Methodology

Two pipelines, two philosophies

Substrate Wire runs two separate editorial pipelines:

• Shadow War pipeline — OSINT signals collected from telemetry sources (GPS jamming monitors, satellite thermal detection, internet outage trackers, aircraft transponder data). All synthesis, hypothesis generation, and narrative writing is done locally by Qwen3.6-35B-A3B on a home CPU. Zero cloud LLM calls.
• Frontier news pipeline — AI, geopolitics, markets, health, space. Articles collected from RSS and web sources, pre-filtered locally, then ranked and summarized with a combination of local AI and cloud-LLM assist (MiniMax M2.7 for editorial scoring and section summarization).

Shadow War pipeline — fully local

Signal sources are machine-telemetry systems that detect anomalies against their own baseline. These are not news articles — they are:

• GPSJam — GPS interference monitor (aircraft transponder data)
• NASA FIRMS — satellite thermal detection
• OONI — censorship and internet access monitor
• Cloudflare Radar — internet traffic anomaly detection
• IODA — internet outage and BGP route monitor
• USGS — seismic event detection
• ADS-B Exchange — military aircraft transponder tracker

The narrative brief and per-signal hypotheses are written by Qwen3.6-35B-A3B (35B parameters, 3B active per token, MoE architecture, Apache 2.0) running locally on a Ryzen 5900X with 64GB RAM. CPU-only inference. No data leaves the local machine for Shadow War editorial passes.

Frontier news pipeline — local + cloud assist

Articles go through a multi-pass pipeline:

1. Collection from RSS and web sources across 11 sections
2. Deduplication and paywall filtering (rule-based, local)
3. Pre-editorial filtering — local Qwen3.6-35B-A3B trims low-signal articles
4. Final editorial ranking, importance scoring, and summarization — MiniMax M2.7 cloud
5. Lead promotion and daytime refresh — local AI + cloud assist on promotion events

Source tiers

Every story is assigned a source tier before editorial scoring:

• Tier A — Wire services: AP, Reuters, AFP. Highest factual weight.
• Tier B — Major national papers, established broadcasters (NYT, BBC, WSJ, FT).
• Tier C — Trade and sector publications (TechCrunch, Breaking Defense, SpaceNews).
• Tier D — Blog posts, speculative analysis, opinion. Shown collapsed.

Tier D items are not promoted to lead or top stories regardless of engagement.

Refresh cadence

The nightly editorial pipeline runs at 02:15 local time. Daytime refresh-merge runs 5 times per day to promote breaking stories and expire stale ones. Shadow War signals are collected on a separate schedule.

Hallucination mitigations

• Every factual claim links to its primary source.
• Direct quotes are only used when pulled verbatim from a cited source.
• Shadow War signals are machine-telemetry anomalies — not news interpretation.
• Items below a minimum confidence threshold are dropped, not published.
• If you spot an error, contact [email protected].

Prediction track record

Shadow War analyst items are timestamped at write time. Each carries a confidence tier (A/B/C/D) and a falsifiable claim. Hit/miss/partial outcomes are logged in the append-only prediction record.

Resolved predictions are logged at /track-record. The raw append-only prediction log (JSONL) is committed to the market-intel git repository alongside the collector scripts — edit history is fully auditable.

What is excluded

• No content generated from scratch — only summarization of sourced articles.
• No unsourced claims about named private individuals.
• No affiliate links, no ads, no tracking pixels.

About  ·  Disclaimer  ·  Support