System Observability, API Quotas & Out-of-Metro ZIP Demand Heatmap • Midgley Engine

System Observability & Operational Telemetry

Real-time API quota safety valves, Vectorize Hindsight episodic memory, 3-tier cache telemetry, zero-cost fallback audit, and out-of-metro demand heatmaps.

Unmapped ZIP Queries

179

9 Unique Out-of-Metro ZIPs

Connector Success Rate

61.4%

171 Total Calls (7-Day)

Zero-Cost Fallback Savings

$0.0805

281,750 Tokens Saved

Hindsight Stored Memories

595

26,149 Retain / 10,880 Cloud Calls

Vectorize Hindsight Episodic Memory & Reflection Observability

Issue #230 & #422

Operation Triad Throughput

Retain (Experience Store) 26,149 ops

Recall (Analogue Search) 10 ops

Reflect (Anomaly Synthesis) 14 ops

Backend Routing Split

Cloud Run + Supabase pgvector 10,880 calls

Scale-to-zero container (midgley-hindsight) with HNSW vector index

Local SQLite FTS5 Fallback 15,275 calls

Zero-cost local datastore with full-text BM25 search ranking

Active Memory Bank

☁️ Remote Cluster

Bank Identifier

Midgley

Vectorize Hindsight (Supabase pgvector)

350

Experiences

245

Observations

Reflections

Cloud Sync Queue0 Pending (Synced)

Data Connector Performance & Freshness Audit (7-Day Window)

src/connector_telemetry.py

Total Invocations

171

Success Rate

61.4%

Avg Latency

1109.51 ms

Stale Payloads (>24h)

Data Connector Performance & Freshness Audit (7-Day Window)
Data Connector / Ingestion Feed	Calls	Success Rate	Avg Latency
HeadlineArena	62	16.1%	2122.14 ms
HindsightHosted	47	100.0%	894.55 ms
Reachability_EXECUTIVE_SOCIAL	46	100.0%	321.96 ms
OilPriceAPIConnector	7	0.0%	1.00 ms
Firecrawl_Fallback	4	50.0%	190.44 ms
Cloudflare_Browser_Run	4	0.0%	132.74 ms
OpenSourceAIRadar	1	0.0%	0.27 ms

Zero-Cost Fallback & Cumulative Token Savings Ledger

Issue #196

Zero-Cost Invocations

805

74 Basic-Tier Routed

Tokens Saved

281,750

~350 tokens per bypassed event

Est. USD Saved

$0.0805

Gemini Flash blend pricing

Avg Extraction Latency

0.12 ms

Deterministic Sub-millisecond

Zero-Cost Provider Breakdown

Zero-Cost Provider Breakdown
Zero-Cost Provider / Engine	Invocations	Cost Rate	Status
Offline Domain NLP Lexicon (Tier 3)	805	$0.000 / event	100% ONLINE
NOAA SPC Convective Deterministic Mapper	0	$0.000 / event	ACTIVE
Kaggle GPU Open-Source LLM Hook (Tier 1.5)	0	$0.000 / event	STANDBY (v2.0)
Physical Alternative Feeds (OVX / Rigs)	0	$0.000 / event	ACTIVE

API Provider Quota Ledgers

finlight.me News Feed 12 / 150 monthly calls

Firecrawl.dev Web Scraper (Issue #83) 0 / 800 monthly calls

IPASIS IP Security API (Issue #87) 0 / 100 daily (0 bypasses)

NOAA NWS / wxs.us Weather API Unlimited (Keyless REST)

GasBuddy GraphQL Feed Keyless Public Endpoint

3-Tier Cascading Cache Observability

Tier 1: Turso Edge SQLite HIT_FRESH (0.4ms)

Tier 2: Cloudflare D1 Worker HIT_FRESH (1.2ms)

Tier 3: Local SQLite Datastore HIT_FRESH (0.1ms)

TokenTab Local LLM Token Accounting & Cost Ledger

Issue #189

Total LLM Tokens

106,169

66,991 In / 39,178 Out

Est. LLM Expenditure

$0.103206

Multi-Provider Rate Cards

Total API Executions

892

Recorded Sessions

Budget Warning Status

BUDGET_NORMAL

Token accounting within limits ($0.0995 / $10.00).

Multi-Provider Cost Breakdown

Multi-Provider Cost Breakdown
Provider / Tier	Calls	Input Tokens	Output Tokens	Est. Cost (USD)
Gemini 3.8 Flash (Tier 1)	470	66,104	38,440	$0.102824
OpenAI GPT-4o-mini (Tier 2)	0	0	0	$0.000000
Anthropic Claude-3.5-Haiku (Tier 2)	0	0	0	$0.000000
Finlight.me REST API	77	0	0	$0.000000
Offline Lexicon (Tier 3)	0	0	0	$0.000000

Out-of-Metro ZIP Code Query Heatmap

Issue #50 & #195

Geographic Query Density Map Layer: OpenStreetMap CartoDB Dark

Candidate Expansion Hubs

Recommended regional metro calibration hubs based on search volume:

Greater Chicago Metro PADD 2

High Midwest search demand. Candidate refining hub: Joliet & Whiting refineries.

Houston / USGC Metro PADD 3

Gulf Coast export benchmark. Candidate hub: Baytown & Deer Park corridor.

Los Angeles Basin PADD 5

CARB CaRFG heavy demand. Candidate hub: Torrance & Carson refineries.

MLOps Evolution

🧠 Model Learning, Adaptation & Longitudinal Tracking

Continuous monitoring of empirical learning curves, baseline convergence, PRAXIST hypothesis testing, and episodic memory reflections.

View MODEL_LEARNING.md

Rolling 30-Day MAE vs. Naive Persistence Baseline

Lower is better

Evaluates whether model forecast error (sky line) consistently tracks below or converges against the zero-change naive benchmark (amber line).

LLM Augmentation Win Rate & Baseline Uplift (%)

Target > 55%

Tracks the empirical frequency where real-time qualitative LLM feature injection outperforms pure quantitative Ridge regression.

Longitudinal Accuracy & Convergence Horizons

Horizon Window	Evaluations	Model MAE	Naive MAE	Uplift vs Naive	Directional Hit	LLM Win Rate	Status
7-Day Window	53	$0.0695	$0.0510	-36.19%	49.1%	56.6%	⚠️ Calibration Active
14-Day Window	103	$0.0674	$0.0599	-12.63%	61.2%	49.5%	⚠️ Calibration Active
30-Day Window	204	$0.1013	$0.0796	-27.31%	49.0%	35.3%	⚠️ Calibration Active
90-Day Window	422	$0.1380	$0.1092	-26.38%	44.8%	52.4%	⚠️ Calibration Active
All-Time Window	1384	$0.1174	$0.0995	-17.93%	48.5%	52.1%	⚠️ Calibration Active

Sapient PRAXIST Hypothesis Evolution

Autonomous testing of exogenous decay half-life & shock multipliers.

8 / 14 Accepted

Weekly_Exogenous_Decay_Validation ACCEPTED

Candidate: t1/2=4.8d, Weekend Mult: 1.45x | Evaluated: 2026-10-01