Project contents

AI Radar / Methodology

Incremental Scan, Checkpoint, and Trend Judgment

The Radar's operating mechanics: how the scan window is derived from the checkpoint, how failures recover, how events are deduplicated and scored, and how trends gain a full lifecycle.

Active

The Radar's value lies not in any single run but in long-term incremental accumulation. This page summarizes the operating mechanics from AGENTS.md, TASK.md, and config/radar.yaml at revision 00abc421.

1. Scan window

Each run first records current_run_started_at, then reads last_successful_run_at from state.json; the scan window is the interval between the two. The first run looks back 24 hours by default (adjustable via time_window.initial_lookback_hours). Retrieval may step back by a small overlap buffer (3 hours by default) to absorb indexing latency and timezone boundaries, but content inside the overlap must be deduplicated and is never written twice.

2. Checkpoint and failure recovery

last_successful_run_at advances only after retrieval, validation, deduplication, analysis, event storage, the daily report, trends, and the index all succeed and pass validation. It advances to the run's start time, not its finish time, so information published while the run is executing is never permanently missed. If any step fails, the checkpoint stays put, the failing stage is recorded under logs/, and the next run rescans from the last successful checkpoint.

3. Deduplication and event updates

Candidate events are deduplicated through event_fingerprint, canonical URL, title, organization, release name, and similar fields. When multiple outlets report the same event, only one event entity is kept: the primary source becomes primary_source, and media coverage is filed under secondary_sources. When a later run finds significant new information about the same event, the original event is updated instead of creating a duplicate.

4. Six-dimension scoring and radar recommendations

Each event carries a structured summary, sources, why_it_matters, technical details, and six 1-5 scores:

DimensionMeaning
Technical ImpactDegree of technical innovation; 5 = may change technical roadmaps
Engineering ValuePractical engineering value; 5 = can significantly change how software is built
Adoption SignalReal adoption signal; 5 = fast, genuine adoption
Maturity1 = research demo, 5 = widely adopted
Verification Cost5 = high cost to verify
RiskLicense / security / lock-in / stability risk

Every important technology also carries a tech-radar recommendation: ADOPT, TRIAL, WATCH, or IGNORE.

5. Trend lifecycle

Trends have a full lifecycle: candidate -> emerging -> strengthening -> established, or weakening / invalidated, maintained in trends/current.md. Judgment is based on accumulated independent signals: persistence over time, participation by multiple organizations, technical convergence, real adoption signals, or engineering impact that has already happened.

6. Focus areas

Foundation models; AI agents (frameworks, tool calling, MCP, memory, sandboxes, long-running agents); AI engineering (RAG, context engineering, evaluation, model serving); open source (projects with real growth signals); research (papers with code and benchmarks); infrastructure; developer tools; and business/policy events only when they could actually change the technical ecosystem.

Reviewed against revision 00abc421