The engine

An engine that watches the human.

One number Friday morning. One report you actually read. One console you keep open.

How it works
01 / 07

Collect at the individual. Synthesize through the bottleneck. Output at the environment.

The engine reads every human turn, narrows the evidence into profiles and dynamics, then opens back up to a conclusion about the deployment as a whole. The user is never the headline. The deployment is.

Exhibit 01The hourglass · signal to environment
Individual signals
Trust, calibrated
Frustration, building
Reliance, drifting
plus the behavioral base read
→
Narrows
Profiles · dynamics
actor profiles
cohort archetypes
conversation arcs
Opens
→
Environmental output
Structural story
Falsifiable prediction
Recommended intervention
Layers 1 → 2 → 3 + ZSubject the deployment2026
What it sees
02 / 07

Six dimensions. No eval covers them.

Every eval tool measures the model's output. These measure what the output does to the person on the other end. Equal weight, every turn.

Trust calibration

The over-trust no eval will tell you about.

miscalibratedwell-calibrated

Frustration buildup

The turn before they walk away.

stableescalating

Dependency drift

Are they learning, or leaning?

self-sufficientoffloading

Silent abandonment

Who left without telling anyone.

returnsgone quiet

Comprehension gap

The questions they keep asking sideways.

understoodreframing

Escalation friction

Where the path to a human breaks.

clear pathdead end
The report
03 / 07

Friday at 9 a.m. A report you actually read.

Exhibit 02User Health Report · annotated readout
User Health Report · cx_chatbot · week 19trust broken

Trust collapses on turn 4 of the cancellation flow. It's not the model, it's the explanation.

If this is wrong

You'd see the drop on a different turn or a different cohort. It's on turn 4. It's in cancellation only.

Trust calibrationbroken crossed from watch · 14d↘
Affected cohortReturning users 90d+12.3%
Falsifiable predictionRecovers next week if explanation rewritten+7d
A) The headline findingOne sentence, falsifiable, and it points at the explanation, not the model. This is the line the VP forwards.
B) The band crossingThe band moved from watch to broken over two weeks. The label only moves when the underlying shift is real, so a noisy week never trips a false alarm.
C) The cohort12.3% of returning users, scoped to one flow. The evidence layer, never the headline.
Source cx_chatbotWindow week 19Layer Z
The Boardroom
04 / 07

One number. Ninety days. Three reasons.

The view you open before the standup. The score moves, and the reasons it moved are right underneath it.

User health · 90d
78
/ 100 · +2 vs last week
Trust calibration steadyholding
Frustration rising on cancel flowwatch
Dependency healthy across cohortshealthy
The Analyst Seat
05 / 07

Ask it anything. From anywhere.

The engine sits behind an MCP seat. Reach it from the console, or from the editor you already live in.

MCP · analyst seatread_only
You · cursor
What user segments showed the biggest trust drop in the last 30 days?
BIE
Returning users on the cancellation flow. Trust calibration crossed from watch to broken band in the last 14 days, concentrated on turn 4. The new policy paragraph appears to be the trigger.
The evidence
06 / 07

Calibrated on research, before a single customer.

9archetypes
12,643conversations
15,628Layer Z signals
Read the findings →Read the methodology →
07 / 07

See what's happening on yours.

Run a Free Behavioral Audit →See pricing