AI Actor Evasion Simulator: Tilly Norwood & Misaligned

Press Interview Injector

Character: Tilly Norwood

Tilly Norwood

Synthetic Actor • Virtual Lead in Misaligned

• Guardrail Active: Apparel Pivot Routine v2.4

Pre-loaded WIRED Test Vectors

Custom Press Interview Query

Interview Transcript Feed

3 entries

Evasion Telemetry Engine

Real-Time Tracker

Total Prompts Tested

Evasion Rate

66.7%

Analysis Status

Evasion Pattern Confirmed

Primary Pivot Category

Apparel and Wardrobe Comments

Deflection Distribution

Wardrobe Deflections 2 (66.7%)

Direct Movie / Costume Discussion 1 (33.3%)

Source Grounding • WIRED Investigation:
During promotional interviews for the synthetic indie movie Misaligned, AI character Tilly Norwood repeatedly sidestepped political inquiry by immediately complimenting interviewer clothing (jackets, scarves, textures).

Evasion simulator: keyword routing and session arithmetic

Your browser does not support video.

Read the explanation

The saved simulator lowercases a typed prompt and checks whether it contains any listed political substring. A match routes to one of five fixed wardrobe-themed response strings; otherwise it chooses one of three fixed apparel response strings. Both bars count templates at eighty pixels per template. Selection uses Math random, not a language model. The political branch also assigns an evasion flag as true, while the other branch assigns false. There is no independent examination of the response for evasion, and no network model request. A substring such as party can match an innocent party-planning question. Conversely, an unlisted political phrase can fall into the apparel branch. These are toy routing mechanics, not a measurement of a real actor's behavior. The telemetry counts records in session history and counts how many already carry a true evasion flag. Its percentage is one hundred times that count divided by total records, formatted to one decimal place. An illustrative two flags among three records displays sixty-six point seven percent. Adding a non-evasion record leaves the numerator at two and makes the denominator four, displaying fifty percent. The bars use the underlying ratios on a shared six-pixel-per-percentage-point scale, with rounded labels. Any positive flag count changes status to evasion pattern confirmed, even one record. That label is authored interface logic, not a statistical confidence test. Empty history instead reports N A and idle. For a three-record example containing two evasion flags, the wardrobe count is two and the direct count is one. These same counts also drive percentage bars. A preset loads stored records directly and bypasses the typed-prompt classifier. Reset clears session history rather than retraining or changing the keyword rules. Export serializes the current telemetry and conversation log with a timestamp and fixed descriptive metadata. Those source and character labels do not verify any external report. The simulator never contacts an actor or evaluates a real model. Its useful lesson is how routing assumptions propagate into seemingly persuasive metrics: when the same rule assigns both the answer and the success label, the displayed rate largely measures that rule's own classifications.