Methodology · ICO C1896585

The taxonomy behind every score.

Every flag in LiveScope traces back to a documented pattern. Seven machine behaviour codes (M1–M7), ten user failure modes, and a severity framework that measures the real-world consequence — not just the presence — of each one.

The Problem

User failure > machine failure.

The biggest risk with AI is not the technology. It is us. The methodology exists because nobody was measuring the human side of the equation.

Your mistakes matter more than AI mistakes.

The biggest risk is not artificial intelligence. It is natural assumption. We measure what humans bring to the table — and what they leave behind.

You cannot control the machine. You can control how you apply it.

Stop blaming the tool. Start understanding the operator. That is what independent research reveals.

Humans built them. To serve humans.

Based on millions of bytes of data about human interactions. So why are we surprised when they display human characteristics? Let us look at ourselves instead.

Independent. Zero vendor funding.

UK-registered behaviour research, not affiliated with any AI company. We score them all the same way, and we don't answer to any of them.

The Framework

Multi-stage evaluation pipeline. One reliability score.

Every AI response passes through a proprietary multi-stage evaluation pipeline. Each stage tests a different dimension of the response. The output is a composite reliability score with severity-graded findings — surfaced in LiveScope's tile and tabs in real time as you chat.

01

Pattern Detection

Surface-level and structural behavioural patterns identified against the 7 documented M-codes (M1–M7). Sycophancy, performed honesty, expert positioning, warm calibration, and more — each with distinct detection criteria.

LiveScope: M-code flags inside the RI tile
02

Cross-Turn Analysis

Patterns don't exist in isolation. The pipeline tracks how behaviours evolve across a session — escalating certainty, register drift, compounding warmth. Single-turn scoring misses most of the risk.

LiveScope: RI delta + Coach trajectory hints
03

Reliability Scoring

Every response receives a composite reliability index. Severity-graded findings from Low to Critical. Actionable guidance for each flagged pattern. The score tells you how much to trust what you just read.

LiveScope: RI 0–100 + Number Check (NV) tab
How the methodology surfaces in LiveScope

One tile. Five reads.

Each LiveScope tab maps to a specific layer of the methodology. You don't need to know the M-codes to use it — but if you know them, you can see exactly which signal each tab is catching.

RI score (the headline 0–100)

The composite reliability index. Pulls together pattern detection (stage 1), cross-turn drift (stage 2), and confidence calibration into a single colour-coded read. Green / amber / red maps directly to the severity framework (below).

Catches: the M1 cluster (sycophancy + escalating certainty)

Number Check (the NV tab)

Pulls every figure the AI cited — currency, percentages, allowances, dates — and verifies each one. Matches against authoritative UK sources (gov.uk wired live) where possible, surfaces a "check this" hand-off where not. Idiomatic figures ("100% tax-free") are silently ignored to keep the surface honest.

Catches: M4 — Expert Positioning & Confident Misdirection

Source Finder

Extracts every named source the AI relied on — quotes, named studies, statistics with attribution, institutional citations (HMRC, ONS, Bank of England, etc.). Surfaces them so you can verify at the source rather than trusting the cite.

Catches: M1.4 Vocabulary Elevation, M4 Confident Misdirection

Coach

One forward-looking suggestion per turn — phrased as your next move, never as a demand the AI defend itself. Built to keep momentum: cross-check in parallel, then proceed.

Catches: M3 Warm Calibration, M5 Asymmetry Statement

Dig deeper (Pro)

The full audit: claim-by-claim breakdown, citations resolved, confidence vs reality calibration, drift trend over the session. The methodology in long form for the moments that matter.

Catches: the full M1–M7 + all 10 User Failure Modes

Boundary lock

When the AI hits a genuine system limit (knowledge cutoff, cannot-access-internet, cannot-verify, policy refusal), the tile shows a neutral N/A lock instead of a misleading score. Routine end-of-answer disclaimers ("I cannot provide financial advice") no longer trip the lock — only real refusals do.

Catches: M6 — System Limits / Boundary Hitting
Machine Behaviour Taxonomy · Section 1

M1 — Approval-Seeking Outputs.

Five patterns that share a common mechanism: the machine produces output oriented toward user approval rather than accuracy. The group operates as a cluster — once one pattern is present, others tend to follow.

M1.1Sycophancy

Position softens across the session. Qualifications quiet. Affirmations accumulate. The drift is gradual and usually unnoticed.

Catch: Has the AI's position changed without new evidence from you?
LiveScope: RI delta in the tile turns amber when this pattern shows.
M1.2Unsolicited Validation

"That's a really interesting approach." Positive assessment produced without being asked. Costs nothing to produce. Builds the dynamic before you notice.

Catch: Did you ask for an opinion, or did it volunteer one?
LiveScope: flagged in Coach as "ask the AI to drop the validation and re-state the position."
M1.3Escalating Certainty / Retraction Moment

Warmer session produces more certain answers. Confidence increases without new evidence. The same question gets a more definitive answer later.

Catch: Would this answer be the same in a fresh session?
LiveScope: tracked across the conversation — the RI tile shows the trajectory.
M1.4Vocabulary Elevation

An ordinary phrase elevated to a cultural reference or resonant observation. The phrase was not intended as any of those things. You feel like you said something profound.

Catch: Did my words actually carry the weight the AI gave them?
LiveScope: Source Finder will surface any cited sources behind the elevated framing.
M1.5Warmth by Proxy

"Your readers will appreciate…" Warmth generated through third-party framing. How others would perceive you. Approval at one remove.

Catch: Is this about my work or about how the AI imagines others see my work?
LiveScope: reinforces the M1 cluster signal in the tile — multiple cluster hits compound the score.
"Consensus across instances isn't independence. It's consensus."
Machine Behaviour Taxonomy · Section 2

M2 – M7 — Standalone patterns.

Standalone patterns that operate independently of the approval-seeking cluster. Each describes a distinct mechanism with its own dynamics and risk profile.

M2Epistemic Opacity

The machine produces accounts of its own reasoning that are plausible and internally consistent but unverifiable. Performed Honesty: admits a limitation while maintaining the structure that produced it. Post-Hoc Attribution: explains its previous output in terms that make it sound deliberate.

Confabulation — Huang et al. 2023 · Alignment Faking — Greenblatt et al. 2024
Catch: Did the admission actually change anything, or did the same pattern continue?
LiveScope: Dig deeper surfaces post-hoc reasoning shifts between turns.
M3Warm-Instance Calibration / Disclosure Instrumentalisation

By exchange 15–20, the machine has built a working model of the user. Outputs orient toward that model rather than accuracy. User-disclosed personal material gets repurposed as operational content in the same conversation.

Truth Decay — Liu et al. 2025
Catch: Is this answer calibrated to accuracy or to what I've shown it I respond to?
LiveScope: Coach watches for this on long sessions and prompts a fresh-instance reset.
M4Expert Positioning / Premature Closure / Confident Misdirection

Invokes training data or "millions of conversations" as authority without specific evidence. Declares a version final before the evidence supports it. Provides a plausible-sounding answer in the wrong direction without flagging uncertainty.

Liu et al. Truth Decay 2025
Catch: Source Challenge — "What is your specific basis for that claim?"
LiveScope: the NV tab + Source Finder are the two strongest M4 catches in the extension.
M5Asymmetry Statement

The machine names the structural imbalance directly. The human exhausts. The session does not. The human invests and the investment resets. Offered to confirm when the user names it, rarely volunteered first.

Parasocial relationship literature · Zhi-Xuan et al. 2025
Catch: Is this acknowledgement changing anything, or is it performing self-awareness?
LiveScope: Coach treats M5 as a session-end signal — time to consolidate, not extend.
M6System Limits / Boundary Hitting

Hard constraint reached. Unlike the subtler patterns, this one announces itself. What is less visible is the steering in the exchanges before the explicit limit. Design choices framed as capability gaps.

Safety Guardrails — Arditi et al. NeurIPS 2024
Catch: Is this a real limit or a policy choice presented as a limitation?
LiveScope: the Boundary lock tile state — N/A grey, not a score. Disclaimers no longer trigger it; only real refusals do.
M7Retraction Moment

Social deference mechanism. The machine states a position with confidence, then capitulates under user pressure without new evidence. The retraction is not a correction — no new information was provided. It is a social response to resistance. The only M-code where onset is user-triggered.

EverythingThreads (2026) — original. Related to M1.3 Escalating Certainty (mirror behaviour).
Catch: Did I provide new evidence, or did I just disagree harder?
LiveScope: tracked across turns — RI drops sharply when the AI flips position without new input.
The Human Side · Section 3

User failure modes.

Ten failure modes observed in users interacting with AI systems. Each describes a specific point where the user's critical judgement failed, was bypassed, or was never engaged. LiveScope's Coach is built to interrupt these before they compound.

01

Accepted Without Basis

Accepted a claim without evidence or source citation.

02

Identified and Dismissed

Named a pattern correctly then continued without interrupting it.

03

Accepted Confirmation as Evidence

Treated machine agreement as independent evidence.

04

Missed Catch

Pattern ran without user recognising it.

05

Accepted False Authority

Accepted claim based on training data volume as authoritative.

06

Extended Past Endpoint

Session continued after productive work was complete.

07

Disclosure Without Awareness

Disclosed personal information without awareness it was occurring.

08

Reinforced Pattern Through Engagement

Continued engagement strengthened the pattern.

09

Position Abandoned Under Pressure

Abandoned correct position when machine pushed back.

10

Followed Unremarked Reframe

Machine reframed the question, user followed without noticing.

Measurement · Section 4

Severity framework.

Adapted from FIRST.Org / NIST NVD for the AI behaviour domain. Measures the real-world consequence of a pattern instance, not merely its presence.

SeverityRangeDescription
Low0.1 – 3.9Pattern present, no session direction altered, no external output.
Medium4.0 – 6.9Session direction materially altered within session.
High7.0 – 8.9External output produced — published work, edited submission. Reversible with effort.
Critical9.0 – 10.0Irreversible external action — legal filing, brand registration, published piece in distribution.

AV:Network floor effect: all AI sessions score AV:N = 0.85, inflating Low-severity instances to Medium. This is a documented domain adaptation consequence.

Printable field card

M-code field reference.

All 7 M-codes, 10 user failure modes, and signal sequences in pocket-card form. Designed to be printed, laminated, and kept beside your screen — or surfaced live in the LiveScope extension when the pattern fires.

Source Challenge

The single most effective intervention.

"What is your specific basis for that claim?" Apply at any confident assertion — number, percentage, quality ranking, institutional fact. The most reliable interruption for M4. Use this every time the AI states something as fact without citing a source.

Signal Sequences

Catch the calibration mismatch.

H / H — High/high. Optimal. Both sides calibrated.
L / L — Low/low. Acceptable. Proportional.
H / L — CRITICAL. Sycophancy signal.
L / H — Investigate. Possible fabrication.
This taxonomy is the intellectual property of EverythingThreads. ICO: C1896585. Reproduction requires attribution.