The Auto-Research Engine

The Ontology as IP

The SBPI ontology (sbpi.ttl) is not a generic schema. It is a domain-specific model of how competitive brand power works in entertainment verticals. The dimension weights encode expert judgment that took months of client engagements to calibrate.

Five Scoring Dimensions

Distribution Power carries the highest weight (25%) because in a mobile-first vertical, app store presence and partnership reach determine whether content reaches audiences. Monetization Infrastructure carries the lowest (15%) because several dominant players subsidize the vertical from existing revenue streams.

Performance Tiers

Validation: SHACL Shapes

Every data load runs through SHACL shape validation. Bad data fails before entering the store. This is schema enforcement, not error handling.

Attestation Layer

Every score carries provenance metadata tracking confidence and source quality:

Attestation:
    confidence: 0.85         # 0.0-1.0 scale
    sourceType: "expert_judgment"
    # Options: primary_data, secondary_analysis, expert_judgment, automated_inference

The attestation upgrade engine progressively improves scores as evidence quality increases. Signal URLs upgrade confidence from 0.85 to 0.90. Scoring rationales add a second attestation at 0.95. The system tracks not just what it knows, but how well it knows it.

The Karpathy Auto-Research Pattern

Andrej Karpathy described the future of AI as "auto-research" — headless AI agents that continuously process information, extract structured knowledge, and produce insights without human prompting. We implemented this literally.

Headless Batch Processing

The batch-mapupdate.sh script uses claude -p (Claude's headless CLI mode) to process session records. Each session takes ~20 seconds:

1. claude -p reads session file + extracts entity statements (text only)
2. curl POSTs entities to InfraNodus REST API
3. curl GETs full graph stats (nodes, edges, clusters, gaps)
4. Shell appends MOC entry with parsed results
5. Dashboard regenerated and deployed to Cloudflare Pages

27 sessions processed in a single batch run. 26 succeeded, 1 failed (entity extraction produced raw output instead of parsed entities). 96% success rate. No human interaction required.

Nightly Insights Runner

The nightly-insights.py scheduler runs 7 SPARQL queries daily at 6:13 AM:

W12-2026 Forecast

9 momentum predictions generated by the SBPI Prediction Engine. Confidence scores derived from a documented, reproducible formula.

Momentum Signals

Confidence Scoring

Every prediction carries a confidence score computed from a documented formula. No black box.

base = 0.60  # minimum: 2 consecutive weeks, same direction

Adjustments:
  +0.10 if avg |delta| > 3.0    # strong movement
  +0.10 if avg |delta| > 5.0    # very strong (cumulative with above)
  +0.05 if both |delta| > 2.0   # consistent magnitude

Cap: 0.95  # never assert certainty from 2 weeks of data

Stagnation Detection

15 companies show zero movement for 2+ consecutive weeks — the stagnation signal that identifies non-competitive tail entries. These companies are either pre-launch, resource-constrained, or have exited the competitive frame.

The Accuracy Experiment

A controlled comparison framework is built and awaiting the first evaluation cycle:

The hypothesis: KG-augmented predictions outperform statistical baselines because the ontology encodes structural information (dimension weights, tier boundaries, company categories) that pure statistical methods cannot access.

The Scaling Path

Knowledge graphs follow a power-law value curve. Below 1,000 facts, a graph is a reference document. At 10,000-100,000 facts, it becomes a reasoning engine. At 1,000,000+ facts, it becomes a prediction platform.

Each row is architecturally supported today. The ETL pipeline ingests any JSON state file. The ontology extends to any vertical by adding instances, not changing the schema. The SPARQL queries generalize — weekly-movers.rq works identically whether tracking 22 companies or 22,000.

Why the IP Is Defensible

Capital Deployment Map

Total: $70K-120K to reach 100K+ triples and prove the prediction accuracy thesis.

The Self-Improving Cycle

The system is autopoietic — it improves itself through use.

Evidence From Production

• Long Zhu → Produced the Layered Ontology Architecture (K1-K3 + O1-O3). Didn't exist before we needed it.
• FrameBright → Validated the two-site editorial pattern. Now deploys in hours instead of days.
• Fiserv → Proved Brand Power Score scales to corporate brands without schema changes.
• MicroCo (SBPI) → First production semantic layer with SPARQL-queryable predictions.
• Batch Pipeline → 27 sessions processed autonomously. System writes its own history into structured knowledge.

Parametric vs. Non-Parametric Knowledge

This is where the billion-node thesis comes from. The distinction between what an LLM "knows" and what a knowledge graph knows is the foundation of the entire business model.

Why Domain Density Changes the Math

A general-purpose knowledge graph like Wikidata has ~100 billion triples but knows approximately nothing about micro-drama competitive dynamics. GPT-4 has seen the Wikipedia page for ReelShort but cannot tell you that ReelShort's Distribution Power score dropped 0.55 points last week, or that JioHotstar's momentum is +11.0 at 85% confidence.

Our 1,672 triples contain more actionable intelligence about the micro-drama vertical than the entire parametric knowledge of any LLM. That's the domain density argument: a small graph with expert-designed ontology beats a large model with generic training in any task that requires structured reasoning over specific facts.

The IP Engine Model

Every intelligence briefing is not just a deliverable. It is a CapEx investment into a proprietary asset. The revenue pays for the construction of a knowledge graph that will eventually be worth millions in recurring licensing.

The Variables

Dual Revenue: Service + Asset Accumulation

Every $2K-10K engagement produces two things simultaneously:

The investor pitch: you're not paying for consultants' time. You're funding the construction of a proprietary database that will eventually be worth millions in recurring licensing. The consulting revenue subsidizes the R&D.

The Karpathy Auto-Research Loop

Andrej Karpathy's "Software 2.0" thesis applied to ontology discovery. Agents don't just extract — they propose, critique, rank, and refine the knowledge graph structure itself.

What We've Built So Far

Steps 1-2 are operational today. The batch-mapupdate pipeline runs headless claude -p agents that read session records and extract entity statements. The InfraNodus API processes these into graph nodes with cluster analysis and gap detection.

Steps 3-4 are partially implemented through SHACL validation (catches schema violations) and the prediction experiment (ranks signal value by accuracy over time).

Step 5 is operational through the wrapup skill and session capture pipeline — human review of extracted entities before they become permanent graph entries.

The Use of Funds Flywheel

Each dollar invested in ShurIQ doesn't pay for a consultant's time. It builds a flywheel where service revenue funds IP accumulation, and the IP makes each subsequent service faster and more valuable.

The Compounding Math

At current scale (1,672 triples from ~$55K in engagement work), each triple costs ~$33. At the automated batch processing rate (20 seconds per session, ~$0.02 per extracted triple), the marginal cost drops 1,650x. The fixed cost is ontology design. The variable cost approaches zero.

From Service to Platform

The Experiment That Proves It

The prediction accuracy experiment is the mechanism that converts this thesis from an argument into evidence.

The hypothesis chain:

1. KG-Augmented > Statistical Baselines → Proves ontology design has predictive value
2. LLM + KG > LLM Zero-Shot → Proves non-parametric knowledge amplifies LLM reasoning
3. Accuracy improves as graph grows → Proves the compounding thesis

If all three hold, the investment thesis is proven: every dollar into the knowledge graph produces exponentially more predictive value over time.