions, configurations and replayable or statistically reproducible outcomes.

instrumentation_reproducibilityentity / population / worldBounded test evidenceLegacy construct correspondence: Instrumentation

Seeded receipts, bounded numeric recording and exact Xylos input replay are implemented and tested locally. Xylos replay suppresses habitat effects and does not rewind the shared world. Independent replication of this new profile and execution of its 25-axis protocols have not occurred.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

trace completeness
seed/replay agreement
independent rerun agreement
provenance and intervention coverage

Acceptance experiment

Archive commit, config, seed/entropy tape, task distribution, budgets and raw traces; replay locally and rerun independently; record numerical tolerances and missing data.

What would weaken the claim?

Claims cannot be traced to raw runs, unlogged randomness changes conclusions, or the displayed score cannot be recomputed.

Repository evidence

Bounded organism-specific input recording and replay src/core/xylos-session.ts
Seeded and recorded-input agreement checks tests/xylos-session.test.ts
Seed-dependent apex receipt and fingerprint checks tests/reproduce.test.ts

Scientific basis: reproducibility2021

25engineering · Engineering and observation context

Visual and embodied scale

Observable detail, responsiveness and inspectability across body and population scales; one complex being is a valid primary target.

visual_embodied_scaleentity / population / worldBounded test evidenceLegacy construct correspondence: Visual scale

Xylos alone is a valid visual subject: its neural state drives membranes and a visible sparse graph, with tests for geometry changes and fixed rendering bounds. This is separate from population capacity; no new hardware-specific latency, frame-rate or comparative aesthetic benchmark was run.

Capability score: uncalibrated

Measures, protocol and evidence

Record these measures

per-entity functional degrees of freedom
inspectable component depth
entity count separately
latency and frame rate at stated hardware

Acceptance experiment

Report single-entity complexity and population capacity as separate measures; inspect sensors, actions and internal state while varying visual load; compare displayed state to simulation traces.

What would weaken the claim?

More particles or organisms are presented as more intelligence, render detail lacks causal counterparts, or the viewer cannot inspect the claimed entity mechanisms.

Repository evidence

Individual-body membranes and live neural graph src/sim/xylos-body.ts
State-to-geometry coupling and bounded presentation tests/xylos-body.test.ts
Organism inspection and experiment controls src/ui/xylos-panel.ts

Scientific basis: reproducibility2021 · lenia2019

What a strong benchmark result must report

Feature plans and capability claims identify affected stable CQM-25 axis IDs and their evidence requirements. Taxonomy changes require an explicit framework version and contract review; evidence updates preserve axis identity and pinned provenance.
Declare the unit of analysis, entity boundary, intended modeling fidelity and applicable axes before evaluation. Treat individual, population and world measurements separately. N/A needs a scope reason; missing or failed evidence cannot be relabeled N/A.
Freeze the implementation, task generator, development and held-out distributions, training exposure, horizon, hardware and resource budget. Preregister primary metrics, meaningful effect sizes, failure conditions and stopping rules.
Use controls appropriate to the claim: random/fixed policy, frozen learning, interrupted recurrence, disconnected broadcast, replayed environment or matched sham intervention. Match compute and other confounds where feasible.
Report independent experimental units and their dependence. Consecutive frames are not independent replicates; population studies must account for shared lineages and use independently initialized runs.
Publish outcome distributions, uncertainty intervals, negative results, missingness and sensitivity analyses. Use multiplicity control for families of inferential tests; do not select only favorable seeds.
Archive the commit, configuration, seeds or recorded entropy inputs, interventions and raw outputs. Reproduce each result within stated numerical tolerances, and seek independent replication of the specific result.
Separate capability measurement, implementation evidence, causal evidence and external validity. A unit-test pass or a named module does not establish the full axis protocol. Strong evidence can support a negative result.
Before cross-system scoring, calibrate task-specific anchors on multiple systems, use independent raters and report agreement. Publish all eligible denominators and resource limits. Do not translate nine historical scores into 25 new scores.
For AGI or superintelligence comparisons, declare task-family coverage, capability depth, reference populations, adaptation budget, tool access, compute, time and human assistance. Report weak domains; neither a single exceptional task nor a profile average establishes general superhuman ability.
For neurometric analogues, identify the estimator, perturbation design, null models and validation domain. Human PCI thresholds, brain-score correlations and animal sentience criteria do not automatically transfer to software.
Reuse relevant existing evaluation resources such as Avida, MODES, EvoGym, Animal-AI, Melting Pot and Brain-Score with their own protocols. This profile organizes several research goals; it is not the first ALife benchmark or a universal standard.

Limits of this edition

Project adoption on 2026-10-22 establishes CQM-25 as the canonical framework. It does not rerun the 2026-10-21 evidence review: its 25 acceptance protocols remain proposed and unexecuted as a unified benchmark; every numeric capability score is null.
The selected axes and grouping are author-defined, correlated and not exhaustive. Primary literature motivates the constructs; it does not establish consensus on this exact profile or equal axis weights.
Implementation and bounded-test labels refer to cited repository mechanisms and narrow tests. Failed V4 and Phase-B advancement criteria remain visible; mechanism presence does not imply generalization.
Life-like organization, cognition, subjective consciousness and valenced sentience are different claims. Theory-linked mechanisms and reward or valence variables do not establish experience.
The profile supplements the original Butlin indicator mapping; it does not replace those 14 indicators or claim their criteria are all met. Human neuroscience and animal behavior evidence require justified transfer to artificial systems.
Engineering breadth, substrate count, photorealism, visual beauty, number of beings and parameter count contribute no points to mental-capability claims. No overall life, consciousness, sentience, AGI or SI score is calculated.
The separate provisional author estimates use an uncalibrated heuristic implementation rubric. Their imported rationales await independent review and are not new protocol results; they never populate canonical capability scores or produce aggregate verdicts.

Provisional author estimates

Uncalibrated heuristic 0–5 judgments — not measured protocol results. Recorded 2026-10-22 against the original evidence reference 339c6daa. All imported rationales are pending independent review; their claims are not newly verified by this publication. Canonical capability scores remain null.

A heuristic author judgment about implementation breadth and integration within each axis and its declared scope. These ordinal anchors are not equal-interval measurements, calibrated capability scores, probabilities or scientific acceptance thresholds.

The incoming values and rationales are preserved verbatim. This rubric declares how the provisional overlay is to be read; the imported values have not been recomputed or independently rated against its anchors. Decimal values indicate author interpolation only, not measurement precision.

Download provisional author estimates and rubric (JSON)

Heuristic rubric: all six anchors

No implementation identified: The author judges that no relevant implementation is present in the inspected scope. Missing evidence alone does not justify zero.
Illustrative component: An isolated demonstration or component suggests the construct, with little integration into the operating system.
Partial implementation: Some relevant mechanisms operate, but major parts of the declared construct or integration remain absent or restricted.
Bounded working implementation: Relevant mechanisms function together in a limited setting; broader behavior, coverage or acceptance evidence remains incomplete.
Substantial integrated implementation: Multiple relevant mechanisms or behaviors are substantially integrated in the inspected scope, while material limitations remain.
Extensive implementation in declared scope: The author judges implementation breadth and integration to be extensive within the declared scope. This is not a claim of scientific validation, unlimited generality or human-level ability.

All 25 values are provisional author estimates, not measured protocol results. Imported rationales may contain unverified interpretation or numerical claims and await independent review.
The original evidence review remains dated 2026-10-21 at the pinned commit. Recording these estimates on 2026-10-22 is not a new execution or review of the canonical acceptance protocols.
Different axes concern different constructs and units of analysis. Values are not calibrated across axes or systems, so no overall score, domain mean, ranking or uncertainty interval is published.
Engineering estimates describe implementation context only. No value is a probability or verdict of consciousness, sentience, AGI or superintelligence, and theory-linked implementation does not establish experience.
Canonical assessment.score remains null for all 25 axes. Missing measurements remain unknown; not-applicable requires an explicit subject and scope reason.

25 separate author estimates. No aggregate or domain mean is published.

Swipe horizontally, or focus the table and use the arrow keys, to read submitted rationales.

Provisional author estimates (0–5)
Canonical axis / kind	Heuristic estimate (0–5)	Submitted rationale
Reproduction and heredity `reproduction_heredity` life	4.0	Imported rationale — pending review Code-grounded 4.0 (GATE-REPRO-SELECT): heritable nW scales auto-fission probability (0.02 + 0.08*nW) and lifespan phenotype in the live EntityManager population (+0.180 separation vs trait-blind control across 500 frames x 8 seeds); held below 5.0 because open instruction-level self-replication and live sexual crossover are not active.
Evolutionary innovation and open-endedness `evolutionary_innovation` life	2.0	Imported rationale — pending review Code-grounded 2.0: directional fecundity selection converges onto a fixed monotone optimum (nW -> 1) and Petri strain selection operates over a fixed 26-form catalog; held at 2.0 because unbounded open-ended novelty and superiority to neutral drift on novel held-out tasks are not established.
Environmental and ecological coupling `environmental_coupling` life	3.4	Imported rationale — pending review Code-grounded 3.4 (GATE-DOME-REFUGE + GATE-XENO-TROPHIC + Xylos habitat): bidirectional Gray-Scott chemical sensing/actuation for single-being Xylos plus 60k edible flora, waste-to-fertilizer recycling, Xenomimic predator-prey regulation, and Big Tree sanctuary refuge (+7.6% equilibrium lift); held below 5.0 because multi-generation plant-genome coevolution is absent.
Morphology and development `morphology_development` life	3.8	Imported rationale — pending review Code-grounded 3.8: Xylos neural-state-driven membrane/graph deformation, SuperBody consciousness/evolution-coupled multi-appendage morphing, and 250 procedural phyla morphotypes; held below 5.0 because morphology is state- and genome-driven rather than evolved via EvoGym/Sims joint body-brain developmental selection.
Embodied dynamics and internal consistency `embodied_dynamics` engineering	3.9	Imported rationale — pending review Code-grounded 3.9: fixed-step deterministic integration, partition-invariant Xylos session stepping, live fused f32 Gray-Scott PDE + Crank-Nicolson wavefunction dynamics, and soft plant-canopy contact under declared xeno-physical laws (Earth-physics replication is not required); held below 5.0 pending a full cross-subsystem PDE convergence audit.
Perception and perceptual integration `perception` cognition	3.8	Imported rationale — pending review Code-grounded 3.8 (GATE-CHEMOTAXIS + Xylos reliability percepts): bilinear 4-cell flora gradient foraging (>3x biomass capture vs blind control) and Xylos precision-weighted chemical/hazard/gradient integration; held below 5.0 because channels consume structured simulation fields rather than raw pixel/occlusion vision.
Learning and plasticity `learning` cognition	4.5	Imported rationale — pending review Code-grounded 4.5 (GATE-BIOLOGIC-LEARN + GATE-MLP + GATE-SELFMODEL + Xylos plasticity): exact Eshkol forward/reverse AD online fitness learning in the live Petri, nonlinear XOR-capable MLP self-model backprop, and Xylos associative plastic predictors outperforming frozen controls.
Memory and continual retention `memory` cognition	3.7	Imported rationale — pending review Code-grounded 3.7: Plate HRR circular-convolution holographic associative memory (context-action binding, unbinding, and decay) plus fast Hebbian plastic weight imprints; held below 5.0 pending multi-task continual-retention benchmarks across interleaved held-out task families.
Predictive and causal world models `predictive_causal_models` cognition	2.6	Imported rationale — pending review Code-grounded 2.6: Xylos action-conditioned sensory predictors reduce forecast error on local feedback loops, but V4 adaptive ecology predictors failed frozen/shuffled controls and Phase-B temporal Predictor-V3 failed all 8 advancement criteria; score explicitly reflects those negative results.
Autonomous goal regulation `autonomous_goals` cognition	4.0	Imported rationale — pending review Code-grounded 4.0 (GATE-VQE-RESOLVE + Xylos autonomous policy): unprompted need-weighted hunger/curiosity/hazard motor policy in Xylos plus Variational Quantum Eigensolver 4-drive frustration minimization via exact parameter-shift gradients in the base population.
Reasoning and planning `reasoning_planning` cognition	3.5	Imported rationale — pending review Code-grounded 3.5: NHI multi-step GOAP planner with material-outcome fact revision plus discrete active-inference expected-free-energy policy evaluation in SuperMind; held below 5.0 because combinatorial detour and novel relational puzzle generalization are unestablished.
Transfer and compositional generalization `transfer_generalization` cognition	2.2	Imported rationale — pending review Code-grounded 2.2: under the frozen 64-seed V4 cross-being causality benchmark (1,152 rows), Titans passed all inference/magnitude/replay gates (1/4 families), while ordinary organisms, Petri biologics, and adaptive predictors failed magnitude or control gates.
Metacognition and self-modeling `metacognition_self_model` cognition	3.7	Imported rationale — pending review Code-grounded 3.7 (GATE-PREDICTIVE-METACOG + GATE-SELFMODEL, indicator-only): arbitrary-order Eshkol Taylor-jet trajectory extrapolation whose remainder causally damps VQE commitment under volatility, paired with online MLP forward self-modeling; held below 5.0 pending difficulty-matched meta-d'/d' calibration.
Social cognition and communication `social_communication` cognition	3.5	Imported rationale — pending review Code-grounded 3.5: 25 Theory-of-Mind organs (6-family ensemble), learned rival-proximity head, Nash game-theory encounters, and V4-verified Titan Prisoner's Dilemma diplomacy causality; held below 5.0 because open compositional signalling to unseen partner policies is unassessed.
Individuality and maintained boundary `individuality` life	3.8	Imported rationale — pending review Code-grounded 3.8: Xylos maintains an individuated recurrent neural manifold, persistent body topology, and recorded session boundary distinct from the shared world; held below 5.0 pending component-turnover and partition-cohesion perturbation assays.
Resource-dependent self-maintenance `resource_maintenance` life	3.4	Imported rationale — pending review Code-grounded 3.4: single-consumption edible resource accounting, energy-coupled foraging/movement in Entities and Xylos, and trophic energy transfer; held below 5.0 because 5-second flora/fauna respawn timers act as a designed habitat floor.
Homeostatic regulation `homeostasis` life	3.6	Imported rationale — pending review Code-grounded 3.6: branching-ratio neural criticality controller (sigma -> 1.0 homeostat) and multi-drive physiological regulation keep internal dynamics bounded across silent and active regimes; held below 5.0 pending whole-organism multi-variable viable-envelope perturbation trials.
Repair and functional resilience `repair_resilience` life	1.8	Imported rationale — pending review Code-grounded 1.8: Xylos implements bounded sparse plastic topology rewiring and state clamping, but unseen structural lesion and functional regeneration assays have not yet been executed.
Recurrent processing `recurrent_processing` theory-linked	3.6	Imported rationale — pending review Code-grounded 3.6 (indicator-only): explicit recurrent pathways in Xylos, SuperMind, and ordinary resource heads with matched-input recurrence-ablation tests demonstrating causal state divergence; computational indicator only, not phenomenal experience.
Global availability and broadcast `global_availability` theory-linked	3.6	Imported rationale — pending review Code-grounded 3.6 (indicator-only): capacity-2-of-6 Global Workspace admission bottleneck and cross-specialist broadcast in Xylos and SuperMind (Eshkol §17 GWT primitives), verified by admission and broadcast ablation tests; computational indicator only.
Causal integration and differentiation `causal_integration` theory-linked	3.3	Imported rationale — pending review Code-grounded 3.3 (indicator-only): exact bipartition quantum/classical integrated information (Phi) and perturbational oscillator complexity lab with shuffled/ablated surrogates; held at 3.3 because live mind-to-lab null separation remains near zero and software proxies do not inherit human PCI validity.
Valence-related control and motivational trade-offs `valence_tradeoffs` theory-linked	3.1	Imported rationale — pending review Code-grounded 3.1 (indicator-only): context-dependent hunger, threat, and uncertainty utility trade-offs modulate hunting, fleeing, and foraging across Entities, Xenomimics, and Xylos; functional control indicator only, not felt pleasure or pain.
Substrate pluralism and model coverage `substrate_pluralism` engineering	4.6	Imported rationale — pending review Code-grounded 4.6 (GATE-XENO-SCHRODINGER + 18 Zig WASM kernels + Tsotchke 17/21 causal ledger): independent live consumers of 5-qubit statevectors, Crank-Nicolson Schrödinger wavepackets, Aaronson-Gottesman Clifford tableaus, QGTL Fubini-Study metrics, Eshkol AD/Taylor towers, libirrep BdG skyrmions, and SbNN MinSR.
Instrumentation and reproducibility `instrumentation_reproducibility` engineering	4.4	Imported rationale — pending review Code-grounded 4.4 (GATE-RNG-PROVENANCE + canonical receipts): live Tsotchke randomness battery provenance (~0.99 vs degenerate stream), deterministic seeded mulberry32 streams, Xylos input-tape replay, and gate-enforced SHA-256 benchmark receipts.
Visual and embodied scale `visual_embodied_scale` engineering	4.2	Imported rationale — pending review Code-grounded 4.2: combines single-being biomorphic visual morphogenesis (Xylos neural-state membranes and live 3D connectome graph, SuperBody shader/appendage morphing, 250 phyla morphotypes: 4.4) with ecosystem instanced scale (60k flora, 10k default / 50k mega entities: 4.0); legacy 9-axis comparison matrix retains 4.0.

Research foundations

These sources motivate individual constructs and evaluation methods. They do not endorse the axis selection or establish universal thresholds.

Open problems in artificial life (research agenda)ALife spans origins, evolutionary potential, and relations to mind; this is an agenda, not a certification standard.
The evolutionary origin of complex features (primary experiment)Digital replication, mutation and selection can be investigated experimentally; complexity is task-relative.
The MODES Toolbox: Measurements of Open-Ended Dynamics in Evolving Systems (measurement framework)Tracks change, novelty, complexity and ecological potential. Finite observations cannot prove unlimited open-endedness.
Evolution Gym: A Large-Scale Benchmark for Evolving Soft Robots (benchmark)NeurIPS 2021 benchmark, arXiv posted 2022. Measures joint body/controller optimization, not consciousness.
Growing Neural Cellular Automata (primary experiment)Demonstrates learned growth, persistence and regeneration; repair is distinct from subjective experience.
Homeostatic reinforcement learning for integrating reward collection and physiological stability (computational neuroscience model)Connects internal-state regulation and reward-driven behavior; a candidate model rather than a universal requirement.
Overcoming catastrophic forgetting in neural networks (primary experiment)Measures retention of previously learned tasks across subsequent training.
The Animal-AI Environment: Training and Testing Animal-Like Artificial Cognition (benchmark)Animal-cognition-inspired testing beyond training environments; no implication of sentience.
Position: Levels of AGI for Operationalizing Progress on the Path to AGI (research framework)Separates performance, generality and autonomy. A proposed framework, not a universally accepted AGI certificate.
Measuring Progress Toward AGI: A Cognitive Framework (preprint framework)Proposes held-out cognitive tasks and cognitive profiles; May 2026 preprint, not established consensus.
Consciousness in Artificial Intelligence: Insights from the Science of Consciousness (theory-derived research framework)Computational indicators drawn from several theories; meeting indicators is not proof of consciousness.
Identifying indicators of consciousness in AI systems (methodological research)Theory-derived indicators address uncertainty and gaming; criteria remain revisable.
Adversarial testing of global neuronal workspace and integrated information theories of consciousness (preregistered primary experiment)Human neuroscience results challenged important predictions of both theories; do not portray either as settled consensus.
A theoretically based index of consciousness independent of sensory processing and behavior (primary neurophysiology experiment)Human TMS/EEG perturbational complexity index; software analogues do not inherit human thresholds or diagnostic validity.
How to measure metacognition (measurement methodology)Distinguishes metacognitive sensitivity, bias and efficiency; match first-order performance before comparison.
Review of the Evidence of Sentience in Cephalopod Molluscs and Decapod Crustaceans (commissioned evidence review)Animal neural and behavioral indicators motivate converging evidence; no validated transfer to software.
Motivational trade-offs and modulation of nociception in bumblebees (primary animal experiment)Context-dependent trade-offs motivate functional assays; reward variables and avoidance alone do not establish felt valence.
Improving Reproducibility in Machine Learning Research (research methodology)Supports explicit code, data, experimental configuration and reproducibility reporting.
Lenia - Biology of Artificial Life (ALife research platform)Continuous cellular automata support varied organism-like patterns; morphology alone does not establish cognition.
Brain-Score: Which Artificial Neural Network for Object Recognition is most Brain-Like? (neural and behavioral benchmark)Neural/behavioral similarity is domain-specific; neither brain similarity nor a composite is a consciousness measure.
Robots that can adapt like animals (primary robotics experiment)Tests recovery of function after unseen damage; distinguishes functional adaptation from structural regrowth.
Scalable Evaluation of Multi-Agent Reinforcement Learning with Melting Pot (benchmark)Tests social interaction generalization to novel partners and situations.
Autopoiesis: The organization of living systems, its characterization and a model (theoretical model)One influential account of maintained organization; not a settled universal definition of life.