Sigma Runtime Standard - Public Specification Notice
Specification License: CC BY 4.0.
Independent implementation is permitted under the public SRS/SRIP terms.
Machine-readable artifacts: Apache License 2.0 where explicitly marked.
Proprietary Sigma Runtime assets and Sigma marks are not licensed by this document.
| Field | Value |
|---|---|
| Information Class | Derived-Public |
| Source Context | Historical and current proprietary runtime implementation context |
| Sanitization | Historical public configuration and method identifiers only; no prompts, private data, deployment, or operational paths |
This document is non-normative. It preserves one historical implementation
shape for SRIP-11 without making its libraries, embedding dimensions, thresholds,
or storage layout requirements of the backend-agnostic public standard.
The retired CMT architecture alias is retained in historical SRIP-11 lineage.
New references use the canonical SMC name. A bounded implementation may keep
legacy wire fields while it migrates, but those fields must not define a second
architecture namespace.
| Surface | Canonical Behavior | Compatibility Behavior |
|---|---|---|
| Configuration input | smc |
A lone cmt block may be accepted and normalized before inheritance or runtime use. |
| Effective configuration | Emit one smc block |
A document containing both smc and cmt is invalid. |
| Persistence and API | Versioned implementation fields | Existing names such as cmt_enabled may remain legacy wire aliases until a separately versioned schema migration. |
| Diagnostic payloads | SMC terminology | Existing cmt payload keys may remain compatibility fields when explicitly documented. |
Accepting a legacy field does not make CMT a second public SRIP name.
| Surface | Reference Choice | Normative Status |
|---|---|---|
| Embedding | MiniLM-compatible 384-dimensional vector | Example only |
| Graph store | NetworkX directed graph | Example only |
| Vector store | FAISS index | Example only |
| Rib Point interval | 10-50 cycles | Calibration example |
| Cluster size | 5-10 Rib Points | Calibration example |
| Anchor buffer | First 8 cycles | Calibration example |
| Fact extraction | First 3 cycles, then every 10 cycles | Calibration example |
| Semantic-loss trigger | 0.25 |
Historical threshold; not universal |
The following examples preserve the concrete defaults formerly embedded in the
canonical SRIP. They describe one historical profile and are not conformance
requirements.
smc:
anchor_buffer:
enabled: true
cycles: 8
fact_extraction:
enabled: true
interval: 10
early_cycles: 3
min_confidence: 0.7
max_facts: 30
domain_adapter: null
Historical adapters included conversational, healthcare, defense, business,
education, legal, and technical extraction profiles. Those names do not create
normative SMC domains.
One identity-specific runtime example used a non-prompt behavior block:
behavior:
compression:
list_tolerance: 0.6
token_policy:
base_limit: 900
min_limit: 400
max_limit: 1200
The historical behavior interpreted list_tolerance >= 0.8 as no list penalty,
scaled lower values by 1 - list_tolerance, and allowed base_limit to override
the default completion limit. These choices are not part of the SMC contract.
One implementation exposed compression feedback through methods named
AEPController.monitor_compression() and
AEPController.trigger_phase_regeneration(). Semantic loss above 0.25 could
queue temporary expansion under that profile. These identifiers and the trigger
are historical implementation evidence, not required interfaces or universal
control semantics.
The same profile used the illustrative retention formula:
retention_weight = (density * coherence) / max(0.01, entropy)
Implementations must not reuse this formula without declaring compatible units,
windows, normalization, and calibration evidence.
Illustrative structures:
@dataclass
class RibPoint:
id: str
cycle_range: tuple[int, int]
vector: list[float] # 384 dimensions in this profile only
summary: str
phase: str
density: float
entropy: float
lineage: list[str]
# Historical profile only:
# retention_weight = (density * (1 - entropy)) / max(0.01, entropy)
@dataclass
class Cluster:
id: str
rib_points: list[str]
centroid: list[float]
theme: str
phase_distribution: dict[str, int]
cohesion: float
Equivalent implementations may use different embedding models, dimensions,
graph stores, vector stores, schedules, and retrieval engines if they preserve
the normative SRIP-11 boundaries, lineage, validity, and observability.
Historical fact-category examples included identity, context, constraints,
relationships, history, commitments, and preferences. Historical rendering
grouped known facts by category before model-visible recall. Both the category
set and rendering format are implementation-profile choices.
The statement 10x reduction in token footprint for sessions above 500 cycles
is a historical target, not a public conformance result. It must not be presented
as a benchmarked outcome without a dataset, baseline, estimator, run manifest,
and evidence reference.