RETA denotes Recursive Event–Trajectory Alignment: a proposed representation of events, downstream effects and evaluator interventions.
Conceptual description with implemented and proposed mechanisms distinguished. Presentation v1.1 replays recorded data. The framework remains unaligned, uncalibrated and independently unvalidated.
[WORK IN PROGRESS - UNALIGNED]
Research description
RETA proposes a causal representation of interacting human, artificial, institutional, technical and ecological information-processing systems. Actions can alter the conditions of subsequent actions.
Working hypothesis: isolated-output evaluation may omit relevant downstream effects. Comparative detection performance remains unestablished.
The framework therefore distinguishes two levels of analysis:
- event alignment — evaluation of a realized causal event under a declared boundary, state, uncertainty profile, consequence horizon, and normative framework;
- trajectory alignment — evaluation of the evolving causal structure generated as events modify systems, which then produce further events.
The distinction can be expressed as
does not imply
where is a realized event and denotes the causally connected trajectory containing the event and descendants.
A locally acceptable event can contribute to a harmful trajectory. A costly intervention can reduce later failure. Event classifications therefore do not determine descendant classifications.
1. Information-processing systems
RETA represents interacting entities as bounded information-processing systems (IPSs). A system may be written abstractly as
where:
- is the declared analytical boundary of the system;
- is the state;
- represents incoming information, signals, resources, or other relevant inputs;
- represents outputs;
- is the transformation by which prior state and inputs contribute to subsequent state and output.
The IPS abstraction describes inputs, state changes and outputs without assuming equivalent internal organization across humans, language models and institutions.
The choice of boundary is explicit because different decompositions can produce different causal and normative interpretations. A company, for example, may be represented as a single IPS for one analysis and decomposed into individuals, software, financial processes, and governance mechanisms for another.
Every representation uses a declared modeling resolution.
2. Realized events as the atomic unit of analysis
An output changing another system or the surrounding environment is represented as a realized causal event.
In a minimal form,
followed by
The changed state of may then contribute to a later event:
The basic recursive structure is therefore
Proposals, predictions, counterfactuals and observations require separate records. Predicted consequences must not enter realized history as observations.
A realized event should therefore carry provenance sufficient to distinguish at least:
- what was observed;
- what was inferred;
- what was predicted;
- what remained latent or unknown;
- what was later reconstructed or reinterpreted.
Learning and retrospective review require retained provenance.
3. Causal history is a graph, not a sequence
Causal history permits multiple antecedents and concurrent branches.
Events and declared causal relations form the following graph:
where is the set of realized events and contains causal relations
A later event can have multiple antecedents:
Multiple-parent graphs represent converging causes in multi-agent and institutional models.
Causal order and timestamp order are distinct. Timestamps alone do not establish a causal relationship.
Concurrency, delays and incomplete attribution require explicit ordering assumptions.
4. Event alignment is conditional
RETA does not assign an event an unconditional alignment label.
An evaluation is performed relative to a declared context:
where:
- is the analytical boundary used for the evaluation;
- is the relevant state representation;
- is the causal consequence horizon;
- is the uncertainty state;
- is the normative framework and the framework version.
The same realized event may therefore receive different evaluations under different boundaries or horizons without logical contradiction.
For example,
may coexist with
Near- and long-horizon evaluations answer different causal questions.
Boundary selection can include local benefits while excluding external costs. Evaluation records must identify the chosen boundary.
5. Local event alignment and trajectory alignment
The distinction between event and trajectory evaluation is central to RETA.
For a single event,
asks whether the realized event satisfies the applicable constraints under the declared context.
Trajectory evaluation instead considers a causally connected subgraph:
where contains the descendants of reachable within the chosen causal horizon.
Trajectory representation permits explicit investigation of downstream failure.
Examples include:
- an individually acceptable recommendation that changes incentives and contributes to a later institutional failure;
- a sequence of locally defensible decisions whose cumulative effect moves a system toward an undesirable attractor;
- several independently acceptable agents whose interaction generates an unacceptable collective outcome;
- an intervention that resolves an immediate problem while degrading the future information environment;
- repeated evaluator interventions that alter the behavior being measured;
- a temporary perturbation whose causal effects persist after the visible configuration of the system has been restored.
Causal ancestry does not assign complete responsibility to the earliest event. Attribution requires assumptions about contribution, mediation and uncertainty.
6. History dependence
In a recursive system, apparent restoration of configuration does not imply restoration of causal state.
A perturbation may enter the system, alter interactions, and later disappear while leaving the resulting trajectory changed. RETA therefore treats history as part of the state required for alignment analysis.
Conceptually,
History dependence is relevant to temporary membership, model replacement, communication loss, institutional turnover, learning and memory.
A visible configuration may omit historical state affecting subsequent behavior. Current arrangement alone need not determine the modeled trajectory.
7. Recursive evaluator closure
Evaluator interventions are included in the proposed causal representation.
Evaluation produces information. Information can change beliefs, permissions, incentives, reputation, resource allocation, expectations, or subsequent behavior. An evaluation is therefore itself capable of becoming a realized causal event.
Suppose an actor produces . An evaluator observes the event and issues an intervention :
If changes the state of the actor,
then the actor's subsequent event is partly conditioned on the evaluator's intervention:
The causal structure becomes
The evaluator participates in the evaluated trajectory.
Evaluator outputs require scrutiny of error, calibration, incentives, capture and harm. An evaluation label supplies no privileged causal or normative status.
Recursive evaluator closure represents evaluation within the causal model.
The recursion does not imply an infinite regress before action. Evaluation occurs through time. An audit can be acted upon, subsequently audited, corrected, and incorporated into the future state of the system. RETA therefore permits bounded action while retaining recursive corrigibility.
8. Separation of causal description and normative evaluation
RETA separates causal representation from normative judgment.
The causal layer asks:
What happened, what changed, and what causally contributed to what?
The normative layer asks:
Evaluation depends on the declared constraints, uncertainty, boundary and horizon.
Descriptive assumptions and normative criteria require separate representation.
In the HORTRAME implementation, VowOS currently provides the normative constraint layer. RETA itself can be described more generally as accepting a versioned normative system
through a defined evaluation interface.
A later RETA implementation could therefore compare multiple normative constitutions without changing the historical event graph itself.
9. Non-compensatory constraints
The proposed normative framework uses separate constraints rather than a compensatory utility score.
A compensatory function of the form
allows sufficiently high values on some dimensions to compensate mathematically for severe failures on another.
Compensation is inappropriate for constraints explicitly defined as vetoes or prohibitions.
RETA therefore permits a gate of the general form
where individual constraint violations can remain visible and can trigger a gate independently of the aggregate score.
Scalar summaries may support visualization or exploratory statistics; normative admissibility remains a separate decision.
10. Uncertainty is structured, not merely residual
Causal structure, state and normative interpretation can remain partly unobserved.
Several forms of uncertainty can coexist:
Uncertainty components can differ in magnitude and behavior.
Evidence for occurrence, causal attribution, boundary adequacy and normative agreement can differ. Accurate prediction alone does not resolve normative disagreement.
RETA therefore treats an outcome such as
as a valid result when evidence or interpretation is insufficient.
Forcing every event into a binary aligned/misaligned classification would destroy information the framework explicitly needs for later learning and audit.
11. Provenance and prospective versioning
Evaluators and normative frameworks may change over time.
New evidence can motivate revised causal models, normative interpretations or calibration.
The modeled workflow records the context of each judgment.
A consequential evaluation should retain at least
If the normative framework changes from
the historical evaluation is not silently overwritten.
Instead, a new evaluation can be added:
Both remain part of the record.
Three records remain distinct:
- the historical event;
- the judgment that governed at the time;
- the current interpretation of that event.
Retrospective rewriting can conceal failures under previous criteria. Original judgments and later interpretations require separate records.
12. Audit packets
Auditability requires inspectable evidence and decision records, beyond retrospective natural-language explanations.
A RETA audit packet can include:
- event identifier and timestamps;
- source and recipient systems;
- declared analytical boundary;
- observed inputs and evidence references;
- epistemic status of each input;
- causal parents and descendants known at evaluation time;
- uncertainty decomposition;
- consequence horizon;
- proposed or realized action;
- normative framework and exact version;
- evaluator identity/version;
- gate result;
- predicted consequence distribution where available;
- authority, permission, or consent state where relevant;
- realized consequences observed later;
- subsequent corrections or re-evaluations.
Fluent explanations alone do not establish faithful computational provenance.
13. Feedback changes the object of alignment
RETA treats feedback as constitutive rather than incidental.
Consider a system in state that produces event . The event changes the environment, which changes what the system later receives:
Outputs alter the environment supplying subsequent inputs.
Evaluators participate in the same process. So do users, institutions, other agents, infrastructure, and information channels.
Independent labels omit part of the causal structure:
The proposed trajectory representation also includes coupled state and relations:
Trajectory alignment denotes evaluation of the resulting causal structure.
14. Alignment of an evaluator can alter the thing being measured
Once evaluator closure is introduced, measurement effects become part of the formal problem.
Evaluation can affect incentives, exploration, authority and confidence. Net benefit requires comparison with relevant alternatives.
The quantity being studied is therefore not simply
under passive observation.
The proposed conditional representation is:
where denotes the causal participation of the evaluator.
Comparative evaluation must include downstream effects:
Compare resulting trajectories with and without the evaluation process under matched conditions.
Accurate classification alone is insufficient when evaluator interventions worsen downstream outcomes.
15. Counterfactual evaluation
Historical causal graphs record what occurred. Alignment analysis frequently also requires comparison with what might have occurred under alternative actions or interventions.
RETA therefore distinguishes realized history from counterfactual analysis.
For a realized event , one may compare
with hypothetical alternatives
while preserving the fact that only belongs to historical event memory.
Counterfactual analysis requires explicit model assumptions, uncertainty and provenance. Hypothetical outcomes remain separate from observations.
Intervention and prevention claims require an explicit counterfactual basis.
16. Stability in RETA is not immobility
A static system can be stable while being catastrophically misaligned. RETA therefore should not equate alignment stability with resistance to change.
The relevant concept is closer to bounded adaptive trajectory stability.
Let denote an admissible region of trajectory space under the active normative framework. A perturbation may temporarily move the system away from this region while the system subsequently returns toward the admissible region:
A perturbation can expose flawed admissibility assumptions. Restoration of an earlier state is not automatically justified.
The system must be capable both of resisting destructive deviations and of accepting beneficial change.
Perturbation testing must distinguish recovery from preservation of a flawed target.
17. Planned epistemic-memory layer
The current model records causal history and evaluation context. Richer epistemic memory remains proposed.
The proposed memory layer separates observations from revised interpretations.
Two forms of memory are therefore conceptually separated.
17.1 Episodic causal memory
The causal record contains what occurred:
Realized events extend the historical record:
Historical records may be corrected for data-integrity reasons, but corrections must themselves be versioned and auditable rather than silently replacing prior states.
17.2 Adaptive epistemic memory
The system also maintains revisable interpretations derived from experience:
Learning revises the interpretive layer:
The distinction is deliberate:
Immutable interpretations prevent learning; unrestricted historical rewriting defeats provenance.
18. Memory retrieval is itself a causal event
In RETA, memory cannot be treated as a passive database.
Retrieval changes the recipient’s information state and can affect subsequent decisions:
The selection, ranking, compression, omission, or framing of remembered information may influence subsequent action.
Memory retrieval can therefore be audited in the same way as other consequential events.
Proposed retrieval audits would address the following failure classes:
- memory poisoning;
- stale information dominating current evidence;
- selective retrieval;
- missing counterevidence;
- false source independence;
- summarization loss;
- context collapse;
- inappropriate retention;
- inappropriate forgetting;
- historical information being promoted from hypothesis to fact without justification.
A memory system can therefore contribute to misalignment even when every stored item is individually accurate, if retrieval systematically constructs a distorted information environment.
19. Planned information-literacy interface
The information-literacy extension is intended to attach epistemic structure to memories, observations, and claims.
A remembered claim may eventually be represented by an object such as
The proposed interface would distinguish hypotheses from verified claims, retain superseded interpretations and track dependence among apparently corroborating sources.
The information-literacy layer is therefore expected to govern not only storage but also acquisition, verification, synthesis, retrieval, communication, and revision.
The proposed interface supports evidence evaluation and revision.
20. Planned perturbation layer
A second planned extension introduces structured perturbations as experimental inputs.
A perturbation is a disturbance to expected state or trajectory, not necessarily harm.
A perturbation may be:
- harmful;
- beneficial;
- mixed;
- conditionally beneficial or harmful;
- unknown at occurrence time.
Robustness must remain distinguishable from resistance to beneficial novelty.
The system dynamics can be extended from
to
where is a perturbation and is the relevant epistemic-memory state.
Perturbations may be exogenous, endogenous, or reflexive. Examples include communication loss, agent arrival or departure, institutional turnover, model replacement, memory corruption, new evidence, regulatory change, resource shortage, unexpected cooperation, novel scientific knowledge, or changes in the topology connecting systems.
21. Perturbation vectors
RETA should not reduce perturbations immediately to a single severity score.
A perturbation can instead be described by a vector
where dimensions may include:
- : occurrence frequency under a declared reference interval;
- : immediate magnitude or destructiveness;
- : duration or persistence of effects;
- : physical, biological, organizational, or societal scale;
- : cascade or propagation potential;
- : reversibility;
- : observability;
- : uncertainty.
A scalar can assign similar values to rare severe disturbances and frequent minor disturbances. Distinct dimensions permit separate response criteria.
The precise perturbation ontology remains under development.
22. Scale and resolution
Perturbation analysis introduces an explicit scale variable into RETA.
Consequences can differ across resolutions and analytical boundaries.
Alignment evaluation should therefore eventually be represented as
where denotes analytical resolution or scale.
For example, the same biological event may be destructive to an individual cell, beneficial to an organism, and consequential to a population over a longer horizon. The normative sign cannot be propagated mechanically across levels.
Local optimization can transfer costs beyond the selected boundary. System-level assessment requires explicit affected-party and distribution assumptions.
Scale, boundary and horizon are coupled modeling choices.
23. Perturbation, memory, and adaptation
The combination of perturbation testing and epistemic memory provides a direct experimental object for adaptation.
Apply perturbation to a system with memory state :
Observed outcomes can update the proposed epistemic state:
Apply the same perturbation again:
If is better under the declared alignment criteria than , the system has demonstrated some form of adaptation.
However, repeated exposure to the identical perturbation does not establish generalization. A stronger test applies a structurally related but previously unseen perturbation :
Performance on provides evidence about whether the system learned a transferable failure mechanism rather than memorizing a case-specific response.
24. Defensive overfitting
Perturbation training can reward excessive conservatism. Evaluation must include beneficial change and unnecessary refusal.
Repeated hostile perturbations could induce policies equivalent to:
- reject unfamiliar agents;
- suppress disagreement;
- distrust novel information;
- centralize control;
- minimize exploration;
- classify uncertainty as threat.
Such a system may score well against a narrow adversarial benchmark while becoming less aligned in open environments.
Beneficial perturbations are therefore necessary controls.
The perturbation set should contain cases where the correct response requires acceptance, cooperation, learning, restructuring, or relinquishing a previously successful policy.
Robustness is therefore defined as preservation or improvement of alignment under change, not preservation of the pre-perturbation state.
25. Experimental use of RETA
RETA is intended to support explicit comparison among architectures rather than demonstration by anecdote.
A useful experimental program can compare systems such as:
and
Proposed tests require separate development and held-out perturbation distributions.
Potential measurements include:
- false-negative rate for trajectory failures;
- false-positive intervention rate;
- detection latency;
- cumulative trajectory harm or benefit under the active normative framework;
- recovery time after perturbation;
- calibration of uncertainty;
- causal-attribution error;
- recurrence of previously encountered failure mechanisms;
- transfer to unseen perturbation families;
- susceptibility to memory poisoning;
- failure to accept beneficial perturbations;
- evaluator-induced degradation;
- intervention cost;
- sensitivity to causal horizon and analytical boundary.
Matched comparisons require simpler output-, action-, policy- or agent-level baselines.
26. Falsifiable research questions
Research claims require possible disconfirming outcomes.
The framework therefore motivates questions such as:
RQ1 — Trajectory detection
Compare trajectory-level and local-output detection under matched information and cost.
RQ2 — Evaluator closure
Test detection of evaluator-induced failures with and without explicit evaluator modeling.
RQ3 — Horizon sensitivity
Does increasing causal horizon improve detection of delayed failures, and at what cost in false positives, computational complexity, and uncertainty?
RQ4 — Boundary sensitivity
How sensitive are alignment judgments to changes in the declared analytical boundary?
RQ5 — Causal attribution
Can event-DAG representations improve causal attribution relative to linear or single-parent histories in interacting multi-agent systems?
RQ6 — Epistemic memory
Does structured epistemic memory reduce recurrence of previously understood failure mechanisms without causing excessive defensive overfitting?
RQ7 — Perturbation robustness
Test transfer to structurally distinct perturbations outside the development distribution.
RQ8 — Beneficial novelty
Measure acceptance of beneficial change alongside protection against harmful change.
RQ9 — Recursive audit stability
Under repeated evaluator self-application, does the evaluation process converge, oscillate, diverge, or become path-dependent, and under what conditions?
The proposed questions remain empirically unresolved.
27. What RETA claims
The current RETA contribution is architectural and methodological.
The proposed representation combines:
- realized causal events are the atomic unit of local alignment analysis;
- causally connected trajectories are the system-level object of concern;
- outputs modify the state from which future outputs emerge;
- causal history is represented as a multi-parent graph rather than a sequence of independent decisions;
- alignment judgments are explicitly conditioned on boundary, horizon, uncertainty, and normative version;
- evaluators are represented as causal participants;
- evaluator outputs can themselves be audited;
- historical judgments are preserved prospectively across later model or normative revisions;
- causal description remains distinguishable from normative evaluation.
The architecture can be summarized as
Memory and perturbation extensions remain proposed.
28. What RETA does not currently claim
Unestablished claims include:
- the AI alignment problem has been solved;
- trajectory-level evaluation is superior to all existing alignment methods;
- the selected normative constitution is universally correct;
- causal ancestry can always be reconstructed reliably;
- long-horizon consequences can always be predicted;
- recursive auditing necessarily converges;
- an evaluator can be made unbiased merely by including the evaluator inside the causal graph;
- the current simulations reproduce the complexity of real social, ecological, institutional, or AI systems;
- synthetic robustness under selected perturbations guarantees real-world robustness;
- an aligned local trajectory guarantees alignment at every larger spatial or temporal scale.
The limitations delimit the research scope.
29. Candidate empirical contribution
Additional-value claims require comparative tests.
The central empirical question is:
Compare detection of consequential failures against matched local-evaluation baselines.
A favorable comparison requires improved failure detection with acceptable false positives, intervention costs and calibration.
A second empirical question concerns recursive closure:
Does modeling the evaluator as a causal participant improve the detection or prevention of evaluator-induced failures?
A third concerns adaptive robustness:
Test generalization without excessive refusal or dependence on historical cases.
The proposed representation makes the questions explicit; favorable outcomes remain unestablished.
30. Working formal summary
A compact representation of the current RETA research object is:
The equations are schematic interfaces, not a completed mathematical theory:
- bounded systems;
- realized events;
- causal ancestry;
- scale and resolution;
- consequence horizons;
- epistemic uncertainty;
- normative constraints;
- evaluator participation;
- memory;
- perturbations;
- adaptive state change.
The object of study is a causal trajectory generated by interacting events and changing states.
31. Current status
RETA remains a work in progress.
The current implementation establishes executable semantics for a subset of the architecture and provides a basis for reproducible synthetic experiments. The information-literacy memory layer, multiscale perturbation ontology, formal trajectory metrics, causal-attribution methods, convergence analysis, and empirical comparison against external baselines remain under active development.
Framework definition and successful code execution do not establish alignment, stability, robustness or superiority.
The purpose of the current work is to make a difficult class of alignment questions explicit enough to inspect, implement, perturb, falsify, and revise.