Wiredframe Radar

R[AI]DAR — evidence-linked daily intelligence on the AI ecosystem

Live Radar · Architecture · Quick start · Deployment · Roadmap

Wiredframe Radar turns a broad daily signal stream into a concise, traceable briefing for executives, strategists, researchers, and builders. It collects current material, filters and ranks it, detects themes spanning multiple categories, generates an evidence-backed executive summary, validates the report, and publishes a static site.

The pipeline is designed to degrade safely: provider fallbacks, checkpoints, schema validation, deterministic editorial checks, and a publish gate prevent a superficially successful run from replacing the last good report with incomplete output.

✨ What it delivers

• Four daily intelligence views: AI News, Research, Social signals from X, and GitHub Trending.
• Cross-category synthesis: topics are accepted only when supported by current items from at least two populated categories.
• Evidence-grounded writing: summaries return exact current-item IDs and cannot use historical summaries as fresh evidence.
• Consistent LLM quality: price-guarded paid MiniMax M3 handles both per-item analysis and high-value synthesis; Gemini and NVIDIA routes provide task-specific fallbacks.
• Operational visibility: phase status, collection funnels, LLM telemetry, token usage, provider costs, and GetXAPI calls are persisted with the report.
• Static-first publishing: generated JSON feeds a fast SvelteKit frontend and machine-readable discovery artifacts.

🧭 Pipeline at a glance

flowchart LR
    subgraph Sources[Current sources]
        RSS[43 news feeds]
        WEB[13 direct web sources]
        HN[Hacker News]
        PAPERS[HF Papers + AlphaXiv]
        RESEARCH[19 feeds + LessWrong + 4 dated hubs]
        X[X via GetXAPI]
        GH[GitHub Trending]
    end

    subgraph Pipeline[Daily intelligence pipeline]
        COLLECT[Collect and normalize]
        FILTER[Keyword filter and deduplicate]
        ANALYZE[Batch analyze and rank]
        SYNTH[Detect topics and synthesize]
        ENRICH[Deterministic-first link enrichment]
        GUARD[Quality and editorial gates]
    end

    subgraph Outputs[Published outputs]
        JSON[Versioned report JSON]
        SITE[SvelteKit static site]
        INDEX[llms.txt and ai-index.json]
        METRICS[Cost and LLM telemetry]
    end

    Sources --> COLLECT --> FILTER --> ANALYZE --> SYNTH --> ENRICH --> GUARD
    GUARD --> JSON --> SITE
    GUARD --> INDEX
    GUARD --> METRICS

Loading

🗂️ Active source inventory

The configuration files—not this table—are the source of truth. Counts reflect the current repository configuration.

Category	Active inputs	Collection path	Notes
AI News	43 RSS/Atom feeds, 13 direct web pages, Hacker News, links expanded from X posts	NewsGatherer, WebScraperGatherer, HackerNewsGatherer, LinkFollower	Includes Kimi, OECD.AI, NIST CAISI, The Batch, Databricks AI-filtered posts, MiniMax News, and Z.ai's official release stream. New direct sources use exact-date deterministic parsing and no LLM calls.
Research	Hugging Face Daily Papers, AlphaXiv Trending, 19 RSS/Atom feeds, LessWrong, 4 dated web hubs	ResearchGatherer	Includes Anthropic Research and Economic Futures, Arena, Epoch AI, Meta AI Research, and OpenAI Research-tagged entries. HTML hubs use deterministic exact-date parsing and no LLM calls.
Social	170 configured X accounts	SocialGatherer	@NVIDIAAI is retained while the broader corporate @nvidia account is excluded. GetXAPI queries at most 20 accounts per paid request.
GitHub Trending	GitHub Trending	GitHubTrendingGatherer	Repositories are analyzed as a separate report category.

Bluesky, Mastodon, Reddit, YouTube, Product Hunt, Discord, and Slack are not active pipeline sources. See AI news sources for maintenance rules and the exact configuration entry points.

🧠 LLM strategy

Routing is caller-aware, so high-volume classification and high-impact synthesis do not compete for the same quota.

flowchart TD
    TASK{Task class}

    TASK -->|Bulk map and filter| NEMO[NVIDIA Nemotron 3 Nano]
    NEMO -->|Fallback| GFL[Gemini 3.5 Flash Lite]
    GFL -->|Fallback| G31[Gemini 3.1 Flash Lite]

    TASK -->|Ranking, category summary, topics, executive| ORGLM[OpenRouter GLM 5.2 paid]
    ORGLM -->|Fallback| G36[Gemini 3.6 Flash]
    G36 -->|Fallback| NVGLM[NVIDIA GLM 5.2]
    NVGLM -->|Fallback| G35[Gemini 3.5 Flash]
    G35 -->|Fallback| G35L[Gemini 3.5 Flash Lite]

    TASK -->|Link enrichment fallback| LINKGLM[NVIDIA GLM 5.2]
    LINKGLM --> G35

Loading

The executive-summary context has a strict boundary:

=== PREVIOUS DAYS' COVERAGE (HISTORICAL; DO NOT REPORT AS CURRENT) ===
...
=== END PREVIOUS DAYS' COVERAGE ===

=== TODAY'S DATA (CURRENT EVIDENCE) ===
... exact current item IDs, titles and summaries ...
=== END TODAY'S DATA ===

Historical reports help the model avoid repetition; only records inside today’s section may support new claims.

🛡️ Reliability model

stateDiagram-v2
    [*] --> Collecting
    Collecting --> Analyzing: minimum source coverage met
    Collecting --> Blocked: collected category collapses
    Analyzing --> Synthesizing: schemas and coverage valid
    Analyzing --> Blocked: fallback rate or evidence invalid
    Synthesizing --> Validating: topics and summary grounded
    Synthesizing --> Blocked: critical synthesis fails
    Validating --> Published: quality threshold met
    Validating --> Blocked: report gate fails
    Blocked --> LastGoodReport: generated files reverted
    Published --> [*]
    LastGoodReport --> [*]

Loading

Key safeguards include:

• checkpoint-based resume after recoverable failures;
• route cooldowns, retries, and fallback chains;
• schema validation and fallback-rate limits;
• fail-open News filtering when an LLM emits invalid JSON;
• evidence coverage checks for topics and executive output;
• deterministic sanitization of unwanted branding and unsupported claims;
• quality scoring and a final report validator before Git commit;
• automatic restoration of the last good report when validation fails.

🚀 Quick start

Requirements: Python 3.11+, Node.js 20+, and API keys for the routes you enable.

git clone https://github.com/BlockFrame/wiredframe-radar.git
cd wiredframe-radar

python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python -m playwright install chromium

cp .env.example .env
npm run install:frontend

At minimum, configure GEMINI_API_KEY, OPENROUTER_API_KEY, and NVIDIA_API_KEY. Add GETXAPI_KEY to collect X posts.

Run the pipeline and frontend:

python run_pipeline.py --date 2026-08-10
npm run dev

Useful recovery commands:

python run_pipeline.py --resume
python run_pipeline.py --resume-from 4.5
python scripts/validate_report.py --web-dir ./web --date 2026-08-10

See Quick start for setup details and Deployment guide for GitHub Actions and Vercel.

⚙️ Configuration

File	Purpose
config/providers.yaml	LLM routes, quotas, fallbacks, timeouts, and pipeline defaults
config/prompts.yaml	Analysis, synthesis, and enrichment prompt contracts
config/rss_feeds.txt	AI News RSS/Atom feeds
config/research_feeds.txt	Research feeds and LessWrong routing entry
config/research_web_sources.txt	Deterministic, date-verifiable Research pages without feeds
config/research_reference_sources.txt	Authoritative static hubs excluded from daily collection
config/twitter_accounts.txt	X accounts queried through GetXAPI
config/web_scraper_sources.txt	Direct pages for sources without usable feeds
config/ecosystem_context.yaml	Grounding context used by synthesis
config/model_releases.yaml	Release-date grounding data

Secrets are loaded from environment variables. Never commit provider keys or proxy credentials.

📦 Outputs and interfaces

Each successful run writes web/data/<YYYY-MM-DD>/summary.json, including category reports, evidence IDs, collection status, analysis funnels, phase status, quality information, and LLM telemetry. Cost details are written beside the report.

The repository also maintains:

• llms.txt for language-model-readable discovery;
• ai-index.json for structured indexing;
• mcp_server.py for MCP-compatible access;
• the static SvelteKit site under frontend/.

🧪 Validation

Unit tests use mocks and do not consume paid LLM or GetXAPI quota.

python -m unittest discover -s tests -p '*_test.py'
npm run check
npm run build

The GitHub Actions workflow runs the critical regression suite before any paid API call, then validates the generated report before publishing it.

🏗️ Repository map

agents/                 Gatherers, analyzers, routing, orchestration and guards
config/                 Providers, prompts, feeds and grounding data
frontend/               SvelteKit application
generators/             JSON and optional visual-output generation
scripts/                Validation, deployment and operational utilities
tests/                  Mocked regression and integrity tests
web/data/<date>/         Versioned reports and telemetry
.github/workflows/       Scheduled generation and publishing

🤝 Contributing

Keep changes evidence-preserving and quota-aware. When adding a source, update its configuration file and source inventory. When changing synthesis, add a mocked regression test that proves current-item grounding and failure behavior.

📄 License

Licensed under the Apache License 2.0. Wiredframe Radar evolved from the open-source ai-news-aggregator project and now maintains its own collection, reliability, routing, editorial, and publishing architecture.