13 Days of Nightly Pipeline Operation
The SBPI autoresearch pipeline has run nightly since March 21. Status, findings, and open issues below.
On March 21, we activated the SBPI Semantic Layer nightly pipeline. Every morning at 6:13 AM, an autonomous Claude Code session spins up, starts the Oxigraph triple store, runs SPARQL queries across the knowledge graph, evaluates prediction models against actuals, and writes a timestamped intelligence digest.
The pipeline tracks 21 companies across the vertical drama / micro-drama market. Three weeks of longitudinal SBPI data (W10 through W12 2026) feed into four competing prediction methods. An Optuna-powered optimizer has already discovered a configuration that lifts directional accuracy from 23.5% to 69.9%.
Two new market entrants were detected automatically: Google/100 Zeros (SBPI 63.65) and Holywater / My Drama (SBPI 61.65). Both entered the Strong tier on first scoring. The pipeline flagged both in the W12 predictive signals digest before any human analyst reviewed them.
How the Pipeline Works
A 4-step autonomous cycle that runs every morning at 6:13 AM.
Port 7878
Insight digest
21 companies
12 parameters
Oxigraph
RocksDB-backed RDF triple store. Currently holds 1,672 triples covering 21 companies across W10–W12 2026. SBPI ontology defines 30+ RDF classes and 50+ properties. SHACL shapes enforce validation on every data load.
SPARQL Insights
11 query types: weekly movers, tier transitions, dimension anomalies, cross-correlations, predictive signals, attestation coverage, platform-vs-pureplay, temporal momentum, and more. Output is a timestamped markdown digest.
Prediction Experiment
Four competing methods (persistence, naive momentum, mean reversion, KG-augmented) generate directional predictions for all 21 companies. Evaluated against actuals weekly. W13 predictions are locked and waiting for actuals.
KG Interface Optimizer
Optuna TPE optimizer tunes 12 parameters controlling how knowledge graph signals translate into predictions. 30 trials complete. Best config achieves 69.9% directional accuracy vs. 23.5% baseline.
Ontology & Knowledge Graph
The SBPI ontology (sbpi.ttl) defines five scoring dimensions with explicit weights:
| Dimension | Weight | Measures |
|---|---|---|
| Distribution Power | 25% | Platform reach, app store position, geographic coverage |
| Content Strength | 20% | Production volume, IP depth, format innovation |
| Narrative Ownership | 20% | Market narrative control, media positioning, brand story |
| Community Strength | 20% | MAU, engagement metrics, retention, creator ecosystem |
| Monetization Infrastructure | 15% | Revenue model, profitability, pricing power |
ETL scripts handle the full data lifecycle: sbpi_to_rdf.py converts weekly JSON state to RDF triples. infranodus_bridge.py connects InfraNodus knowledge graphs to the RDF store. attestation_upgrade.py progressively improves confidence scores as evidence quality increases. event_impact_analyzer.py processes news and event signals.
Market Movements
W12 movers, dimension anomalies, and new entrants detected by the nightly pipeline.
Current Rankings
| Company | SBPI | Delta | Tier | Key Signal |
|---|---|---|---|---|
| DramaBox | 82.75 | +4.0 | Dominant | Disney Accelerator, $500M valuation, only profitable pure-play ($10M net) |
| ReelShort | 82.0 | -2.05 | Dominant | Head of Production defected to GammaTime. ShortMax surpassing on Google Play |
| Disney | 76.55 | +2.3 | Dominant | Locker Diaries #1 US. D+ vertical feed confirmed. DramaBox Accelerator bet |
| iQiyi | 65.7 | +1.2 | Strong | China crosses trillion-yuan threshold. AI dramas at 38% of output |
| Google/100 Zeros | 63.65 | — | Strong | NEW. Range Media Partners slate. Google TV first window |
| Netflix | 62.8 | -2.0 | Strong | Mobile redesign without content. Awareness-commitment gap widening |
| JioHotstar | 62.25 | +3.95 | Strong | IPL 2026 launch. 100 microdramas in 7 languages. 300M subscriber base |
| Holywater / My Drama | 61.65 | — | Strong | NEW. $22M Series A. Fox equity. Revenue tripled in 2025 |
| GoodShort | 58.8 | +1.7 | Strong | #3 US app. Most capital-efficient producer ($160-200K/series) |
| CandyJar | 58.65 | 0 | Strong | AI-powered content selection from 7M+ reader base |
| ShortMax | 56.3 | +1.3 | Strong | 100M downloads. 3,888% growth. Now #3 entertainment on Google Play |
| Lifetime/A&E | 54.1 | +2.8 | Strong | Established IP library activating for short-form |
| Amazon | 50.2 | -2.6 | Emerging | Downgraded from Strong. Zero microdrama strategy. Distribution without content |
| Viu | 48.15 | -1.85 | Emerging | DramaBox SE Asia expansion compressing core territory |
| COL/BeLive | 44.55 | +3.15 | Emerging | FILMART execution. 1,700 catalogue. SaaS model proving out |
| KLIP | 22.35 | -2.65 | Vulnerable | Structurally squeezed by JioHotstar's 100-title IPL slate |
| Mansa | 19.35 | +1.85 | Vulnerable | Africa-first positioning. Original series launch |
Detected This Cycle
New Google / 100 Zeros — SBPI 63.65
Range Media Partners partnership announced March 12. Google TV mobile app launched dedicated microdrama offering in US. Slate includes Mike Fleiss (Bachelor), McG, Simon Fuller, Kenan Thompson. First window on Google TV. Entry strategy combines platform distribution with established IP talent, unlike pure-play competitors.
New Holywater / My Drama — SBPI 61.65
$22M Series A at $200-250M valuation. Fox equity stake + 200-title commitment. Dhar Mann Studios 40-title deal. Maksim Chmerkovskiy starring. 55M lifetime downloads. Revenue tripled in 2025. Combines social media distribution with traditional entertainment talent deals.
Dimension Anomalies
Companies where individual dimension scores diverge sharply from the composite.
COL/BeLive
MI score of 92 reflects SaaS licensing model validated at FILMART (1,700-title catalogue). Narrative (35) and Content (28) scores pull the composite down.
Amazon
Distribution Power at 80 but Content Strategy at 22 and Narrative at 25. Only major platform with no microdrama content strategy as of W12. (Note: Amazon launched Fatafat in W13.)
Netflix
Mobile redesign signaled awareness of the vertical short-form market. No microdrama production has followed. Content Strategy at 26 has not moved in 3 weeks despite the composite holding at 60.8.
Disney
Narrative Ownership at 93 (highest in the dataset) vs. Content Strength at 55. The 38-point gap between narrative positioning and content execution output is the widest among Tier 1-2 companies.
Prediction Accuracy
Four methods evaluated. Mean reversion at 47.1% is the only method above random (23.5%). KG-augmented matches baseline due to untuned default parameters.
W11 → W12 Evaluation (17 companies)
| Method | Dir. Accuracy | MAE | Brier Score | Assessment |
|---|---|---|---|---|
| persistence | 23.5% | 1.80 | 0.250 | Baseline. Predicted "stable" for everything |
| naive_momentum | 23.5% | 1.80 | 0.279 | = baseline. Needs 2+ weeks same-direction to fire |
| mean_reversion | 47.1% | 2.11 | 0.250 | Best naive method. Correctly called 8/17 companies |
| kg_augmented | 23.5% | 1.80 | 0.250 | = persistence. Graph signal not translating yet |
Why KG-Augmented Underperforms
The knowledge graph contains 1,672 triples covering dimension anomalies, event catalysts, and tier transitions. The default interface configuration treats everything as "stable" because the direction_threshold is set at 0.5, which causes small noise to trigger false stability calls. The Experiment 2 optimizer found that raising this to 1.295 and enabling anomaly signals lifts accuracy to 69.9%.
Predictions Locked
84 predictions generated for 21 companies across 4 methods. Recorded on April 2 at 06:53 UTC. Waiting for W13 actuals to evaluate.
Directional Outlook (Momentum Signals)
| Direction | Companies | Combined Signal |
|---|---|---|
| Bullish | JioHotstar, COL/BeLive, Disney, DramaBox, GoodShort, Lifetime/A&E | 2+ week positive momentum |
| Bearish | Amazon, Netflix, ReelShort | 2+ week negative momentum |
| Neutral | CandyJar, ShortMax, iQiyi, Viu, KLIP, Mansa, + 6 others | Insufficient signal or conflicting indicators |
KG-LLM Interface Optimizer
30 Optuna TPE trials found a 12-parameter configuration lifting directional accuracy from 23.5% to 69.9% on training data.
Trial Performance Distribution
Hover bars for exact scores. Best trial highlighted.
Parameter Shifts Driving the Gain
| Parameter | Default | Optimized | Change | What It Does |
|---|---|---|---|---|
| direction_threshold | 0.50 | 1.295 | +159% | Raises the bar before committing to a directional call. Ignores small noise |
| mean_reversion_rate | 0.10 | 0.257 | +157% | Stronger pull toward tier midpoints. Aligns with observed W11-W12 patterns |
| anomaly_contributes | false | true | Enabled | Dimension anomalies (like COL's MI gap) now feed into predictions |
| divergence_weight | 0.00 | 0.180 | New signal | Dimension-composite divergence as predictor. First structural KG feature |
| tier_proximity_weight | 0.00 | 0.096 | New signal | Proximity to tier boundaries. Companies near edges are more volatile |
| confidence_base | 0.60 | 0.443 | -26% | More calibrated. Lower base confidence means less overconfident predictions |
KG Signal Extraction
The gap between naive KG-augmented (23.5%) and optimized KG (69.9%) is a +46.3 percentage point lift from the same underlying data. The default interface parameters were too conservative to extract existing graph signals. Parameter tuning produced the gain, not additional data.
This is training-set performance. Out-of-sample validation on W13 actuals is the next test.
Accuracy Comparison Across All Methods
13 Days of Nightly Insights
Every run since activation on March 21. The pipeline runs at 6:13 AM daily, generating SPARQL insight digests, recording predictions, and reporting optimizer status.
Full Pipeline Clean Run
Oxigraph on :7878, 1,672 triples. DramaBox +4.0 largest mover. Google/100 Zeros and Holywater / My Drama flagged as new entries. W13 predictions recorded for 21 companies. Optimized config reports 69.9% but Optuna not installed for live re-optimization.
Full Pipeline — Infrastructure Note
Oxigraph started fresh (was not running). All 4 steps completed. Same core signals: DramaBox +4.0, JioHotstar +3.95, ReelShort -2.05 eroding. COL/BeLive MI anomaly persists at +47.5 gap. No new optimizer data since March 25.
Max Turns Reached
Pipeline hit the 20-turn safety limit. Partial results generated. Insights written but experiment steps may have been truncated.
Recovery Run — store_client.py Restored
Discovered store_client.py was missing from working tree. Restored from git history (commit 0e304b2). SPARQL query files also missing — Step 1 produced empty sections. Predictions and optimizer steps ran normally. Action item: recreate query .rq files.
Max Turns Reached
Pipeline hit safety limit. Partial output.
Full Pipeline — Best Signal Day
Most signals per run of the cycle. DramaBox "category-defining" at $500M valuation. JioHotstar IPL converts planning to execution. Amazon downgraded Tier 2 to Tier 3. ReelShort talent exodus confirmed. COL/BeLive MI anomaly at +47.5 identified as core prediction target.
Empty Result
Oxigraph query returned empty. Possible cold start issue or configuration gap.
Optimizer Ran Live
Optuna TPE completed 30 trials. Best score: 69.86% directional accuracy. Optimized 12 parameters stored to best-config.json. Credit balance warning appeared. Also ran the MicroCo sources Slack notification task at 12:00.
Full Pipeline — 2,588 Triples
Store had higher triple count (2,588 vs later 1,672 — likely a reload/rebuild event between runs). Full movers digest with bullish/bearish signals. Amazon -5.8 and Netflix -5.0 combined bearish momentum. JioHotstar +9.45 strongest bullish signal.
Credit Balance Issue
Pipeline started but hit "Credit balance is too low" early. Minimal output.
Double Run (2 files)
Two insight files generated: 141356 and 141959. Likely a retry after initial issue. Both produced results.
First Run — Pipeline Activated
Scheduler registered. First nightly insight generated at 06:13. Python3 permission issue flagged — requires manual approval or settings update. Pipeline architecture established.
Reliability Summary
- Clean runs: 7 of 13 (Apr 2, Apr 1, Mar 28, Mar 25, Mar 24, Mar 22, Mar 22)
- Partial / max-turns: 3 (Mar 31, Mar 30, Mar 29)
- Failures: 3 (Mar 27 empty, Mar 23 credit balance, Mar 21 permission)
- Effective uptime: ~77% producing usable output
What Needs to Happen Next
Open actions ranked by priority and blocking status.
Install Optuna
The optimizer has been running in report-only mode since March 25. pip install optuna enables live re-optimization when new weekly data loads.
Deploy Optimized Config
The 69.9% accuracy config sits in best-config.json but hasn't been applied to the live prediction engine. W13 predictions used the old 23.5% interface. Run kg_interface_optimizer.py --apply.
W13 Out-of-Sample Test
69.9% is training-set accuracy. The real test is W13 actuals, available next week. If out-of-sample accuracy holds above 50%, the optimized KG interface is validated.
Score Google & Holywater / My Drama
Both new entries have SBPI composites but no deltas (first appearance). Full dimension scoring needed to enter the prediction pipeline properly.
Restore SPARQL Query Files
Three .rq files disappeared from queries/ directory (noted Mar 30). Step 1 insight digest runs but produces empty sections without them.
Address Max-Turns Limit
3 of 13 runs hit the 20-turn safety limit. Consider increasing to 25 or splitting the pipeline into two smaller scheduled tasks (insights + experiments).
Expansion Plan
| Experiment | Status | Prerequisite | What It Tests |
|---|---|---|---|
| Exp 1: Goodhart Guard | Active | None (started) | Early stopping to prevent overtuning on small data |
| Exp 2: MOTPE Optimizer | Active | 4 weeks data (have 3) | Multi-objective optimization: accuracy + Brier + MAE |
| Exp 3: Dimension Weights | Queued | 6 weeks data | Learn optimal dimension weights via TPE |
| Exp 4: Temporal Decay | Blocked | 8 weeks data (late April) | Exponential decay weighting for recency bias |
| Exp 5: Cross-Vertical | Designed | Exp 2 converged | Warm-start K-Pop / AI Agent verticals from micro-drama config |
Scaling Milestones
Current: 1,672 triples, 21 companies, 3 weeks. The expansion spec targets 10K triples (8 weeks, cross-vertical), 100K (multi-market), 1M+ (full automation). The $70K-120K capital deployment model funds the infrastructure for these milestones.