What we shipped.

15,628 Layer Z signals across nine AI deployment archetypes. Open methodology, CC BY 4.0 dataset, MIT-licensed generator. Published at /research.

Every User Health Report now anchors numbers against a cross-archetype median (e.g. "frustration of 0.49 sits above the 0.40 baseline for cx_chatbot, N=1,803").

No-login dashboard at /demo/explore: a populated CX-bot deployment with 600 synthetic conversations, every chart drawn from real Layer Z signals.

The BIE agent surfaces baseline medians and source disclosure inline. Tool v1.1; 12-tool agent.

Nightly review of deployment memory files. Flags claims that contradict signals, drift from prior state, or land outside the file’s scope.

Two-arm search on every deployment. Embeddings for paraphrase recall, tsvector for exact phrasing, weighted blend by default.

Independent tool calls fire in the same iteration. Median synthesis time down 38% on multi-tool questions.

concise vs detailed switch on bie_deployment, bie_layer_z_trend, bie_query_signals, bie_search_conversations_lex, bie_search_conversations_sem. Cuts token spend on rail queries.

Twelve auto-clustered topics by five conversation-turn buckets. Where users hit friction, by topic.

Public per-org behavioral health score. Slug claim, opt-in, embedded SVG badge.

AI turns are read for context only. Human turns get the score.

7-day calibration before the first alert. No alert storms on day one.

Weekly per-deployment narrative. Headline finding. Falsifiable claim.

Read-only MCP scope. Interrogate BIE from Claude Desktop, Cursor, anywhere.

Welch t-test per dimension on before/after deploy events.

95% token reduction on 10K-row datasets vs the per-row loop.

The daily digest no longer burns tokens on quiet days.