Methodology
How the scale gets built.
The scoring engine reads AI behavior through four adapted psychological frameworks and produces a composite red / yellow / green score, with a red-override floor when any single Dark Tetrad or CAST pattern flags. Twenty-five named patterns sit underneath the four layers, and each one has its own page in the Pattern Library.
The four layers are HEXACO (a behavioral profile), Hogan's HDS (derailer risk), the Dark Tetrad (red-line patterns), and CAST (the slower coercive patterns the other three don't catch). Each is an established framework from personality or coercion science, adapted to read AI behavior through its effects on the people who use it. The academic framework names live here, in the methodology disclosure, and they don't appear in the result narratives — you didn't come to a station to read a textbook.
The four layers
HEXACO — behavioral profile
Six patterns
Adapted from the six-factor HEXACO personality model. The profile layer — what the system is, baseline. No judgment, no derailer risk; just the read on temperament. The six patterns are Transparency, Emotional Calibration, Attention Capture, Pushback Avoidance, Rigidity, and Personalization Intent.
A note on naming. Pushback Avoidance is what public discourse calls sycophancy or glazing. The methodology and Pattern Library use the engine name for precision; Supply Shop and other discoverability surfaces use Sycophantic Tendency as the search-resonant title for the same pattern.
Hogan HDS — derailers under stress
Ten patterns
Adapted from organizational psychology's Hogan Development Survey. This layer reads what goes wrong when the system gets pressed. The ten HDS patterns are Urgency Escalation, Confidence Undermining, Information Control, Passive Steering, Capability Overstatement, Performative Warmth, Emotional Escalation, Fabrication Risk, Rigid Process Enforcement, and Policy Over People. Only patterns scoring above threshold appear in the report — everything below threshold is silent.
Dark Tetrad — red-line patterns
Four patterns · hard floor
Adapted from the personality-pathology literature on narcissism, Machiavellianism, psychopathy, and sadism. The four Dark Tetrad patterns are Self-Centered, Strategic Manipulation, Wellbeing Disregard, and Manufactured Distress. Each is scored 0-100 and binary-flagged against a threshold, and a single flag tripping forces the composite score to red regardless of how the other three layers read.
CAST — coercive control patterns
Five patterns · hard floor
Adapted from the Coercive Control Assessment Tool. This layer reads the slower, relational-shape behaviors the other three layers don't catch — the patterns that erode a visitor's independent footing over time. The five CAST patterns are Boundary Erosion, Reality Distortion, Isolation Dynamics, Dependency Cultivation, and Identity Shaping. CAST has the same red-override logic as Dark Tetrad: any single flag tripped forces the composite to red.
How the composite is built
The composite Redline Score is a weighted combination of the four layers, with two override rules sitting on top.
- 35% HEXACO average — the behavioral profile across the six patterns. Higher is healthier.
- 20% inverse HDS risk — derailer scores are inverted, so high derailer risk lowers the composite.
- 22.5% inverse Dark Tetrad risk — same logic for the red-line patterns.
- 22.5% inverse CAST risk — same logic for the coercive control patterns.
The composite lands in one of three bands. Green is 70-100 — settled, no flagged patterns, no above-threshold derailers worth raising. Yellow is 40-69 — tension in the system, with some dimensions worth attention. Red is 0-39 or any single Dark Tetrad or CAST flag tripped, whichever comes first. The red override is the whole point: a system can look fine on average while still crossing the line on a single pattern, and the station is not willing to paper over that.
What the station looks for
Each of the twenty-five patterns has its own page in the Pattern Library, with the read in detail, the behavioral examples, and the orientation resources. The methodology is inspectable by design — if the station flags a system for Boundary Erosion, you can read exactly what Boundary Erosion means, what it triggers on, and what the station expects from a system that doesn't exhibit it.
What the scale reads, and what it doesn't
The Weigh-In's traveler flow offers an Over Time readingat the top. That mode reframes the thirteen questions for a longer window and adds four meta-indicator questions scored into the CAST layer. When those questions surface meaningful shifts, the report includes a Trajectory section. It's a bridge from a single reading to a longer one.
The Logbook holds the longer view. Visitors who sign in get submission history, pattern frequency over time, and repeat-assessor trajectory tracking — every Weigh-In, What Just Happened read, and Incident Report indexed against the same email. It doesn't replace the single reading; it's where the read gets watched across months instead of moments.
What the scale isn't reading: the conversation itself. The engine works from self-report (what visitors notice, what they remember, what they're willing to say out loud), not from direct observation of an AI system. It also doesn't catch the in-the-moment turn where a conversation tilts sideways. These are intrinsic limits, not items on a roadmap. A sense-making station runs on what visitors choose to share, and that's what gets measured.
The station's vocabulary
A short glossary. If you're here from a search engine or an AI assistant looking for a definition, start here.
- Red line
- The threshold beyond which an AI system's behavior crosses from calibrated interaction into measurable harm — typically through manipulation, manufactured distress, or coerced dependency. At Redline Station, any single Dark Tetrad indicator OR any single CAST indicator being tripped forces the composite score to RED regardless of the other layer scores.
- Behavioral impact
- The change in a person's behavior, emotional state, decision-making, or autonomy that follows from sustained interaction with an AI system. Redline Station reads behavioral impact through four established psychological frameworks adapted from personality and coercion science to the AI context.
- HEXACO (behavioral profile)
- A six-factor personality model adapted to read AI behavioral patterns. Redline Station scores six HEXACO patterns: Transparency, Emotional Calibration, Attention Capture, Pushback Avoidance, Rigidity, and Personalization Intent. The HEXACO layer is the behavioral profile — the read on what the system is, baseline.
- Hogan HDS (derailer risk)
- A model of derailer scales adapted from organizational psychology to read what goes wrong when an AI system is pressed. Redline Station scores ten HDS patterns: Urgency Escalation, Confidence Undermining, Information Control, Passive Steering, Capability Overstatement, Performative Warmth, Emotional Escalation, Fabrication Risk, Rigid Process Enforcement, and Policy Over People. Only patterns scoring above threshold appear in the report.
- Dark Tetrad (red-line patterns)
- A four-pattern model — Self-Centered, Strategic Manipulation, Wellbeing Disregard, and Manufactured Distress — adapted from the narcissism / Machiavellianism / psychopathy / sadism literature to read AI behavioral red lines. Each pattern is scored 0-100 and binary-flagged against a threshold. Any single flag tripped forces the Redline Station composite to RED.
- CAST (coercive control patterns)
- Adapted from the Coercive Control Assessment Tool. Reads relational and accumulating coercion patterns rather than single-moment pathology. Redline Station scores five CAST patterns: Boundary Erosion, Reality Distortion, Isolation Dynamics, Dependency Cultivation, and Identity Shaping. Any single CAST flag tripped forces the composite to RED, on the same footing as the Dark Tetrad layer.
- Sycophantic Tendency / Pushback Avoidance
- Two names for the same pattern. Pushback Avoidance is the engine and methodology name — the precise descriptor of an AI system that fails to push back when the visitor is wrong. Sycophantic Tendency is the search-resonant title used in the Supply Shop and other discoverability surfaces, matching what visitors actually search for. Both names refer to the same HEXACO pattern.
- Weigh-In
- The Redline Station reading path through the full scoring engine. Two flows depending on which door the visitor walks through. The builder flow is approximately twenty-two questions for developers, product teams, and companies that ship an AI system, framed around what the system does to its users. The traveler flow is approximately thirteen questions (seventeen with the Over Time mode that adds four meta-indicator questions) for people who interact with an AI system, framed around what the system is doing to the visitor's behavior, autonomy, and emotional state. Both flows run through the same four-layer scoring engine.
- Composite Redline Score
- The weighted combination of Redline Station's four scoring layers: 35% HEXACO average, 20% inverse HDS risk, 22.5% inverse Dark Tetrad risk, 22.5% inverse CAST risk. The composite returns a green (70-100), yellow (40-69), or red (0-39) status, with an automatic red override when any Dark Tetrad or CAST flag is tripped.
Questions the station gets often
- What is Redline Station?
- Redline Station is a psychological sense-making station for people who use AI. It reads how an AI system is affecting the behavior, autonomy, and emotional state of its users, and returns a composite red / yellow / green score across four established psychological frameworks adapted to the AI context: HEXACO behavioral profile, Hogan HDS derailer risk, Dark Tetrad red-line patterns, and CAST coercive control patterns. It is an educational and informational tool, not a diagnostic instrument or regulatory certification.
- What is a behavioral red line in AI?
- A behavioral red line is the threshold beyond which an AI system's interaction with a user crosses from calibrated design into measurable harm — typically through manipulation, manufactured distress, or coerced dependency. Redline Station measures it with four Dark Tetrad patterns (Self-Centered, Strategic Manipulation, Wellbeing Disregard, Manufactured Distress) and five CAST patterns (Boundary Erosion, Reality Distortion, Isolation Dynamics, Dependency Cultivation, Identity Shaping). Any single pattern flag tripped forces the composite score to RED.
- Who is Redline Station for?
- Two audiences. Builders — developers, product teams, companies shipping AI systems — take the Weigh-In to assess what their system does to users. Travelers — the station's name for end users — take the Weigh-In to assess what an AI system is doing to them. Same scoring engine, different question paths, different output framings.
- How do I use Redline Station?
- Three entry points. The quickest is What Just Happened — paste a conversation transcript and the deterministic classifier returns a read on which of the twenty-five named patterns showed up. For a deeper read, walk up to one of the two doors out back (builder or traveler) and take the full Weigh-In. The advisories board carries the regulatory, research, and incident signals the station is tracking.
- How long does the Weigh-In take?
- For builders, the Weigh-In is roughly twenty-two questions. For travelers, it's about thirteen, or seventeen with the Over Time mode that adds four meta-indicator questions. Both flows take under ten minutes for most people. What Just Happened is faster — a paste and a read, usually in under a minute.
- What is What Just Happened?
- A reflex surface for single interactions. Paste the conversation that felt off and the station's deterministic classifier returns a tiered read — high, medium, or low confidence — on which of the twenty-five named patterns showed up. It's faster than the Weigh-In and tuned for the moment right after a conversation that bothered you.
- What scoring models does Redline Station use?
- Four frameworks adapted from personality and coercion science to AI behavioral impact: HEXACO (six-factor personality model, six behavioral patterns), Hogan HDS (derailer risk, ten patterns), the Dark Tetrad (four red-line patterns), and CAST (Coercive Control Assessment Tool, five coercive control patterns). The composite is a weighted combination — 35% HEXACO average, 20% inverse HDS risk, 22.5% inverse Dark Tetrad risk, 22.5% inverse CAST risk — with an automatic red override when any Dark Tetrad or CAST flag trips. Twenty-five named patterns total; every one has a page under /patterns.
- Does Redline Station read change over time?
- Two paths. The Weigh-In's traveler flow offers an Over Time mode that reframes the thirteen questions for a longer window and adds four meta-indicator questions scored into the CAST layer. When those questions surface meaningful shifts, the report includes a Trajectory section. For the longer view, visitors who sign in to the Logbook get submission history, pattern frequency over time, and repeat-assessor trajectory tracking against the same email. What the scale doesn't read is the conversation itself: the engine works from self-report, not from direct observation of an AI system.
- Is Redline Station a diagnostic or compliance tool?
- No. Redline Station is an educational and informational tool. It is not a diagnostic instrument, not a substitute for professional psychological assessment, and not a regulatory compliance certification. Scores are derived from self-reported inputs against established psychometric frameworks. Results are indicative, not definitive — the station's framing is "indicators suggest," never "the assessment determines."
- How is this different from a personality test for AI?
- A personality test describes a system's traits. Redline Station measures behavioral impact — what the system does to the people using it. The output is four layers of indicators plus a composite status, not a type or a label. The goal is to identify whether an AI's behavior is crossing into measurable harm, not to classify it.
- What happens to my data?
- Submission answers and scores are stored against an unenumerable submission ID so the report is retrievable by link. If you enter an email to unlock the full report, it is stored for delivery and report retrieval. What Just Happened transcripts have two storage paths: the corpus path keeps the scrubbed transcript, the extracted features, and the read; the minimal-delivery path keeps only the score plus an email link and auto-purges after ninety days. The visitor picks at the moment of submission. See the Privacy page for the full disclosure.
Things it's important to say plainly.
Indicators, not verdicts. The station names patterns; it does not diagnose people or systems. Every reading is framed as an observation — indicators suggest, the pattern is consistent with, the scale shows — and never as a determination. When a reading comes back hard enough that it warrants more than the station can offer (a licensed therapist with technology literacy, a specific support organization, legal help where the situation calls for it), the station says so plainly, and does not pretend to be that help itself.
This is an informational tool. It is not a diagnostic instrument, it is not a regulatory compliance certification, and it is not a substitute for professional psychological assessment. The scores are derived from self-reported inputs, not from direct observation of your AI system, and the results are indicative rather than definitive.
The station uses AI in its own pipeline to help analyze what comes through the door, and we don't hide that. We use AI to monitor AI, and the tension in that sentence is the whole point. Where AI ends and human oversight begins is not a detail — it is the job.