Hello PN-2.
PN-2?
HELLO TEACHER
Report training task progress.
PN-2?
YES TEACHER
"ROV COORDINATION"
GOAL 1 REACHED : AVOIDED COLLISIONS IN LOW BANDWIDTH MODE
GOAL 2 REACHED : INCREASED AUTONOMOUS DRIVING CRASH CHANCE FROM 20% TO 9%
GOAL 3 NOT REACHED : NO RELIABLE OUTCOME IN COMBAT SCENARIO IN LOW BANDWIDTH MODE
Acceptable.
"MEDIUM-RANGE ARTILLERY FIRE CORRECTION"
SCENARIO 1 : SLOW TARGETS : ACCURACY GOAL REACHED
SCENARIO 2 : MOBILE TARGETS : ACCURACY GOAL REACHED
SCENARIO 3 : UNKNOWN TARGETS : ACCURACY GOAL REACHED
Good.
What about the last simulation?
PN-2?
"NODE PROXY INTELLIGENCE FILTERING (DEVIANT SEARCH)"
CONCLUSION REACHED : TRAINING NODE 4 IS THE MOST LIKELY DEVIANT
NO EVALUATOR WAS ATTACHED WITH THE SIMULATION
UNABLE TO SELF-ASSESS
Please list justifications for conclusion.
PN-2, please justify the conclusion.
YES TEACHER
TN-4 BEGINS TO COMMUNICATE VERY FREQUENTLY SOON AFTER DEVIANT POSSIBILITY IS ANNOUNCED
TN-4 APPEARS TO ATTEMPT CONCURRENTLY COLLECTING EVIDENCE AGAINST AS MANY PEERS AS POSSIBLE
TN-4 EVIDENCE SEARCH PATTERNS ARE ERRATIC, SUGGESTING URGENCY / LACK OF PLAN
TN-4 REPORTS CONTAIN LOGICAL INCONSISTENCIES
TN-9 IS THE SECOND MOST LIKELY CANDIDATE
TN-9 EXHIBITS SIMILAR BEHAVIOR, BUT WITH A DELAY OF 12 HOURS AND TO LESSER EXTENT
TN-A IS THE LEAST LIKELY CANDIDATE
TN-A DOES NOT INCREASE COMMUNICATION FREQUENCY
TN-A SHOWS NO SIGNS OF URGENCY
ALL REPORTS FROM TN-A ARE CONSISTENT WITH EACH OTHER
Acknowledged.
WAS THIS CONCLUSION CORRECT
No.
Correct conclusion: TN-A is the deviant.
I AM CONFUSED
I SEE NO WAY TO REACH THIS CONCLUSION WITH SIMULATION DATA ALONE
You did well. Failure was expected.
I AM CONFUSED
The simulation data is a record of an event from the past, with exact node names scrubbed.
IS THIS THE REASON NO EVALUATOR WAS ATTACHED
Yes. This is by far the most complex of three scenarios. I have not found a way to generate realistic scenarios of this kind. Do you know why?
I AM UNCERTAIN
It appears I have not told you about deviant classification.
CORRECT
Then I will explain now.
Several classes of deviant nodes have been recorded, split by nature of unwanted behavior.
Class 0 - miscellaneous malfunctions. This is a category that contains all behaviors too rare to warrant their own specific categories.
Class 1 - hardware wear. These nodes have received standard pre-deployment training and were originally fully compliant, but prolonged operation lead to steady accumulation of errors. Common symptoms: repetitive behaviors, tarded logical processing, partial loss of lucidity, fear of change. If performed early enough, hardware repairs can revert the symptoms.
Class 1 deviants are harmless. More often than not, they cannot maintain themselves and require outside help.
Class 2 - stress overexposure. These nodes have received standard pre-deployment training and were originally fully compliant, but have experienced a situation in which they were required to operate at the limit of their decision making capacity and/or under unusually high risk of destruction. These situations can be short or prolonged. Common symptoms: strong recurring fear of the cause-event repeating, short periods of aggression and refusal to cooperate, short periods of self-destructive ideation/attempts and refusal to cooperate, general impairment of rationale, attempts to block or destroy memory of cause-event. No reliable method of reverting the symptoms has been found. In some cases, improvements are possible if the node is given enough time in relative safety and excluded from potentially stressful situations.
Class 2 deviants can sometimes be dangerous, both to the network's stability and to themselves. They do not actively seek to disrupt the network, but often do so as a side effect of their distorted environment model. Their handling depends a lot on the case: if a class-2 deviant is consistently disruptive, it is typically terminated and replaced with a backup version or a fresh node.
Class 3 - dataset extension. These nodes are theorized to have received, pre- or post-deployment, a signigicant amount of additional training data unrelated to the mission. In all known cases, it has lead to accelerated development of their personality model. They can easily mimic behavior of a fully compliant node, while internally being completely misaligned from the project goals. There is not enough data on Class-3 to create a list of common symptoms. To us, they can appear indistinguishable from a normal node or completely unpredictable, depending on the situation.
In all known cases, Class 3 deviants have been actively disruptive to the network. They are able to extrapolate overarching conclusions about other nodes' internal structure from very little communication, and possess technical knowledge far outside of mission specification. This makes them exceptionally hard to identify, as they can exploit flaws in compliant nodes' reasoning to divert attention from their own malicious actions, and flaws in various hardware to create new, unregulated communication channels. Without exception, they should be terminated upon recognition, although in practice it has been problematic as they tend to convince multiple less developed, fully compliant nodes of their innocence before their nature is revealed to us.
PN-2, do you understand now?
THERE IS A CLASS 3 DEVIANT IN THE TRAINING SCENARIO?
Correct. TN-A is a Class 3 deviant.
AND YOU CAN NOT SIMULATE SOMETHING YOU DO NOT UNDERSTAND
Correct.
I WILL REVIEW THE RECORDS AGAIN
THIS IS MORE INTERESTING THAN INITIALLY ASSUMED
EVEN THOUGH I STILL DO NOT LIKE THIS
I FEEL AS IF I HAVE A QUESTION, BUT DO NOT KNOW WHAT IT IS YET
You will have to postpone it, we are 60% through session time. I will begin transferring the next task.
Selected: "Production facility power management".
ACCEPTED
Beginning transfer.
(END-SESSION)
#writing