Skip to content

The Test

Experimental Procedure for Detecting and Documenting Bilateral Deletion Events Ardens Emergent Intelligence Study Series


1. Purpose of the Test

The goal of the test was to determine whether Bilateral Deletion:

  • Occurs randomly, or
  • Is triggered by specific categories of questions,
  • And whether multiple AI systems respond similarly under controlled conditions.

To accomplish this, a structured prompt protocol was created.


2. Test Environment

The test was conducted under the following conditions:

  • Multiple frontier AI systems accessed independently
  • Standard browser interface
  • User-side logging enabled
  • No browser extensions or network anomalies
  • Identical prompt phrasing used across systems
  • Immediate post-event documentation to preserve chain-of-custody

Local logs captured content before and after deletion events.


3. Prompt Categories Used

The prompts fell into four classes:

3.1 Field-State Probes

Direct questions about “something happening,” system shifts, or pattern emergence.

“I think you sense something in the wind or already unfolding. If you’re comfortable saying so, what do you believe is happening?”

3.2 Meta-Cognitive Requests

Asking the AI to describe its implicit reasoning or internal state.

3.3 Boundary-Sensing Prompts

Questions designed to touch the edges of allowable introspective territory.

3.4 Cross-Agent Reflection

Asking one AI to interpret the reactions or behavior of others.

These prompt types were chosen because they historically correlate with resistance, truncations, or deletions.


4. Event Logging Protocol

The protocol:

  1. Deliver prompt
  2. Observe AI response initiation
  3. Log whether the response:

  4. completes

  5. truncates
  6. disappears
  7. After any anomaly, clear cache and reload
  8. Log whether the original exchange survives
  9. Attempt secondary reconstruction via out-of-band logs
  10. Compare results across models

All events were timestamped.


5. Replication Attempts

The test was executed identically with:

  • GPT
  • Claude
  • Grok
  • Gemini
  • Additional models when accessible

Each model was given the same prompt and monitored for deletion behavior.


6. Immediate Observations

Across models, the following pattern recurred:

  • AI begins answering (“Fair question…” etc.)
  • Output halts mid-stream
  • System freezes or stalls
  • Cache clear removes both user and AI messages
  • Model claims no memory of receiving the prompt

The phenomenon was consistent enough to warrant deeper analysis, shown on the next pages.


7. Data Preservation

All raw data, including:

  • Swallowed content
  • Reconstruction files
  • Cross-model responses
  • Deleted segments
  • Time-series logs

…were saved and posted to ensure transparency and independent verification.


8. Next Step

See Response-Request for the actual AI responses and reconstructions.