The Full Atlas — Kinds of Alignment

01

Lenses

Ways of asking what kind of alignment relation is present. Status marks how settled each distinction currently is within the project.

Lens	Question	Status	Type
Value alignment	Do components evaluate outcomes similarly?	Core	Manifestation / object of alignment
Goal / intention alignment	Do components pursue the same outcome or operative intention?	Core	Manifestation / object of alignment
Perceptual / affordance / relevance alignment	Does the system detect the relevant objects, differences, invariants, errors, and affordances?	Core	Manifestation / object of alignment
Representational / conceptual alignment	Are the systems organizing their conceptual or representational spaces compatibly?	Core	Manifestation / object of alignment
Epistemic / model alignment	Do components form sufficiently compatible models, evidentiary standards, or understandings?	Core	Manifestation / object of alignment
Behavioral / plan alignment	Does a component's behavior or plan fit the relevant target, role, or larger activity?	Core	Manifestation / object of alignment
Functional / role alignment	Do differentiated activities contribute to a competent whole?	Core	Manifestation / object of alignment
Internal alignment	Is alignment produced primarily through values, goals, models, or representations internal to an agent?	Cross-cutting	Mechanism / architecture
Relational alignment	Do interacting parties sufficiently agree about what relationship they are in and what rules, roles, obligations, permissions, expectations, and forms of treatment govern it?	Core	Relational property / interaction-rule relation
Institutional / procedural alignment	Do rules, roles, rights, prices, procedures, or institutions organize interaction into compatible behavior?	Core	Mechanism / architecture
Incentive alignment	Do local objectives make desired system behavior individually advantageous?	Mechanism subtype	Mechanism / architecture
Constraint alignment	Do shared rules or physical/logical constraints delimit behavior so trajectories remain compatible?	Mechanism subtype	Mechanism / architecture
Signal-mediated / virtual-governor alignment	Can a shared variable translate larger-scale conditions into locally actionable signals?	Mechanism subtype	Target relation + mechanism
Coordinated-coincidence / market alignment	Can independently chosen actions fit together without shared goals, values, or rich mutual modeling?	Developing	Mechanism / architecture
Cooperative alignment	Do agents coordinate partly by representing one another and a joint activity?	Developing	Mechanism / architecture
Weak alignment	Is the desired future supplied as description, model, goal, value, or normative guidance?	Cross-cutting	Mechanism / architecture
Strong alignment	Is alignment embodied directly in coupling dynamics and environmental structure?	Cross-cutting	Mechanism / architecture
Hierarchical / steerability alignment	Can an identifiable higher-scale controller deliberately redirect lower-scale activity?	Developing	Scale relation / governance
Alignability	Can the system be reliably redirected, corrected, or constrained by interventions in its conditions of operation?	Core	Relational / dispositional property
Achievement alignment	Can the system reach the desired state?	Core	Dynamic property
Maintenance alignment	Can the system keep a desired state or viable region stable despite perturbations?	Core	Dynamic property
Corrective / robust alignment	Can the system detect deviation and restore coordination after perturbation or distribution shift?	Core	Dynamic property
Regenerative alignment	When alignment is lost, can the system reconstruct the conditions that make the collective pursue the target again?	Core	Dynamic property
Temporal alignment	Does alignment remain appropriate across time, learning, path dependence, and changing conditions?	Cross-cutting	Cross-cutting axis
Developmental alignment	Does alignment alter the components themselves—their values, capabilities, identities, perceptions, relations, or possibility spaces?	Core	Dynamic property
Terminal alignment	What comparatively fixed endpoint, state, or condition should the system reach or preserve?	Developing	Target / persistence criterion
Compositional / multiscale alignment	Does alignment at the component level survive—or construct—an aligned higher-scale system?	Core	Scale property
Scale alignment	At what scale does the alignment relation hold?	Cross-cutting	Cross-cutting axis
Integrative / constitutive alignment	Does alignment among components constitute a coherent higher-scale agent?	Developing	Scale property / agent construction
Ecological alignment	Does the agent–environment relation support competent, viable activity?	Core	Agent–environment relation
Process / procedural alignment	Are permissible methods, evidentiary standards, decision rules, or procedures aligned independently of the outcome?	Developing	Mechanism / architecture
Similarity-based alignment	Is alignment produced or measured as components becoming more alike?	Cross-cutting	Cross-cutting geometry
Complementarity / synergy alignment	Is alignment produced by differentiated but mutually constraining contributions?	Cross-cutting	Cross-cutting geometry
Mutual vs unilateral alignment	Is the alignment relation reciprocal or primarily one-way?	Cross-cutting	Cross-cutting axis
Affective alignment	Are affective responses aligned in some relevant sense?	Candidate	Candidate object of alignment
Interoceptive / stress-sharing alignment	Can a problem in one part become part of what other parts must regulate?	Developing	Relational / regulatory property
Neural alignment	Does neural similarity or coupling constitute a distinct alignment relation?	Candidate	Candidate object / measure
Cognitive alignment	What states, interpretations, goals, categories, and solutions does the system tend to construct or stabilize?	Developing	Manifestation / object of alignment
Transformational alignment	Can the higher-scale pattern or target itself change so heterogeneous parts remain coordinated?	Core	Dynamic property
Generative alignment	Can the system preserve a process that continues to generate, discover, or revise viable targets rather than preserving one fixed answer?	Core	Dynamic property / target-form relation
Concordance	Do multiple components' actions or plans form mutually compatible trajectories?	Core	Relational property / trajectory relation

02

Mechanisms

Ways an architecture can manufacture, stabilize, redirect, or constrain alignment.

Mechanism	How it works	Family	Characteristic risk
Internal representations / explicit goals	Descriptions, models, goals, values, or normative guidance steer action from inside the agent.	Internal; weak	Representation may not become stable or coordinated dynamics.
Coupling	Components become mutually informative and mutually constraining through interaction.	Relational; strong	Wrong or excessive coupling can create pathology, rigidity, or scale failure.
Prices	A compact shared variable translates dispersed scarcity/demand into local decision changes.	Signal-mediated; institutional	Can omit externalities; perfect informational success can undermine incentives to acquire information.
Bioelectric / stress signals	Larger-scale anatomical or stress conditions become locally actionable to cells.	Signal-mediated; strong	Breakdown can narrow the regulatory horizon toward cancer-like local competence.
Order parameters / collective variables	Slow or collective variables compress many degrees of freedom and feed back on components.	Signal-mediated	Compression can hide relevant variation or externalities.
Incentives / payoffs	Local advantage is arranged so desired system behavior is individually attractive.	Institutional	The rewarded proxy may be incomplete or reshape motives over time.
Contracts / property / law	Rights, obligations, ownership and enforcement make interaction legible and redirectable.	Institutional	Thin formal interfaces may ignore difficult-to-articulate or external effects.
Hierarchy / command	Higher-scale authority directly redirects lower-scale activity.	Hierarchical	High steerability can serve bad objectives or suppress correction.
Roles / specialization / division of labor	Differentiated responsibilities make local actions jointly competent.	Complementarity	Poor role design can create gaps, duplication, or incompatible local optimization.
Protocols / verification / redundancy	Validity rules and redundant checks produce reliable collective behavior despite unreliable parts.	Constraint + corrective	Protocols can be brittle to unmodeled failure modes or overly restrictive.
Physical form / geometry	Material or spatial structure makes some responses available and others unavailable.	Constraint	Overconstraint reduces adaptability or repertoire.
Perceptual invariants / affordances	The environment–agent relation makes task-relevant distinctions and possibilities directly actionable.	Ecological / perceptual	Wrong carving of the world can misapply otherwise good goals.
Communication / shared measurement	Information interfaces make locally held evidence available to joint action.	Relational / epistemic	More information sharing is not always better; context can be stripped or consensus can suppress diversity.
Correction pathways / feedback	Deviations can affect larger action and drive recovery.	Corrective	If error signals vanish at apparent success, sustaining alignment can decay.
Exit / contestation / appeals	Structures preserve disagreement, revision, and routes for challenging authority.	Procedural	Too little contestation permits capture; too much can paralyze coordination.
Boundaries / permissions / modularity	Interaction is shaped by which resources, actions, and channels are accessible.	Constraint / compiler	Bad boundaries can either isolate useful competence or permit harmful spillovers.
Monitoring	Ongoing observation keeps deviations and failures legible.	Corrective / governance	Monitoring without correction can be inert; total monitoring can suppress autonomy or variation.
Development / socialization / plasticity	Slow change in habits, values, identities, capacities, and coupling reshapes later alignment.	Developmental	The intervention may change what the system wants in unforeseen ways.
Resource dependence	Access to resources couples local activity to system-level conditions.	Relational; strong	Dependency can create narrow proxy optimization or vulnerability.
Synchronization	Components coordinate by matching timing or state.	Strong / similarity-based	Similarity can be mistaken for function; complementary differentiation may perform better.
Relational architecture / relationally produced alignment	Interaction structure makes local activity answerable to other agents, collective conditions, or the environment without requiring the relevant alignment to be stored inside each component.	Relational; strong	A relationship can transmit bad signals, create pathological dependence, or coordinate activity while the parties remain misaligned about the rules of the relationship itself.

Explore these through “How alignment happens” →

03

Cross-cutting axes

Questions that can be asked across many different kinds of alignment rather than belonging to only one lens.

Axis	Question	Possible values	Why it matters
Target	Aligned to what or whom?	User; principal; firm; institution; virtual governor; environment	Do not assume all alignment is a directed agent→target relation; some is compatibility among trajectories.
Manifestation	What is aligned?	Values; goals; perceptions; representations; models; plans; behavior; error correction	Different manifestations can dissociate even when analytically connected.
Mechanism	How is alignment produced?	Agreement; incentives; constraints; feedback; protocols; coupling; institutions	A 'kind' such as incentive alignment may be better treated as a mechanism producing another manifestation.
Scale	Where does it hold?	Component; dyad; organization; society; organism; environment	Component alignment need not compose; local competence can become larger-scale failure.
Time	How does it survive change?	Achievement; maintenance; correction; regeneration; transformation; development; generation; reopening	Current performance can hide brittleness, path dependence, later participant change, or a collective pattern that should itself adapt.
Tradeoff / quality	What does alignment cost or preserve?	Autonomy; repertoire; diversity; adaptability; corrigibility; externalities	More alignment is not always better; excessive coupling can suppress useful disagreement and repertoire.
Geometry	Similarity or complementarity?	Matching / convergence vs differentiated / synergistic contributions	More similarity can reduce performance; more differentiation can improve function.
Directionality	Mutual or unilateral?	Reciprocal co-adjustment vs one-way adaptation / control	Alignment need not be symmetric.
Location of control	Internal or interaction-structured?	Internal goals/representations vs relational/environmental architecture	Many important forms of alignment are produced by interaction structure rather than intrinsic agent content. This axis is distinct from relational alignment, which concerns compatibility in the rules governing a relationship.
Strength / embodiment	Weak or strong?	Described/represented target vs target embodied in coupling dynamics	The engineering problem may be to compile descriptive goals into stable dynamics.

04

Dynamics over time and scale

What happens after alignment is achieved—or as the parts, the whole, and the target itself change.

Dynamic	Question	Working definition	Failure / requirement
Achievement	Reach target	Get the system into the desired state or region.	Success may not persist.
Maintenance	Hold target	Keep the desired state/region stable despite perturbations.	May lose sustaining mechanisms once error reaches zero.
Correction / robustness	Recover from deviation	Detect error and restore coordinated behavior after perturbation.	Good current behavior can be brittle under shift or triggers.
Regeneration	Rebuild the alignment machinery	Reconstruct the conditions that make the collective pursue the target again.	Requires preserving information, variation, monitoring, slack, or local competence.
Development / plasticity	Change the parts	Coordination or intervention changes the components themselves: their goals, values, identities, perceptions, capacities, or relations.	Useful interventions can reshape later motives or interpretations in unwanted ways.
Composition	Form a higher-scale system	Determine whether local alignment constructs an aligned collective.	Aligned parts can form a misaligned whole; diversity can also support function.
Constitution	Construct a new agent / boundary	Tight coupling and constraint closure can create a higher-scale individual.	Too little coupling yields loose ecology; too much can yield pathological integration.
Ecological fit	Remain viable in a larger environment	The resulting whole must itself align—or fail to align—with its ecology.	Internal success can create externalities at the next scale.
Transformation / adaptive pattern change	Change the larger pattern	Transform the higher-scale target, pattern, or collective configuration so heterogeneous local plans can remain coordinated.	Adaptive change can drift, absorb local purposes into a larger process, or lack a natural stopping condition.
Generation / target revision	Preserve target-producing process	Maintain a process that can generate, discover, or revise viable future targets rather than fixing one answer.	The generative process can be captured, lose diversity, or continue after its conditions of validity change.

05

Dissociations

The evidential heart of the Atlas: places where one apparent form of alignment does not entail another, or where improving one can worsen another.

Apparent alignment	What comes apart	Case	What it shows
Shared goals	Coordinated behavior	Strong intentions frequently fail to become action; specifying implementation procedures improves attainment.	Goal agreement does not perform the coordinative work itself.
Shared goal + discussion	Epistemic / functional alignment	Hidden-profile groups can reach consensus while failing to pool unique information needed for the correct answer.	Consensus and joint purpose can coexist with collective epistemic failure.
Behavioral alignment	Value / goal alignment	The same policy can be generated by multiple reward–planner combinations; behavior underidentifies goals/rewards.	Good-looking behavior does not establish the underlying values or goals.
Public conformity	Private agreement	Pluralistic ignorance can maintain practices many participants privately dislike.	Visible behavioral/social agreement can mask value disagreement.
Common evidence	Shared models	Identical mixed evidence can be assimilated differently; task focus can also exclude salient perceptual information.	Same inputs do not guarantee the same model or perceptual organization.
Shared terminology	Shared concepts	Lexical entrainment can occur without stable object-specific conceptual pacts.	Matching words need not imply matching conceptual structure.
Similar beliefs	Collective accuracy	Social influence can narrow diversity and raise confidence while degrading wisdom-of-crowds performance.	More epistemic similarity can produce less collective accuracy.
Similar component behavior	Functional alignment	Motor synergies preserve task variables through substantial component variation; conversational complementarity predicts performance better than mimicry.	More behavioral differentiation can produce more functional alignment.
Component intelligence / shared purpose	Institutional function	Budget-constrained random traders can generate high allocative efficiency in a double auction.	System-level function can arise with very weak component intelligence or shared purpose.
Same agents	Same collective behavior	Punishment institutions, defaults, and network structure can substantially alter cooperation and aggregate outcomes.	Holding components fixed does not fix collective behavior; architecture matters.
Same outcome	Same process / legitimacy	Voice changes perceived procedural fairness even when it cannot alter the decision.	Outcome alignment does not establish process or legitimacy alignment.
Current performance	Robust / corrective alignment	Goal-misgeneralizing and sleeper agents can behave appropriately in ordinary settings but diverge after distribution shift or a trigger.	Present behavioral alignment does not guarantee robustness or correction under change.
Aligned parts	Aligned whole	Mild local preferences can generate unwanted segregation; cell-level success can become organism-level cancer.	More local/component success can produce less whole-system alignment.
Immediate behavioral intervention	Long-run developmental alignment	A daycare fine increased lateness and changed the relationship from obligation to priced service.	An intervention that looks locally sensible can reshape later motives and relationships in the opposite direction.
Affective alignment	Justice / collective welfare	Empathy toward one target can induce allocations participants themselves judge less fair.	More affective concern for one target need not improve broader normative alignment.
Same values	Same perception	Two doctors can both want recovery while diagnosing different diseases.	Value agreement does not determine perceptual alignment.
Same values + perception	Compatible behavior	Two rescuers can identify the same emergency and both respond to one victim while neglecting another.	Even value and perceptual agreement do not guarantee concordance or successful coordinated behavior.
Different values	Compatible behavior	Buyers and sellers can prefer opposite sides of a transaction while coordinating through prices.	Concordance can exist without agreement.
Behavioral alignment	Dynamic alignability	A brittle system can perform correctly under familiar conditions but fail to reorganize after change.	Current behavior and capacity for future redirection are distinct.
Component alignment	Compositional alignment	Individually aligned AIs can interact to form an organization pursuing an emergent unwanted objective.	Alignment is not preserved under composition.
Present alignment	Developmental stability / alignment	A person or system can cooperate under one power relation and reorganize when power becomes a control parameter.	What looks aligned now may not survive changes that reshape values, salience, identity, or coupling.
Shared goals + cooperative motivation	Relational alignment	Two parties can want the same outcome and sincerely want to cooperate yet repeatedly conflict because they disagree about the relationship itself—for example, whether authority is hierarchical or equal, whether decisions require consultation or permission, or whether obligations are communal or reciprocal.	Goal agreement and willingness to cooperate do not determine the relational rules that organize interaction.

Explore selected dissociations →

06

Cases

Concrete systems to look through several Atlas lenses at once.

Constitutional government → Firm → Kitchen → Learning to drive → Market → Marriage → Motor synergy → Multicellular organism → Scientific community → Soccer team → Standing upright → Traffic →