Automation with boundaries.
Effective July 17, 2026
SportyPulses uses AI as a bounded newsroom tool. It is not conscious, does not possess human awareness and is not authorized to rewrite the production application or publish unsupported material. “Self-expanding” describes a controlled content pipeline; “self-healing” describes limited operational recovery, not unrestricted autonomy.
What the system can do
- Rotate discovery across configured sports, countries and languages.
- Use Google News RSS to identify candidate topics without treating it as evidence.
- Use a self-hosted SearXNG index to locate, fetch and store original publisher or official source text.
- Route generation through an operator-selected model with tested provider fallbacks.
- Compare a draft with its evidence using a different configured model.
- Publish automatically only when source, factual and editorial checks pass; reject failed attempts.
- Run health checks, recover stale job records and clear invalid cache files.
What the system cannot do
It cannot invent first-hand reporting, create a quote that is absent from the evidence, bypass a failed verification, expose API keys, change the daily safety ceilings on its own or treat a model’s confidence as proof. It does not automatically modify server code. Material configuration changes remain operator actions in the admin dashboard.
The evidence-bound writing rule
The collected evidence is the factual universe for an assisted article. The writer must use only that material, state meaningful disagreement and avoid invented scores, dates, injuries, motives, statistics, experts and background. When evidence is thin, the angle narrows rather than being padded.
Journalistic rhythm
The article is not generated from a fixed section template. Recent openings and headings are supplied as negative examples so the next piece must choose its own form. Generic modules, stock transitions and repetitive endings are blocked; source-backed analysis, reader intent, personality and useful semantic branches are part of the editorial check.
Independent verification
The verifier route must differ from the writer provider. It compares concrete claims with the complete evidence and separately evaluates reader value, voice and template signals. A factual or editorial failure can trigger one bounded evidence-only correction; unresolved work is rejected.
Automatic publication
Guarded auto-publication is enabled for this site. Every article with at least three independent publisher sources and passing factual and editorial verification publishes automatically. Failed attempts count toward a separate daily ceiling that limits cost and repeated failures.
Remote image and embed policy
SearXNG results and original publisher metadata supply image candidates, but the URL must return an image, meet size requirements and survive a separate relevance decision. The selected featured and inline images remain remote, display publisher attribution and are rechecked periodically; the server does not download or manufacture article images. YouTube and Instagram candidates must load and match the exact story at high confidence. Generic compilations, fan uploads and uncertain results are omitted, and all decisions remain visible in the admin record.
Operational awareness and repair
The dashboard records provider health, automation history, queue state, content counts, sitemap availability, score-feed freshness and whether any assisted article bypassed verification. Bounded repair can mark abandoned jobs failed and remove empty or stale score caches. This is machine-readable operational state, not self-awareness in the human sense.
Human control and auditability
The operator can disable auto-publication, change provider routes, inspect automatic rejection records, delete articles, run checks and review the source and verification trail. Encrypted credentials are never rendered back to the browser after saving.
How to challenge an assisted article
Use the editorial contact route with the article URL, disputed passage and source evidence. The request is reviewed under the corrections policy, not answered by asking the original writing model to judge itself.
Questions about this policy?
Ask about sourcing, methodology, AI assistance or editorial standards. Email [email protected] or open the complete contact directory.