Sigma Stratum Documentation – License Notice
This document is part of the Sigma Runtime Standard (SRS) and the
Sigma Stratum Documentation Set (SRD).It is licensed under Creative Commons Attribution–NonCommercial 4.0
(CC BY-NC 4.0).The license for this specific document is authoritative.
For the full framework, see/legal/IP-Policy.
In Sigma Runtime, an attractor is a stabilized pattern of meaning, continuity, and behavioral orientation that persists across turns or cycles.
Publicly, attractors should be understood as a way of describing how interaction becomes coherent over time, not as hidden characters or unrestricted autonomous entities.
An attractor is a recurrent configuration inside the active interaction field.
It can preserve:
Attractors therefore explain why a runtime can remain coherent over extended interaction instead of behaving like unrelated single-turn completions.
Publicly, attractors should not be interpreted as:
An attractor is a bounded control-and-continuity concept inside the runtime, not an excuse to suspend product or safety boundaries.
Attractors are useful because they help the runtime:
At a public level, they are one of the main ways Sigma Runtime explains continuity without reducing everything to plain transcript replay.
Publicly, it is sufficient to distinguish several broad attractor roles:
| Class | Public role |
|---|---|
| Coherence-stabilizing attractor | Holds interaction together around continuity, identity, and bounded intent. |
| Exploratory attractor | Supports extension, variation, and controlled expansion into new themes or motifs. |
| Reflective attractor | Supports evaluation, meta-interpretation, and self-correction under pressure. |
| Transitional attractor | Bridges one stable field configuration to another without abrupt collapse. |
| Containment-oriented attractor | Narrows symbolic amplitude and helps isolate unstable regions during recovery. |
This is an explanatory taxonomy, not a promise that public users can directly select or force every attractor mode.
Attractors are not static.
They typically pass through a bounded lifecycle:
The important public point is that persistence is conditional, not absolute.
Attractors do not operate outside the rest of the runtime.
They remain bounded by:
This means Sigma Runtime does not simply reinforce every recurrent pattern.
It differentiates between helpful continuity and destabilizing persistence.
Attractor language is only useful if the system also recognizes its failure modes.
Publicly, the main ones are:
These are the reasons attractors must stay under bounded runtime control.
Attractors should not be understood as permanent traps. A stable runtime also
needs a controlled way to introduce variation when the current field becomes
too repetitive, over-converged, or locally optimized.
SRIP-15 defines this as controlled perturbation: a bounded deviation from the
current attractor that explores an alternative trajectory without suspending
safety, identity, drift, density, or memory constraints.
Publicly, controlled perturbation means:
Perturbation can be internal, such as reframing or motif re-weighting, or it can
use SRIP-14 retrieval as an exploratory signal. In both cases, the perturbation
remains subordinate to the runtime control stack.
SRIP-10 AEP adds a public way to reason about whether attractor persistence is
still adaptive.
In attractor terms, AEP helps distinguish:
AEP is not a mechanism for arbitrary novelty or uncontrolled style churn. Its
role is to provide bounded evidence about when the attractor field may need to
soften, narrow, recover, or transition.
SRIP-16 adds a bounded self-modeling view over attractor evolution.
RSM may observe when an attractor becomes too rigid, too diffuse, or too
self-referential, and may emit reflective evidence for the control layer.
This does not create a hidden agent or independent attractor authority.
The self-modeling layer records how the runtime is changing; it does not decide
by itself that a change should be persisted, amplified, or exposed to the user.
At a public level, attractor analysis is valuable when it helps answer questions like:
Exact internal telemetry may evolve, but the explanatory role remains the same:
attractor analysis helps distinguish stable continuity from unstable amplification.
References:
Tsaliev, E. (2025). Attractor Architectures in LLM-Mediated Cognitive Fields — DOI: 10.5281/zenodo.17629926