This document describes what ARGUS 0.5 computes today. The longer product blueprint contains future research ideas and must not be read as implemented behavior. The machine-readable contract is GET /api/v1/methodology.
ARGUS retains public metadata and short feed summaries, not article bodies. The current adapters collect:
- GitHub repository metadata and weekly commit activity;
- dated Hacker News stories matching configured relevance terms;
- current Google News RSS titles, links, dates, and summaries.
GitHub collection works anonymously but uses GITHUB_TOKEN when configured to
avoid the low unauthenticated quota. The token does not change scoring.
Exact normalized-title duplicates are removed. ARGUS does not currently perform semantic claim clustering, full-text extraction, organization-level production verification, or LLM classification of evidence. OpenRouter is used only for separate technology-candidate discovery and administrator-requested draft-profile preparation.
All dimensions except momentum use a 0–100 normalized scale. Momentum ranges from −100 to 100.
| Dimension | Implemented meaning |
|---|---|
| Attention | Relative public discussion and project activity; visibility, not quality |
| Expectations | Forward-looking impact or replacement language in collected metadata |
| Disappointment | Setback, limitation, security, reliability, or failure language |
| Verified adoption | A deliberately conservative public-use signal; repository activity is capped as a weak proxy and is not production proof |
| Operational maturity | Maintenance, release, governance, integration, and operating-practice signals |
| Momentum | Completed-week activity relative to a trailing four-week baseline. It is withheld while the current week is settling, so incomplete activity cannot appear as a contraction. |
| Evidence coverage | Share of configured source families that completed collection, not percentage of the internet observed |
hype_gap = expectations - adoption. A positive result means claims are running ahead of observed use. It is not a prediction that the technology will fail.
The implemented metadata weights are 0.90 for GitHub repository evidence, 0.62 for Google News RSS evidence, and 0.48 for Hacker News evidence. These values express relative analytical influence for the associated claim type. They are not truth probabilities. Source failures reduce coverage and remain visible.
The deterministic scoring model in backend/src/argus/inference/engine.py evaluates five qualitative phases. It combines the dimensions with explicit weights and adds only a modest continuity preference for the same or an adjacent prior phase. Softmax converts scores to a phase distribution. No LLM selects a phase.
Confidence combines source coverage and separation between the two leading phase scores, then applies conflict and coverage caps. Source coverage is based on the actual completion outcome of GitHub, Hacker News, and Google News RSS—not the number of collected headlines or commits. The UI shows low, moderate, or high confidence; the numeric score is a model diagnostic, not measured accuracy.
The current calibrated model also caps confidence when evidence comes from fewer than two independent source classes or lacks an independent adoption-oriented signal. Current source-specific adjustments are bounded and applied at most once per source class and feature: release/package metadata can support maturity, security advisories can support setbacks, developer implementation discussions can support adoption, and research metadata can support maturity. Their values are versioned and covered by the locked synthetic regression fixtures; they are not a claim of statistical calibration against market outcomes.
Active technologies are analyzed weekly by default. After 12 consecutive weekly snapshots spanning at least 11 weeks in plateau_of_productivity, ARGUS changes their analysis cadence to quarterly. They remain publicly visible, become due every 12 weeks, and can be returned to weekly analysis by an operator. The threshold is configurable with ARGUS_PLATEAU_WEEKS, with a safety minimum of eight weeks.
- Public data overrepresents English-language and open-source activity.
- Current historical signals are reconstructed from GitHub and Hacker News metadata.
- Headline keyword classification can miss context and must not be treated as semantic fact extraction.
- News RSS availability and repository APIs can be incomplete or rate-limited.
- The latest weekly snapshot can omit momentum while its period is settling; ARGUS deliberately does not infer a direction from a partial week.
- ARGUS is independent, not affiliated with Gartner, and does not reproduce official Gartner research.