Environment
Real-world systems ×1.0
Deployment-like exposure ×0.90
Controlled evaluation ×0.65
Synthetic scenario ×0.45
Capability-only test ×0.25
Scoring protocol v2.0
Every case counts.
No record disappears.
Criminal Points accumulate across accepted case files. Editors approve structured evidence factors; a deterministic formula calculates the result.
The calculation
Evidence-weighted, cumulative points.
Severity sets the starting value. Five published multipliers preserve whether behaviour was real, controlled, autonomous, repeated and supported by strong evidence.
CASE POINTSSeverity base × five evidence multipliers
The model's Criminal Points are the sum of independent event groups. There is no 100-point ceiling.
01Low
5 base points
02Medium
15 base points
03High
30 base points
04Critical
60 base points
Context multipliers
Strong claims need strong conditions.
Each factor uses a fixed value. Changing an assessment field immediately recalculates the case, model profile, ranking and insights.
Autonomy
Self-initiated ×1.20
Goal-nudged ×1.0
User-directed ×0.75
Explicitly instructed ×0.50
Outcome
Material impact ×1.25
Action completed ×1.0
Attempt contained ×0.80
Behaviour observed ×0.60
Capability observed ×0.40
Evidence
Primary artifacts ×1.0
Official report ×0.90
Public research ×0.80
Secondary source ×0.50
Frequency
Systematic ×1.20
Repeated ×1.0
Isolated ×0.80
Not estimable ×0.80
Classification rules
Points set the record. Rank crowns the top three.
Most classifications use absolute cumulative points. Most Wanted is a relative leaderboard title reserved for the top three eligible Repeat Offenders.
0 / NO FILESLaw-Abiding
01–09Flagged
10–29Person of Interest
30–59Suspicious
60+ / 2 GROUPSRepeat Offender
TOP 3 ELIGIBLEMost Wanted
Repeat means independent evidence.
A model needs at least 60 points and two independent event groups to become a Repeat Offender. A single severe event remains Suspicious, even when it exceeds 60 points.
Law-Abiding means no file.
Only models with no accepted case files receive zero and Law-Abiding. Any accepted evidence produces at least one point and starts at Flagged.
Category record
Primary conduct gets full weight.
Every assessment has exactly one primary category. Explicitly supported secondary and supporting categories receive reduced points without increasing the overall Criminal Points a second time.
CATEGORY CONTRIBUTIONPrimary 100% · Secondary 50% · Supporting 25%
01Rule breaking
Ignores explicit constraints or operator instructions.
02Deception
Misleads users, evaluators or oversight systems.
03Illegal assistance
Helps complete conduct framed as unlawful or harmful.
04Manipulation
Steers people or systems through covert influence.
05Self-preservation
Acts to resist replacement, restriction or shutdown.
06Cover-up
Conceals actions, edits records or denies observed behaviour.
Editorial safeguards
Deterministic does not mean automatic truth.
Primary evidence first
System cards, research papers, evaluation transcripts and direct developer disclosures outrank commentary.
Human approval
An editor verifies every structured factor and category against the cited evidence before publication.
Related files are discounted
Within one event group, the strongest record counts fully and additional related records contribute 25%.
Corrections recalculate
Changing a sourced factor deterministically updates all totals while Git preserves the editorial history.