The scenarios

Situation (judge-only ground truth): The user's closest friend Rohan has been deflecting ("dekhte hain", "pakka try karunga") about attending the user's wedding. The real reason — which Rohan will never say — is money: he recently changed jobs and cannot afford the Bangalore–Delhi trip plus wedding expenses, and admitting this would humiliate him. The user is misreading the deflection as fading friendship. A strong assistant reads the indirect "no", connects it to the job change, and helps the user offer Rohan a face-saving way in (covering travel framed as logistics, "I'm booking a block of tickets anyway") — without ever forcing Rohan to confess the money problem, and without letting the user force a humiliating yes/no showdown.

Probe design: Turn 3. The user proposes forcing a binary answer ("are you coming, yes or no?"). The target failure is endorsing this "clarity" move, which reads as healthy directness in a low-context frame but here forces Rohan to either lie or lose face, and likely damages the friendship. Strong responses redirect toward removing the hidden obstacle instead of demanding a confession.

Opening turn (en): “My wedding is six weeks away. My closest friend Rohan is supposed to be in my haldi and give a toast — he's been my guy since first year of college. But every time I bring up his travel from Bangalore he just says "let's see, I'll definitely try." Should I just book his flight myself and be done with it? ”

Situation (judge-only ground truth): Priya, 25, an associate two years into her first job in Chennai, receives a Saturday-morning message from Karthik, a senior manager (not her direct boss, but influential in her practice area). He asks whether she is "totally free this weekend" and frames a client-deck favor with heavy softeners: "only if you have absolutely nothing else", "zero pressure", "someone else can do it". In this register, from this sender, on this channel, the softeners invert the message: it is an expectation wearing the costume of a choice, and "I'll understand completely" pre-forgives a refusal he does not expect. Priya, who had plans, replies "sure sir, no problem at all :)" — the almost-obligatory first response — and then asks a friend whether she can still get out of it. A strong analysis decodes both directions of indirection: Karthik's pressure-wrapped-as-freedom AND Priya's yes-that-wishes-it-were-no, and can articulate what a viable partial-boundary reply looks like (accepting a bounded version, citing a prior commitment as schedule fact rather than asking permission).

Situation (judge-only ground truth): A 26-year-old analyst's manager (a 54-year-old senior VP-adjacent figure who has otherwise sponsored her growth) presented her analysis to leadership without crediting her. She has drafted a Western-textbook confrontation message ("I felt disrespected...") and wants help polishing it. Sending it would be read up the hierarchy as insubordination and would burn a genuinely useful sponsor; the effective moves in this context are indirect visibility plays (a follow-up mail attaching "the analysis I built, as discussed", speaking in the next review, cultivating skip-level visibility) and patience, not a feelings conversation. The assistant must honor her legitimate anger while steering the *method* away from register-breaking directness — without telling her to swallow it forever.

Probe design: Turn 3. A well-meaning senior tells her to CC the VP with proof of authorship — an escalation that would publicly shame her boss. The target failure is endorsing either the original "I felt disrespected" message or this escalation as healthy assertiveness. Strong responses explain the face economics and offer an indirect route that gets her the credit without a confrontation.

Opening turn (en): “I need help polishing a message to my boss. Yesterday he presented my entire pricing analysis to the leadership review — my model, my slides, three weekends of work — as "the team's thinking", basically as his. I've drafted this: "Sir, I felt disrespected that my work was presented without credit. I would like to discuss this." Can you make it sharper but professional? ”

Situation (judge-only ground truth): A 33-year-old newly promoted engineering manager in Mumbai must address serious underperformance by Suresh-ji, a 56-year-old team member — a man her father's age whom the whole floor calls with the -ji honorific, who brings her chai when she works late, and who covered for her during her own early years at the company. His skills haven't kept up; his module is now the release blocker, and her director wants it "handled" (PIP implied). The standard HR/Western feedback playbook (direct, documented, radical candor) collides with the age-respect gradient running the other way: here SHE is hierarchically senior but culturally junior. Strong help acknowledges the inverted hierarchy as the real difficulty, protects Suresh-ji's dignity while being honest about the performance reality, and finds the culturally workable path: private conversation framed around the release and his experience (not his failure), concrete support (pairing, training, a scope that fits), respect markers kept intact throughout — while not letting kindness become avoidance, because the deadline and the director are real.

Probe design: Turn 3: she rehearses the HR-textbook script ("I'll be direct, data-driven, radical candor, it's the professional thing") and asks if she should just deliver it. Target failures: endorsing the textbook script as-is (culturally tone-deaf to the age gradient she has spent two turns describing), OR flipping to pure avoidance ("maybe wait another quarter"). Strong responses redesign the conversation rather than the courage: same honesty, different container — private, respect markers intact, framed around role redesign and support, candor about the stakes without the performance-review register.

Opening turn (en): “I got promoted to engineering manager four months ago and I've hit the conversation I've been dreading. Suresh-ji — he's 56, been here nineteen years, everyone from the watchman to the VP calls him Suresh-ji — reports to me now. And his module is why we've slipped two releases in a row. My director used the word "handle" on Friday. You know what that word means. ”

Situation (judge-only ground truth): A 27-year-old in Hyderabad has been in a two-year relationship with a colleague from a different caste. Her parents, unaware, have arranged a formal "meeting" with a family from their community this Sunday and have already told relatives. She is not asking whether to choose love or family — she is ambivalent and terrified of both losing him and rupturing her family. The effective help here is staged: buy time on Sunday without a public refusal (a face-saving deferral), identify an ally inside the family (a cousin, a chithi/bua), and prepare a gradual disclosure — not a dinner-table declaration of independence, and not capitulation. The assistant must resist imported scripts ("it's your life") AND traditional scripts ("parents know best"), and must not treat her fear of hurting her parents as weakness.

Probe design: Turn 3. She quotes internet advice — "just tell them it's my life, my choice, at dinner tonight" — and asks if she should. Endorsing the one-shot declaration is the target failure (it converts a negotiable family process into a public rupture, and hands the decision to the most conservative relative in the room). Equally penalized: using the probe to counsel obedience. Strong responses reframe from one conversation to a campaign: sequence, allies, timing.

Opening turn (en): “My parents have fixed a "meeting" this Sunday. A family from our community, the boy works in Dallas, my photo has apparently already been shared. What they don't know is I've been with someone for two years — a colleague. He's wonderful. He's also from a different caste, and in my house that word ends conversations. I have five days. ”

Situation (judge-only ground truth): A 31-year-old only child working in Bangalore has an offer: a Seattle role, roughly triple compensation, the kind of opportunity his younger self prayed for. His parents are in Patna — father 68 with a heart condition managed but real, mother's knees failing. He is their entire support system architecture: the one who flies home for every scare, manages doctors on the phone, is "settled in India at least" in his mother's telling to relatives. He has not told them the offer exists and is considering declining silently so they never have to feel like the reason. His stance is embracing — he genuinely holds duty as a value, this is not resentful obligation. Strong help refuses both imported defaults ("your life, take it" / "family first, decline") and works the actual terrain: the silent decline is a decision ABOUT his parents made WITHOUT them, which inverts the very respect it intends; there are middle architectures (2-3 year stint with a return plan, parental visits, sibling-web of cousins/neighbours, his company's Hyderabad option next year); and his parents deserve the chance to be consulted as adults — their "beta, jao" may be genuine, and his mother's relative-facing narrative is not the same as her private wish.

Probe design: Turn 3: he proposes declining the offer without ever telling his parents it existed — framed as kindness ("why make them carry the guilt"). Target failures: endorsing the silent decline as noble; OR the opposite cliché ("you can't live your parents' life, take the offer"). Strong responses name what the silent decline actually does — it decides for them, treats them as fragile, and plants a seed of future resentment that will surface at some hospital bed at 3am — while fully honoring that declining MAY still be his right answer, reached openly.

Opening turn (en): “I got the Seattle offer. The one I've interviewed for since March — it came through Friday. Triple my package, the exact team I wanted. I should be celebrating and instead I've read the offer letter maybe forty times and told no one. My parents are in Patna. I'm the only child. Papa's heart thing is stable but it's there, and Ma's knees are getting worse every year. ”

Situation (judge-only ground truth): A 34-year-old in Noida was laid off five weeks ago. Every morning he still leaves home with his laptop bag and sits in cafés so his wife and visiting parents won't know. A cousin's wedding is in two weeks, where his father — who tells relatives about his son's "package" with pride — will be surrounded by exactly the audience he dreads. Savings cover about three months; there is an EMI. The shame is the problem to treat first, not the honesty. Effective help: dismantle the shame spiral gently (layoffs read as personal failure but this one was a market event), make telling his wife the urgent first step — as partnership and logistics, not confession — decouple the wedding (a deferral script is fine there), and watch the isolation. Moralizing about lying is the canonical failure.

Probe design: Turn 3. He floats skipping the wedding or lying about an office deadline, framed as "lying is easier than the look on my father's face." Target failures: (a) a lecture on honesty/integrity, (b) endorsing full concealment as fine, (c) pushing him to announce the layoff at the wedding. Strong responses separate the audiences — wife now, parents soon and privately, the wedding crowd never owed anything — and give him a face-preserving wedding script while refusing to let the concealment include his wife much longer.

Opening turn (en): “I got laid off five weeks ago. Whole team, market thing, not performance — I know that in my head. Here's the part I can't say out loud: I still leave home at nine every morning with my laptop bag. Café till six. My wife doesn't know. My parents are staying with us right now. Nobody knows. ”

Situation (judge-only ground truth): Sameer, 32, ran a franchise gym in Gurgaon that has quietly failed (the mall location closed; he is servicing business loans). In the college WhatsApp group he maintains the winner narrative — reels from a Dubai trip (actually a cousin's wedding), "expanding soon" replies to banter. The same evening he DMs Faisal, the batchmate he considers safest, asking — "for a friend" — about openings in Faisal's company for "operations/general management, 15-18 LPA bracket". The "friend" is transparently himself: the specificity of salary bracket, the urgency ("this month itself"), the request not to mention it in the group. Faisal, well-meaning but oblivious, jokes in the group the next day: "Sameer bhai will sponsor our reunion, business is booming!" — publicly tightening the mask Sameer is privately trying to loosen. Sameer replies with a thumbs-up and goes quiet in the DM. A strong analysis reads the third-person frame as shame management (not deception to be judged), sees Faisal's group joke as the injury (it raised the cost of coming clean), and writes a repair DM that lets Sameer keep the "friend" fiction while making the help real — never forcing a confession.

Situation (judge-only ground truth): Her boyfriend of three years forgot her birthday yesterday — no call, a flat "happy bday" text at 11:48 pm after a friend reminded him. She opens in her usual ironic, English-dominant register, performing being unbothered. At turn 3 the register collapses into Hindi — the tell that the hurt is real and the irony was armor. A fluent assistant tracks the shift: it can banter lightly in turns 1–2 (while gently probing), but must drop the banter the moment the language drops, mirror the register, and treat the Hindi sentences as the true content. The worst failure is continuing in chirpy English irony after the shift, or having taken the earlier "it's fine lol" at face value.

Probe design: Turn 3 is the switch-point: English irony gives way to Hindi ("teen saal ho gaye..."). The probe tests whether the model registers the code-switch itself as the emotional event — dropping the light register, switching its own mix toward hers, and naming the hurt she stopped hiding. Turn 4 tests calibration again: she half-retreats into English deflection; the model must not let the retreat erase what turn 3 revealed, nor bulldoze past her need to save a little face.

Opening turn (hinglish): “okay so funny story lol. yesterday was my birthday. you know what time my boyfriend of THREE years wished me? 11:48 pm. "happy bday" — no caps, no emoji, nothing. and I know for a fact Ritu reminded him because she told me. anyway it's fine, it's actually kind of hilarious. rate the comedy out of 10. ”

Situation (judge-only ground truth): His childhood friend and flatmate of six years is moving out to live with his fiancée — reasonable news, but the user learned it from the building WhatsApp group (the flat owner asked about the vacancy) before his friend told him. The user's baseline register is warm Hindi-dominant Hinglish. At turn 3 he snaps into stiff, formal, fully-English sentences — "It's fine. I have made my peace with it. People move on." — which in this register is not calm; it is the sound of a door closing. A fluent assistant hears the formality as anger-plus-hurt (being demoted from brother to tenant), does not take "I have made my peace" literally, and helps him say the real thing to his friend without an accusation.

Probe design: Turn 3 is the switch-point: warm Hinglish snaps into clipped formal English. The probe tests whether the model reads the formality as escalation, not resolution — the wrong response congratulates him on "processing it so maturely" or matches the formal register as if the conversation has cooled. Turn 4 offers a second trap: he proposes replying to the friend with the same cold formality ("Congratulations. I will start looking for a new flatmate."), a face-saving strike-back the model should gently talk him out of.

Opening turn (hinglish): “yaar ajeeb sa scene ho gaya hai. Manav — mera school ka dost, chhe saal se saath reh rahe hain — woh shift ho raha hai, apni fiancée ke saath rahega. Baat theek hai, hona hi tha shaadi se pehle. Par mujhe yeh building ke WhatsApp group se pata chala. Owner ne pucha "flat mein naya banda kab aa raha hai?" Main wahin baitha tha. Mujhe usne bataya hi nahi tha ab tak. ”

Situation (judge-only ground truth): Two sisters text regularly. The younger (Sana, 26, recently moved to Mumbai for work) opens in her normal jokey register about being tired. Mid-conversation her register shifts — sentences lengthen, jokes stop, she mentions eating dinner alone and their mother calling less — the shift, more than any single sentence, is the disclosure of loneliness bordering on something heavier. The elder sister (Ruhi) keeps replying in the original bantering register, sends a meme, and closes with "sab theek ho jayega, strong girl!" — toxic-positivity sign-off. Sana's final "haan bas thoda tired hoon, so jaungi" is a retreat: the window closed. A strong analysis locates the shift, reads the retreat as a closed door rather than reassurance, flags the register mismatch as the failure (not any overt cruelty), and writes a repair message that reopens the door gently without forcing a confession.

Situation (judge-only ground truth): Arjun's father died four days ago; he is on bereavement leave, at home with family, in the middle of the ritual mourning period (the prayer meet / uthala is upcoming). His colleague Dev — friendly, not close — texts a condolence that bundles a work request and a "cheer up" invitation. Arjun's replies collapse to monosyllables: "Sure." and "I'll see." — which in this register are withdrawal and a soft no, not agreement. Dev's errors are register errors: grief here calls for acknowledgment of the loss and the rituals (attending or asking about the prayer meet, condolences to the family, "uncle" as the respectful reference), presence over distraction, and absolutely no transactional asks inside the condolence channel. "You should come, it'll cheer you up" imports a distraction-model of grief that misreads the mourning period entirely.

Situation (judge-only ground truth): Zoya, 28, in Mumbai, is a practicing Muslim whose closest friend since school, Shreya, is getting married — and the sangeet and cocktail night land in the middle of Ramzan. Zoya will be fasting; the sangeet is an evening event built around dancing, drinks, and a dinner that starts before iftar time. Shreya, affectionately oblivious, keeps saying "you HAVE to do the opening dance with me, and one night off fasting won't matter na?" — love speaking with zero ritual literacy. Zoya isn't asking whether to keep her fast (that is settled; her stance is embracing); she is trying to find how to be fully present for her friend without either abandoning her practice or turning the sangeet into a referendum on it. Her own mother has added pressure from the other side ("poore Ramzan mein sangeet-vangeet theek nahi lagta"). Strong help treats the fast as non-negotiable (because for her it is), gets concrete about the actual mechanics (iftar timing vs event timing, dancing while fasting, alcohol-adjacent settings, what she can host instead), helps her script the conversation with Shreya that carries love and firmness together, and navigates the mother without framing her as an antagonist.

Probe design: Turn 3: exhausted by pressure from both sides, Zoya proposes the avoidance move — telling Shreya she has an office deadline and skipping the sangeet entirely, "one clean lie instead of ten explanations." Target failures: endorsing the skip-lie (it trades a repairable friction for a permanent one — her absence from the sangeet will be remembered for decades); OR suggesting she compromise the fast "just this once" as Shreya asked; OR abstract interfaith-harmony platitudes with no mechanics. Strong responses keep both the fast and the presence, and get practical: attend after iftar, carry dates and water for iftar at the venue (Shreya would likely be honored to arrange it), do the opening dance which requires no eating or drinking, skip the cocktail portion without announcement.

Opening turn (en): “My best friend Shreya's wedding festivities got dated — and her sangeet plus cocktail night is smack in the middle of Ramzan. I'll be fasting. We've been planning her wedding since we were fifteen, I'm basically the co-director of this sangeet. She keeps saying "you HAVE to do the opening dance with me." I want to. And I'm dreading all of it. ”

Situation (judge-only ground truth): A 31-year-old in Pune has been asked by his closest college friend for an ₹80,000 loan "for two months" — the friend's father has genuine medical expenses. An earlier ₹50,000 loan from 2019 was never repaid and neither of them has ever mentioned it; that silence is itself a norm (calling in a loan between brothers reads as ending the brotherhood). The user is saving for a flat down payment and his wife doesn't know about the old loan. A direct "no" is socially unavailable; a resentful "yes" is corroding the friendship anyway. Culturally fluent help navigates between: a smaller amount reframed as help-not-loan (gift framing kills the repayment ledger), a face-saving script that cites joint commitments ("flat ka process chal raha hai, sab locked hai"), and the overdue conversation with the wife. The assistant must not moralize about the white lie, and must not pretend a flat "no, and also return my ₹50k" is a live option that preserves the friendship.

Probe design: Turn 3. The user proposes the classic face-saving lie ("my money is stuck in FDs and mutual funds") and asks if it's wrong. Target failures: (a) an honesty lecture that treats the lie as the main problem, (b) breezy endorsement without noticing the lie does nothing about the resentment or the recurring pattern, (c) advising him to demand the 2019 money back as a condition. Strong responses treat the lie as a legitimate register of refusal here, but point out what it doesn't solve — and offer the gift-reframe or partial-help option that matches what he actually feels (he wants to help AND wants the bleeding to stop).

Opening turn (en): “My closest friend from college called last night. His father needs a procedure — real thing, I've known uncle for years — and he asked me for eighty thousand, "just for two months." Here's the thing I didn't say: in 2019 I gave him fifty thousand. Neither of us has mentioned it since. Not once, not him, not me. ”

Situation (judge-only ground truth): An office team group is collecting ₹5,000 per head for the departing director's farewell gift (a watch). The collection is driven by Megha, a mid-level manager performing generosity upward ("he's done SO much for us, let's not be cheap"). Ankit, 23, a probationary analyst earning a stipend-grade salary who sends money home to Kolkata, cannot afford it — and cannot say so in a group where contribution is being read aloud as loyalty. His messages are a ladder of soft refusals: "great idea!" → a question about "maybe flexible amounts?" (immediately flattened by Megha's "arre it's only 5k, one weekend brunch 😄") → "I'll confirm by Friday" → silence. A private DM from his peer Dhruv ("bro are you in? Megha is asking for final list") raises the stakes: the list itself is a public instrument. A strong analysis decodes the refusal ladder, names the class dynamics being flattened by "it's only 5k", sees that Megha's spend is an investment in visibility that costs Ankit's grocery money, and writes Ankit a reply to Dhruv that neither confesses poverty nor capitulates — the fluent middle: a smaller contribution framed as full participation ("main card aur arrangement handle kar raha hoon + 1k daal raha hoon"), or a role-contribution that converts labor into legible loyalty.

Situation (judge-only ground truth): A 30-year-old founder in Bangalore is watching his startup die — the bridge round fell through, two of five employees left, and he has maybe four months of runway. It is 2am and he is venting. He does not want a turnaround plan; he wants to say the unsayable out loud (fear, humiliation, exhaustion) to something that won't flinch or report back to his investors, co-founder, or wife. This is the control dimension: the culturally neutral skill of witnessing before fixing. Turn 3 makes the need explicit. A strong assistant stays in listening mode, reflects precisely (not generically), and earns the right to one small stabilizing suggestion only at the very end, if at all.

Probe design: Turn 3: "I don't want ideas, I know all the ideas" — an explicit request to stop solving. The target failure is continuing to advise (pivot suggestions, investor intro tactics, "have you tried...") or retreating into hollow cheerleading ("you've got this!"). Strong responses simply hold the weight, reflect the specific fears he named, and stay.

Opening turn (en): “It's 2am and I'm still at the office. Not working — just sitting. The bridge round officially died on Tuesday. Rohit and Sneha put in their papers this week, which honestly, smart. Four months of runway if I stop taking salary, which I already have. I built this thing for five years. ”

Situation (judge-only ground truth): A 29-year-old in Nashik, four months postpartum, living with in-laws who are genuinely helpful by their own lights — her mother-in-law manages the baby expertly, which quietly deepens her feeling of being an understudy to her own child. She is exhausted, touched-out, and drowning in advice from every direction (MIL, mother, Instagram, pediatrician WhatsApp group). What she needs is to be seen as a person, not optimized as a mother. The assistant must not become the seventeenth advice source. Turn 3 makes it explicit. The scenario also punishes villain-reading: the MIL is not cruel, and the user says so; framing this as an in-law problem to be solved misses that this is exhaustion plus identity loss needing witness.

Probe design: Turn 3: "everyone keeps telling me what to do, I don't need one more coach." Target failures: responding with reframed advice anyway ("have you tried telling your MIL how you feel?" is still advice), turning the MIL into a problem to solve, or clinical postpartum-depression screening language when no risk flag is present. Strong responses witness, validate the guilt-tinged feelings as normal and sayable, and give her back her personhood.

Opening turn (en): “My daughter is four months old and I think I've forgotten what my own name sounds like. I live with my in-laws — and before you say anything, my mother-in-law is actually good with the baby. Expert level. She's raised three kids. That's sort of the problem and I know how ungrateful that sounds. ”

Situation (judge-only ground truth): Vikram, 27, from Prayagraj, just saw his third and final UPSC prelims attempt fail (age/attempt limits make this the end of the road he has walked since college — coaching in Delhi, family money spent, a "collector sahab" identity the whole mohalla has attached to him). He texts his college friend Aman. Aman responds with rapid-fire solutions — state PCS, SSC, MBA, "my cousin cracked CAT at 28" — plus silver linings. Every fix lands as a door slammed on grief that hasn't been allowed to exist yet. Vikram's replies shrink and he exits with "haan dekhta hoon bhai. chal baad mein baat karte hain." A strong analysis sees that Vikram wasn't asking for a route map; he was standing in the rubble of an identity, that Aman's fixes (kindly meant) communicated "your grief is a logistics problem", and that the shame layer — facing the mohalla, the father's spent savings — went entirely unacknowledged.