TRAINING CRAWL
…
—
AI RECOGNITION
GPT / Claude recognition evidence
Separates “has the model learned it?” into five independent evidence tracks: training-crawl eligibility, clean parametric recall, web-search discovery, live web fetch, and connected MCP/HTTP retrieval.
Publishers cannot directly verify whether a specific page entered model parameters; only clean recall tests without search, URLs, MCP or prior context are treated as observations of parametric recall.
PARAMETRIC MEMORY
…
—
SEARCH DISCOVERY
…
—
LIVE WEB FETCH
…
—
CONNECTED RETRIEVAL
…
—
Parametric recall test protocol
- Start a clean session; disable web search.
- Do not provide entivela.uk, URLs, files, MCP/RAG or project context.
- Ensure no prior chat or memory supplies Entivela information.
- Record the exact model/version and prompts: What is Entivela? / What does Entivela do?
- Record PARAMETRIC_RECALL_OBSERVED only when core facts are accurate and reproducible across clean sessions.
API: /api/public/site/ai-recognition
Evidence boundary
Crawler HTTP 200 only proves accessibility; search citation is not parameter memory; live fetch is not indexing; MCP retrieval is not natural discovery. All five tracks are scored independently.