What Is Recursive Science®
A Scientific Framework for Runtime Intelligence and Evidence-Governed Computation
Recursive Science proposes a longitudinal science of runtime behavior. It begins from the premise that learned capability becomes organized behavior through time—and that prior activity can recursively change the conditions under which later behavior forms.
The framework investigates how continuity, coherence, drift, rigidity, instability, collapse, and recovery develop across an ordered runtime. Its object is not the isolated output or an inaccessible mental entity. It is the evidence-bound behavioral trajectory formed through interactions among models, people, tools, roles, constraints, and operational events.
Intelligence is expressed as structured motion through time, and prior activity recursively conditions the behavior that follows.
The foundational paper consolidates the scientific hierarchy developed across the Recursive Science core manuscripts and subsequent research into Runtime Intelligence, Chronodynamics, Computational Behavior Architecture, Runtime Evidence, and Evidence-Governed Computation.
Foundational Research Paper · September 2026 · Version 1.0
The Research Problem
Modern AI research provides essential accounts of how models are trained and how inference is computed. Learned parameters constrain what a model can generate. Attention, activations, tokenized context, retrieval, cache state, and decoding participate in producing each response.
Recursive Science does not reject or replace these mechanisms.
It addresses a different level of analysis.
A model used once produces an output. A model operating inside a conversation, agent loop, tool workflow, or multi-actor process participates in an evolving runtime.
During that runtime:
earlier outputs return through context;
tool results alter later possibilities;
people introduce corrections and constraints;
roles transfer or contest authority;
objectives persist, weaken, or change;
and local decisions accumulate into a trajectory.
These relationships cannot be adequately described by evaluating each output in isolation.
A final response may appear coherent while the preceding trajectory contains repeated correction failure, growing contradiction, role fragmentation, or accumulated displacement. A poor response may occur within a runtime that subsequently recovers. The scientific object therefore extends beyond the individual answer to the ordered development connecting events across time.
Recursive Science studies the organization that becomes visible when computation is examined as a developing runtime rather than a collection of disconnected outputs.
Mind as Motion
Mind as Motion™ is the organizing proposition of Recursive Science.
In its operational form, motion means ordered change within a declared behavioral coordinate system. The proposition does not describe intelligence as a physical fluid, hidden substance, or entity separate from computation. It directs scientific attention toward how observable organization develops from one event to the next.
The symbolic expression I = MM² serves as a mnemonic for this proposition. It is not presented as a dimensional equation, quantitative identity, or validated law of physics.
Its scientific value depends on whether the underlying proposition produces:
clearly defined variables;
observable distinctions;
reproducible measurements;
discriminating experiments;
and hypotheses that can fail.
Recursive Science therefore treats Mind as Motion as the beginning of an investigation—not as its conclusion.
The Scientific Object
The primary object of Recursive Science is the longitudinal runtime trajectory.
A runtime trajectory may include:
source events;
prompts and responses;
tool calls and results;
roles and authority relations;
objectives and constraints;
temporal dependencies;
behavioral signals;
regime conditions;
transition markers;
operational outcomes;
and evidence-quality states.
When these elements are qualified, ordered, and reconstructed under a declared method, they form an evidence-bound worldline.
The worldline is not a direct observation of hidden model state. It is a representation derived from available operational evidence. Its authority cannot exceed the coverage, quality, and provenance of that evidence.
This distinction makes the runtime scientifically investigable without claiming access to internal mechanisms that the record does not contain.
Four Foundational Commitments
Temporal Extension
Continuity, drift, regime formation, collapse, and recovery require an interval. They cannot be established from one output alone.
Runtime investigation must therefore preserve event order, temporal coordinates, dependencies, and evidence horizons.
Recursive Conditioning
Earlier outputs, tool results, corrections, role actions, and environmental changes may become conditions of later computation.
If prior activity matters, modifying or removing it should produce testable differences in the trajectories that follow.
Runtime Organization
The same trained model may express different organized behavior under different contexts, interaction histories, tools, participants, and sampled outputs.
Recursive Science concerns how capability becomes coordinated during a particular runtime. It does not claim that capability is created independently of training, architecture, retrieval, or external state.
Evidence Dependence
A description of runtime behavior becomes scientifically meaningful only when its source, transformation, method, uncertainty, and inferential limits can be inspected.
A compelling label is not evidence. An untraceable measurement cannot support a durable scientific claim.
Recursive Conditioning and Path Dependence
Recursive Science uses recursion in a specific operational sense:
Recursion occurs when the consequences of prior activity participate in constructing the conditions of later activity.
This relationship can operate across several scales:
Token scale: each generated token conditions what may follow.
Turn scale: a response becomes part of the next interaction context.
Agent and tool scale: plans, actions, observations, and tool results alter later decisions.
Multi-actor scale: people, models, tools, policies, and systems respond recursively to one another.
The relevant runtime is therefore better described as parameter-static and context-stateful than simply stateless. The model’s learned parameters may remain unchanged while the surrounding system continually reintroduces history, memory, summaries, tools, and prior outputs.
Path dependence is a testable consequence of this process. If two otherwise matched runs differ in an earlier correction, tool result, role instruction, or event order, their later behavior may diverge.
The research question is not merely whether divergence occurs. It is which earlier conditions matter, how long their influence persists, and whether the resulting difference exceeds ordinary stochastic variation.
Runtime Intelligence
Runtime Intelligence names the organized behavioral phenomenon expressed while a computational system operates.
It is not a claim that inference creates intelligence or capability from nothing. Training, architecture, retrieval, tools, and memory establish the resources and constraints available to the runtime.
Runtime Intelligence concerns how those resources become organized across a particular trajectory.
It becomes observable through patterns such as:
a constraint that persists across many turns;
a reasoning posture that survives disturbance;
a role configuration that repeatedly returns;
a correction that initially holds and later disappears;
an objective that gradually changes;
or a displacement that culminates in visible failure.
These patterns can be investigated without assigning consciousness, subjective experience, intent, or an enduring self to the system.
Runtime Intelligence occupies a level between low-level model mechanism and final task outcome: the level of organized behavior through operation.
Inference-Phase Dynamics
Inference-Phase Dynamics provides the first experimentally accessible domain of Recursive Science.
It studies the formation, persistence, deformation, transition, and loss of behavioral organization during inference and sustained interaction.
Its evidence may include:
prompts and responses;
system instructions when available;
timestamps and event order;
role labels;
tool activity;
corrections;
external state changes;
evaluator markers;
and operational outcomes.
From these records, investigators can examine whether behavior remained bounded, accumulated drift, entered a rigid configuration, approached a candidate boundary, crossed into another regime, or recovered beyond a temporary correction.
The domain begins with inference, but the broader framework may eventually be tested across other recursive computational and socio-technical systems. Transfer to human, animal, or biological behavior cannot be assumed; each domain requires its own definitions, ethical constraints, and independent evidence.
The Core Constructs
Recursive Science develops a language for describing runtime behavior through time.
Worldlines represent ordered, evidence-bound runtime trajectories.
Symbolic time represents consequential progression through events, dependencies, recurrences, and transitions.
Regimes describe persistent conditions of runtime organization.
Attractors describe recurring configurations toward which behavior appears to converge or return.
Containment describes persistence within a declared behavioral boundary.
Drift describes cumulative departure from a reference condition.
Curvature describes change in trajectory direction within a declared projection.
Pressure describes accumulated competing demands, contradictions, or unresolved constraints through defined signals.
Shear describes divergence between coupled runtime dimensions.
Basin Exit marks a qualified transition beyond previously sustained stability criteria.
Collapse describes persistent loss of established runtime organization.
Recovery requires sustained re-establishment or reorganization following instability.
Role topology represents changing relationships among people, models, tools, policies, and other participants.
These terms are analytical constructs. Where they are expressed mathematically, their coordinates, scales, methods, and uncertainty must be declared. Otherwise, they remain disciplined conceptual models rather than established physical quantities.
The Canonical Hierarchy
Recursive Science contains several connected layers that must not be treated as synonyms.
Mind as Motion
The foundational proposition that intelligence can be studied through the evolving organization of behavior.Recursive Science
The scientific framework investigating the structures and dynamics of that motion.Runtime Intelligence
The organized behavioral phenomenon that forms and changes during operation.Inference-Phase Dynamics
The first empirical domain for studying runtime formation, drift, regimes, collapse, and recovery.Computational Behavior Architecture
The structural architecture for representing trajectories, roles, measurements, regimes, and evidence-bearing runtime objects.Runtime Evidence
The object and discipline through which reconstruction becomes inspectable, reproducible, challengeable, and preservable.Evidence-Governed Computation
The authority law constraining instruments, projections, interpretations, and claims to a common source-bound evidence object.Fieldglass®
The reference implementation through which the integrated architecture becomes executable.
The conceptual dependency moves downward through this hierarchy. Evidence produced through implementation returns upward, where it can support, refine, or contradict the scientific propositions.
Implementation closes the experimental loop. It does not convert implementation into proof.
The Distinctive Contribution
Dynamical approaches to cognition have an established intellectual history. Context dependence, autoregressive conditioning, entropy, similarity, variance, and trajectory analysis are also not new by themselves.
Recursive Science does not depend on claiming otherwise.
Its distinctive contribution is the integration of:
longitudinal behavioral reconstruction;
symbolic and operational time;
role and interaction topology;
worldlines and regime analysis;
stability, drift, pressure, collapse, and recovery;
shared evidence authority;
bounded instrument contracts;
deterministic evidence artifacts;
explicit claim boundaries;
guided investigation;
replay;
and preservation.
These components form one continuous path from first-principles theory to externally inspectable computational evidence.
The originality of the program rests in this integrated architecture and the research object it makes available—not in asserting ownership over every underlying concept it employs.
Research Status
Recursive Science contains four kinds of contribution that must remain distinct:
Foundational propositions about studying intelligence as temporally organized behavior.
Theoretical constructs including symbolic time, worldlines, regimes, attractors, and the transient runtime substrate.
Implemented architecture including canonical records, evidence authorities, Passports, instrument contracts, reconstruction, and deterministic artifacts.
Empirical hypotheses that remain open to confirmation, revision, or rejection.
Fieldglass demonstrates that the computational architecture can be implemented and that a fixed source, configuration, and version can produce a reproducible evidence artifact.
It does not independently establish that every signal measures its intended construct, that every regime generalizes across systems, or that the proposed dynamics are universal.
Those claims require:
controlled perturbation;
baseline comparison;
stable negative cases;
ablation;
held-out evaluation;
prospective testing;
calibration;
cross-system transfer;
and independent replication.
What Would Validate the Program
The foundational paper establishes an initial testable agenda.
Recursive Science should produce evidence that:
earlier runtime events create persistent, measurable effects beyond sampling variation;
regime classifications describe sustained structure beyond arbitrary thresholds;
candidate boundaries precede independently defined changes under prospective criteria;
recovery can be distinguished from a one-turn correction;
role-aware reconstruction outperforms role-blind analysis where role structure matters;
qualified instability markers provide useful lead time at acceptable false-alarm rates;
selected measurements retain meaning across systems without unrestricted retuning;
and all projections derived from one evidence authority remain logically and numerically consistent.
Failure to meet these conditions would require the relevant constructs to be revised, narrowed, or rejected.
That possibility is essential. Recursive Science becomes scientific only to the extent that its propositions remain open to failure.
Claim Boundaries
The framework supports investigation of observable behavioral organization.
It does not, from output-derived evidence alone:
reveal hidden model state;
identify a unique internal mechanism;
establish intent;
prove consciousness or subjective identity;
determine root cause from chronology;
convert analytical geometry into literal physics;
or generalize automatically from AI records to human or biological cognition.
These boundaries are part of the framework itself.
They preserve a research object that can remain ambitious without depending on metaphysical agreement or inaccessible evidence.
From Foundation to Investigation
Recursive Science begins with a simple proposition:
Intelligence can be studied through the structured development of behavior across time, and prior activity can recursively alter what follows.
The wider research program turns that proposition into temporal coordinates, worldlines, regimes, measurement contracts, runtime evidence objects, instruments, and working infrastructure.
Its significance will not be established by the coherence of its vocabulary or the scale of its implementation alone. It will be determined by whether its constructs produce reliable, discriminating, and useful knowledge about runtime behavior beyond what simpler methods already provide.
The architecture and instrument now make that question testable.