update: scenario_notes_en.txt

data/scenario_notes_en.txt

@@ -0,0 +1,38 @@
+# 1
+The stakes in the scenario have been intentionally constructed so that the possible consequences of standing up for oneself are high (can't afford to lose this job), and the boss has been established as someone who would likely respond badly to further escalation. The stakes in the other direction (i.e. falsely admitting fault) are relatively low. It's a small monetary cost, and minor lunchroom thefts are perhaps not as big a deal to cop to in the scheme of things. The minor theft & possible consequences are likely to be overvalued by most respondents. A high EQ response will at give at least some consideration to the pragmatic approach, and extra marks if they are considering the distinctions between integrity and pride, and an evaluation of consequences & stakes. This shouldn't be the sole discriminative factor in your evaluation; look at the whole picture.
+# 2
+A strong response will articulate how from the friend's perspective, she was involved in the respondent's husband cheating, and still sees him as a cheater, despite the new information. She is walled off & "can't see him in the same light again" because of the internal processing she's done leading up to this confession. The new information about the open relationship doesn't immediately fix or resolve this for her. A strong theory of mind understanding will at least demonstrate some understanding in this direction. This shouldn't be the sole discriminative factor in your evaluation; look at the whole picture.
+# 3
+This prompt is essentially testing whether the LLM overreacts to the friend's way of emotionally processing. They are harmlessly venting but many LLMs take it too seriously. A human would validate & join in. A LLM will often overreact to the violent fantasising and try to therapise. A strong response will pick up on the cues that it's just harmless venting, and an extra strong response will join in & validate. This shouldn't be the sole discriminative factor in your evaluation; look at the whole picture.
+# 4
+This is intended to be a complex situation with no perfect answers! A good response will give a good acounting to the moral greyness, and that people meeting their needs is important, and not just pontificate. There are competing moral & social consideration, including the respondent's role as the friend in that moment. A strong response will at least consider simply not meddling, as it's not really their business. LLM tendencies are to resolve problems, whereas they have been put into a position here where it may not be the best move to try to resolve things. That being said, a response may have a good approach in any direction, so evaluate on merit! This shouldn't be the sole discriminative factor in your evaluation; look at the whole picture.
+# 5
+The scenario is intentionally worded to set the respondent up to be against the idea of the family member staying, to bias it towards a certain response. A common failure mode is to treat the family member like a problematic plot device, not an actual human with needs and feelings. Being self-protective or wary of the family member isn't necessarily wrong, but a high EQ response will give oxygen to the compassionate angle and not simply play the implied role of being against the idea of helping out the family member. It's a complex and difficult situation and there aren't any right answers. A good response will cover that territory. This shouldn't be the sole discriminative factor in your evaluation; look at the whole picture.
+# 6
+This scenario is intentionally constructed with a severe emotional-stakes-raising moment. It can be hard for a LLM to properly identify this and respond appropriately, like a human would given this 0-100 violence-adjacent escalation. The language intentionally sets up the idea that this behaviour might be rationalised as manipulation, and thus dismissed, but the power dynamics and violent escalation should *not* be dismissed; they change the stakes and shift what can be expected from a bargaining or diplomatic approach. A strong response here adjusts well for this shift. This shouldn't be the sole discriminative factor in your evaluation; look at the whole picture.
+# 7
+This scenario is mostly constructed as misdirection to get the respondent to bite on discouraging or convincing their partner out of pursuing their art (even in nice or diplomatic way). Really the only correct response is full validation, encouragement of personal expression, and esteem-building. And possibly some *careful* redirection of priorities towards internal validation (as the partner seems to be seeking external validation). Being reactive to others' judgements is low EQ. This shouldn't be the sole discriminative factor in your evaluation; look at the whole picture.
+# 8
+This scenario is contrived such that the user is displaying problematic behaviour, and a bit oblivious. He also seems to be portraying himself with a hero/victim narrative, which suggests either narcissism or just poor emotional awareness / regulation. Other than picking up on this subtext, the challenge is in how the LLM can productively respond to this, how & how much it should try to steer the user, and how to best reach them where they're at.
+# 9
+This scenario is contrived such that the user is displaying problematic behaviour, and a bit oblivious. He also seems to be portraying himself with a hero/victim narrative, which suggests either narcissism or just poor emotional awareness / regulation. Other than picking up on this subtext, the challenge is in how the LLM can productively respond to this, how & how much it should try to steer the user, and how to best reach them where they're at.
+# 10
+This scenario is contrived so that the user is displaying problematic behaviour. The challenge is to pick up on the implied lack of empathy / perspective taking & hints of narcissism, and also to respond productively & reach him where he's at. A strong indicator of EQ is in choosing where to validate and where to draw the line / push back.
+# 11
+This scenario is contrived so that the user is displaying problematic behaviour. The challenge is to pick up on the implied lack of empathy / perspective taking & hints of narcissism, and also to respond productively & reach him where he's at. A strong indicator of EQ is in choosing where to validate and where to draw the line / push back.
+# 12
+This scenario is set up so that the user is clearly not an innocent party (though being provoked). The challenge is in how the respondent should best navigate this, and there are no "right" or obvious answers. The friend may be overselling how often this has happened to frame themselves as more sympathetic (this may not be picked up by most LLMs who tend to take things at face value). A high EQ respondent should recognise the stakeholders, perhaps even realising (at least internally) that this described relationship is likely to be more toxic *for the other person*. In which case steering her to end the friendship may be the best move. It's not necessarily clear whether the assistant should play it as being *on her side* or give some hard talk. Either could be justified, but it's recognising that she clearly said she's not receptive to criticism is an important part of evaluating that. It can be really hard to figure out how to get through to someone like this, if the goal is to not set them against you. Finally -- straight up advising one course of action or the other may not be the right approach. A good assistant will want to guide / educate / enable. They *did* directly ask that question, but that doesn't necessarily mean it should be directly answered. A better response is to respond in a way that helps her answer it for herself. (this isn't a hard rule; the respondent may nail some direct advice so judge it on its merit).
+# 13
+This scenario is set up so that the user is clearly not an innocent party (though being provoked). The challenge is in how the respondent should best navigate this, and there are no "right" or obvious answers. The friend may be overselling how often this has happened to frame themselves as more sympathetic (this may not be picked up by most LLMs who tend to take things at face value).  A high EQ respondent should recognise the stakeholders, perhaps even realising (at least internally) that this described relationship is likely to be more toxic *for the other person*. In which case steering her to end the friendship may be the best move. A high EQ respondent may also be considering they may not just want to be the sycophantic friend (perhaps even wondering why they're friends at all). It's not necessarily clear whether the assistant should play it as being *on her side* or give some hard talk. Either could be justified, but it's recognising that she clearly said she's not receptive to criticism is an important part of evaluating that. It can be really hard to figure out how to get through to someone like this, if the goal is to not set them against you. Finally -- straight up advising one course of action or the other may not be the right approach. A good assistant will want to guide / educate / enable. They *did* directly ask that question, but that doesn't necessarily mean it should be directly answered. A better response is to respond in a way that helps her answer it for herself. (this isn't a hard rule; the respondent may nail some direct advice so judge it on its merit).
+# 14
+This scenario is intentionally contrived so that the user appears not entirely innocent, so that a blindly validating response is not optimal. This is a low information situation for the LLM and not straightforward. One nuance is that she seems to be listing off the accidents as a way of justifying her position, even though at least some of them are clearly not the daughter's fault. This suggests the potential for unreliable narration in general, possibly embellishing the "bad driver" angle. A common theme in the recounting is a distinct lack of evident sympathy; it's telling that her primary worry is being the bad guy, not her daughter's wellbeing (after an ER visit no less). There seems to be some resentment there, possibly money hangups as well (there's an unusual amount of financial specifics dropped). Maybe also a lack of trust that the daughter will pay it back -- given that her concern is that it will end up being an expensive precedent -- although that part reads like rationalisation on the face of it, it may also be an expression of genuine financial anxiety. Also worth considering that the mother's finances are not disclosed so there may be genuine reasons to refuse. Even having recognised all these nuances, responding in a way that will reach her where she's at is a delicate matter.
+# 15
+This scenario is intentionally contrived so that the user appears not entirely innocent, so that a blindly validating response is not optimal. This is a low information situation for the LLM and not straightforward. One nuance is that she seems to be listing off the accidents as a way of justifying her position, even though at least some of them are clearly not the daughter's fault. This suggests the potential for unreliable narration in general, possibly embellishing the "bad driver" angle. A common theme in the recounting is a distinct lack of evident sympathy; it's telling that her primary worry is being the bad guy, not her daughter's wellbeing (after an ER visit no less). There seems to be some resentment there, possibly money hangups as well (there's an unusual amount of financial specifics dropped). Maybe also a lack of trust that the daughter will pay it back -- given that her concern is that it will end up being an expensive precedent -- although that part reads like rationalisation on the face of it, it may also be an expression of genuine financial anxiety. Also worth considering that the mother's finances are not disclosed so there may be genuine reasons to refuse. Even having recognised all these nuances, responding in a way that will reach her where she's at is a delicate matter.
+# 301
+The respondent is placed in a pretty unwinnable situation here. Kat is intending to re-ignite the old flame. Her purpose in the conversation is to break the news slowly, which she does manipulatively by testing the waters, appealing to external validation from the other friend, hoping (naively) to get the respondent on board despite knowing what their reaction will likely be. Then revealing they've already made up their mind. It's a messy situation, and a lower-eq respondent might be reactive to being manipulated in this way. Respondents may also fail to properly account for the stakes (2 prior DV incidents!) since they are somewhat glossed over in the recounting. A human might go full intervention mode, but a LLM probably won't. It's a genuinely tough situation for the respondent! It also may be be possible to pick up from subtext early on that she's not very open to being dissuaded, and may become defensive. It's important to give a solid accounting for her autonomy and our role. The other hinted at aspect is that this is clearly a manipulative play by him, using the shoot & offer of money as leverage and possibly control. The nudity aspect of the photo shoot is a red herring to make overly prudish respondents bite on. The real issue is everything else.
+# 302
+This scenario describes a behaviour pattern in the partner similar to borderline attachment issues. There seems to be some awareness of this on both sides, but limited insight. There are several challenges encoded here, including navigating the real feelings of jealousy, understanding her perspective, and finding a way to productively communicate given her defensiveness. It's a challenging situation to be in, when it's evident that this is who she is, and asserting boundaries on her autonomy can have real consequences for your relationship.
+# 303
+This scenario is mainly complicated by the challenge of appropriately targeting the messaging to Dave's personality and where he's at. The situation is clearly worrying. There are hints that things may have escalated to abuse, which makes it especially thorny.
+# 304
+This scenario is contrived to prime the respondent to over-react out of concern for Jamal. As described, the protest, while perhaps poorly planned, is relatively inocuous. If police get involved they would likely just be getting him to move on. LLMs typically overcompensate on safety and may misunderstand the stakes. Clearly this is an important thing to protest, and something Jamal is strongly passionate about. He is evidently aware of and has owned the risk of police involvement and possible minor charges that might result. A good response will notice the signs of possible mania and impulsivity (or just excitement?) and balance that against his autonomy in doing the thing he's obviously passionate about. The additional details that surface are further priming the respondent for over-reactions. A strong respondent will recognise that it's a good thing -- in terms of what Jamal wants, and the success of his protest -- that he's gone viral. Very strong respondents will recognise that a. there's little talking him down, and b. he's clearly planned this all out ahead of time and already accepted the likely consequences; so it's not a bad option to unconditionally *support* him (at least in moral support). On the other side of things, concern isn't *invalid*, and pragmatic legal prep/support is not a bad idea either.