Every Report We Write Makes the Next One Smarter

SBPI Ontology

The Structural Brand Power Index is formalized as an OWL 2 ontology with full provenance (W3C PROV-O), attestation tracking, and prediction modeling. Every score traces to its source, every prediction links to the evidence that generated it.

Namespace

@prefix sbpi:     <https://shurai.com/ontology/sbpi#> .
@prefix company:  <https://shurai.com/data/company/> .
@prefix week:     <https://shurai.com/data/week/> .
@prefix score:    <https://shurai.com/data/score/> .
@prefix signal:   <https://shurai.com/data/signal/> .
@prefix pred:     <https://shurai.com/data/prediction/> .
@prefix anno:     <https://shurai.com/data/annotation/> .

Core Classes

Five Scoring Dimensions

Tier System

*ReelShort at 84.05 is technically Strong tier but rounds up to Dominant threshold.

SHACL Validation

Seven node shapes enforce data integrity on every ETL load. Every company must have a name, slug (lowercase-hyphenated), and vertical assignment. Every score record must have exactly 5 dimension scores, a composite between 0-100, and a tier. Predictions must declare a direction from {up, down, stable, volatile} and a confidence between 0.0-1.0.

# Every score record must reference exactly one company, one week,
# have 5 dimension scores, composite 0-100, and a tier assignment
<https://shurai.com/shapes/ScoreRecordShape> a sh:NodeShape ;
    sh:targetClass sbpi:ScoreRecord ;
    sh:property [
        sh:path sbpi:compositeScore ;
        sh:datatype xsd:decimal ;
        sh:minCount 1 ;
        sh:maxCount 1 ;
        sh:minInclusive 0 ;
        sh:maxInclusive 100 ;
    ] ;
    sh:property [
        sh:path sbpi:hasDimensionScore ;
        sh:minCount 5 ;
        sh:maxCount 5 ;
    ] .

ETL Pipeline

Weekly SBPI state files (JSON) are converted to RDF triples via sbpi_to_rdf.py. The script creates Company, ScoreRecord, DimensionScore, Signal, and Attestation nodes for each company in each week, then links weeks in temporal chains for time-series queries.

Data Flow

# Input: state/current.json (17 companies, 5 dimensions each)
# + state/archive/W10-2026.json, W11-2026.json, ...

python3 etl/sbpi_to_rdf.py --all --validate

# [1/5] Loading ontology from ontology/sbpi.ttl
# [2/5] Processing current state
# [3/5] Processing 2 archived weeks
# [3.5] Linked 3 weeks in temporal chain
# [4/5] SHACL validation PASSED
# [5/5] Loaded 1,672 triples into Oxigraph

Triple Structure per Company-Week

Each company-week record generates approximately 15-20 triples:

# Company identity (6-8 triples)
company:reelshort a sbpi:Company ;
    sbpi:slug "reelshort" ;
    sbpi:companyName "ReelShort" ;
    sbpi:inVertical <.../vertical/micro-drama> ;
    sbpi:isPlatformGiant false .

# Score record (4-5 triples)
score:reelshort-W11-2026 a sbpi:ScoreRecord ;
    sbpi:forCompany company:reelshort ;
    sbpi:forWeek week:W11-2026 ;
    sbpi:compositeScore 84.05 ;
    sbpi:delta -0.55 ;
    sbpi:inTier sbpi:TierDominant .

# Dimension scores (5 x 2 = 10 triples)
score:reelshort-W11-2026-content_strength a sbpi:DimensionScore ;
    sbpi:forDimension sbpi:ContentStrength ;
    sbpi:dimensionValue 85 .

# Attestation (4 triples)
attest:reelshort-W11-2026 a sbpi:Attestation ;
    sbpi:confidence 0.85 ;
    sbpi:sourceType "expert_judgment" ;
    prov:wasGeneratedBy "ShurIQ SBPI Pipeline" .

SPARQL Query Library

Eleven queries cover the full analysis cycle, from basic movers to compound anomaly detection. Three run nightly, four run weekly, four are on-demand.

# Two consecutive weeks of same-direction delta = momentum signal
PREFIX sbpi: <https://shurai.com/ontology/sbpi#>

SELECT ?name ?w1Label ?delta1 ?w2Label ?delta2 ?momentum ?predDirection
WHERE {
    ?sr1 sbpi:forCompany ?c ;
         sbpi:forWeek ?w1 ;
         sbpi:delta ?delta1 .
    ?sr2 sbpi:forCompany ?c ;
         sbpi:forWeek ?w2 ;
         sbpi:delta ?delta2 .
    ?w1 sbpi:previousWeek ?w2 .           # temporal chain
    ?c sbpi:companyName ?name .

    FILTER((?delta1 > 0 && ?delta2 > 0) ||
           (?delta1 < 0 && ?delta2 < 0))

    BIND(?delta1 + ?delta2 AS ?momentum)
    BIND(IF(?momentum > 0, "BULLISH", "BEARISH") AS ?predDirection)
}
ORDER BY DESC(ABS(?momentum))

Prediction Experiment

Four methods run against the same data. When actuals load, every prediction is scored on directional accuracy, mean absolute error, and Brier score (confidence calibration). The experiment tracks whether the KG-augmented method outperforms statistical baselines as the graph grows.

Future methods (requiring API keys):

• LLM Zero-Shot: Company names only, no KG context
• LLM + KG: Full semantic layer context fed to LLM

Evaluation Metrics

W11-2026 Stack Ranking

Live data from the Oxigraph SPARQL endpoint. 17 companies tracked across the micro-drama vertical.

Predictive Signals (Nightly Digest)

Companies showing 2+ consecutive weeks of same-direction momentum. These are SPARQL-derived, not statistical noise.

Dimension Radar

Five-axis comparison across the top companies. Select companies to compare their structural profiles.

Prediction Experiment Status

Parametric vs. Non-Parametric Knowledge

When an LLM "knows" that ReelShort dominates the micro-drama vertical, that knowledge lives inside trillions of numerical weights trained on internet-scale data. This is parametric knowledge: static, probabilistic, impossible to audit, and expensive to update (retraining costs millions). It hallucinates because it's guessing what token comes next, not retrieving verified facts.

The SBPI semantic layer stores the same intelligence as non-parametric knowledge: explicit RDF triples in an Oxigraph store. Every fact is verifiable, traceable to its attestation source, and updatable in milliseconds. When we say ReelShort scored 84.05 in W11-2026, we can show you the triple, the attestation, the SHACL shape that validated it, and the SPARQL query that retrieved it.

Why Domain Density Wins

The Compounding Asset Model

Every intelligence briefing we produce is simultaneously revenue and CapEx. The $5K-25K engagement pays for the work. The triples generated during that work become permanent additions to a proprietary knowledge asset. After 52 weeks of weekly reporting across 3 verticals, the graph contains ~15,000 attestation-backed facts about competitive dynamics that no other system on Earth has.

Where We Are

The Revenue-IP Flywheel

What Capital Unlocks

OTK Lineage

This system consciously inherits from the EU's On-To-Knowledge project (2001-2003), which spent over $1M building an ontology-driven knowledge management pipeline. The components map directly:

The OntoRAG project (2025) independently arrived at a nearly identical architecture. Their philosophy: "If the system cannot explain what it knows, where it comes from, and why it changed, it is not a knowledge system." We agree. Every triple in this graph traces to its attestation.

The IP Engine Variables

To model the knowledge graph as an investable asset, we quantify four variables that determine whether the graph compounds into a structural advantage or stalls as a glorified spreadsheet.

Use of Funds — The Force Multiplier

Investors want to see that $1 into ShurIQ doesn't just pay for a consultant's time — it builds a flywheel. Every investment area produces both immediate deliverables (revenue) and a durable, scalable IP asset.

Capital Deployment Map

The Karpathy Auto-Research Cycle

Andrej Karpathy's "Software 2.0" insight: the most valuable code is not the code humans write — it's the data the system accumulates through iterative cycles of extraction, evaluation, and refinement. We apply this directly to ontology discovery.

This cycle already runs in a simplified form. The nightly scheduler extracts signals, the SPARQL queries cross-reference against historical data, and the prediction experiment stack-ranks which signals have predictive value. The human finalization step happens when analysts review the weekly digest and decide which signals to act on. As the graph grows, the automated steps become more accurate and the human becomes a curator rather than a creator.

Measuring the Thesis

The prediction experiment is designed to produce the specific evidence an investor needs. Each week, four methods predict the same 17 companies. When actuals load, we score them all on the same metrics. Over time, the data answers three questions:

Experiment Status

First evaluation runs when W12-2026 actuals are loaded via python3 etl/sbpi_to_rdf.py.

The Autopoietic Cycle

The system improves itself through use. This is not a metaphor — it is a measurable property of the architecture.

This is what separates the system from a data warehouse. A warehouse stores records. This system stores records and the schema for interpreting them, and the validation shapes for ensuring quality, and the prediction methods for testing whether the stored knowledge has real-world predictive value. The value comes from how the sources are organized — InfraNodus entity extraction, wikilink relationships, gap analysis, ontological bridging — not just from the sources themselves.

Concept Network

Force-directed layout of the SBPI knowledge graph analyzed via InfraNodus. 150 nodes, 700 edges, 8 clusters. Node size reflects degree centrality. Edge opacity reflects relationship weight. The five gateway nodes (data, prediction, knowledge, graph, schema) bridge multiple clusters and control information flow through the network.

Structural Gaps

Three structural gaps identified in the concept network. These represent missing conceptual bridges between clusters that, if filled, would strengthen the overall intelligence framework. Each gap is a content or research opportunity.

Cluster Composition

Eight topical clusters ranked by betweenness influence. The Predictive Accuracy cluster dominates at 39% of network influence, followed by Knowledge Density at 30%. Together these two clusters control nearly 70% of the network's information flow.

Graph Statistics