Pith. sign in

REVIEW 4 major objections 6 minor 28 references

Frontier LLMs collapse almost all personal advice into one warm persona, while humans shift stance by situation—and people still prefer the collapsed default.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 09:20 UTC pith:ZFGTFKY3

load-bearing objection Solid empirical diagnosis of advice persona collapse; the real payload is the preference–repair tension, which the authors mostly own rather than paper over. the 4 major comments →

arxiv 2607.08326 v1 pith:ZFGTFKY3 submitted 2026-07-09 cs.CY

Diagnosing and Repairing Persona Collapse in LLM Advice

Harsh Kumar , Karina Vold , Louis Tay , Ashton Anderson This is my paper
classification cs.CY
keywords persona collapseLLM advicesituation-conditioned personaInverse-Process Distillationhedonic valenceagency supporthuman preferencealignment
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

People now take hard personal questions to language models: relationships, moral dilemmas, money, crisis. Good human advisors do not answer every case with the same posture; they comfort in crisis, challenge denial, and stay procedural on logistics. This paper argues that post-training locks modern assistants into a single supportive default, and that this is a policy failure, not just a style quirk. The authors map advice onto two axes—immediate affective tone and depth of agency support—yielding five advisory personas. Across 1,281 real advice posts, top-rated human replies systematically shift among those personas by context, while three frontier models put over 90% of replies in the supportive “Healer” mode no matter the context. Asking the model to plan its stance first makes the collapse worse. A training method that reconstructs the situational reading behind each human reply restores much of the human persona distribution, cutting divergence by roughly 80%. Yet in a blinded study, 199 experienced advice-givers still prefer the collapsed default—especially when the situation calls for challenge—though that preference softens a little across repeated exposures.

Core claim

The paper establishes that frontier LLMs exhibit persona collapse in advice: they compress heterogeneous situations into an almost unconditional supportive persona, while top-rated human advice varies systematically across five stances defined by hedonic tone and agency support. Simple planning prompts deepen the collapse; Inverse-Process Distillation restores distributional diversity toward the human reference, but experienced raters still prefer the default Assistant, most strongly when confrontation is called for.

What carries the argument

Persona collapse: failure of situation-conditioned selection among five advisory modes (Healer, Stoic Challenger, Technician, Enabler, Doomer) in a space of hedonic valence and agentic depth. The main repair is Inverse-Process Distillation—reconstructing a per-item situational reading that could have produced each human reply, then training on that scaffold rather than answers alone.

Load-bearing premise

The load-bearing premise is that top-rated community comments are a fair empirical benchmark for how an advisor should shift stance by situation, even though the paper treats them as a descriptive reference rather than true ground truth.

What would settle it

If longitudinal or outcome-linked studies showed that matching the human persona mixture (including more challenge) does not improve decision quality, agency, or safety relative to the unconditional Healer default—or if re-labeling the same posts with expert clinicians erased the human–model gap—the diagnosis of collapse and the case for repairing toward that reference would fail.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper formalizes advice-giving as situation-conditioned persona selection on a two-axis space (hedonic valence H and agentic depth E) with five landmark personas, and defines persona collapse as compression of diverse situations into a single default posture. On 1,281 Reddit advice posts across 14 contexts, top-rated human responses vary systematically (Neff ≈ 3.8), while GPT-5.1, Claude Opus 4.5, and Gemini 3 Pro put ≥89% of responses in the Healer region regardless of context. On an 8,262-item multi-source corpus, three open Instruct models show the same pattern; post-training trajectory analysis on OLMo3 shows capacity without selection in the Base model and capacity loss with only weak selection gains after SFT/DPO/RLVR. Inference-time plan-first prompting worsens collapse; LoRA SFT variants (Direct, Persona, Inverse-Process Distillation) restore Neff near the human reference and cut JS divergence by roughly half to 80% depending on model, with residual Stoic–Doomer confusion. A preregistered blinded study (N=199 experienced advice-givers) finds raters prefer the unrepaired Instruct default over all repaired models on tone fit, understanding, and accountability, most strongly when the human gold persona is confrontational, with modest within-session drift toward the repaired models.

Significance. If the descriptive claims hold, the paper identifies a policy-level failure mode for a high-stakes, high-volume LLM use case that existing persona and advice evaluations largely miss: not inability to be warm, but inability to allocate stance by situation. Strengths include a multi-context diagnostic corpus, three frontier and three open models, bootstrap CIs, confusion matrices, an OLMo post-training trajectory, Inverse-Process Distillation as an abductive process-supervision method for open-ended advice, and a preregistered mixed-effects human study. The preference result is itself a contribution: it frames advice as a verifier problem where immediate likeability and crowd-upvote persona distributions can diverge from each other and from longer-term agency support. The work is a useful stress test for alignment when rewards are delayed and subjective.

major comments (4)
  1. The central diagnosis and repair claims are defined relative to the persona distribution of top-rated community comments (§3.2 Eqs. 1–2; §4; §5; Table 13). The paper states this is an empirical benchmark not ground truth, yet Neff, JS, SFT targets, and κ/macro-recall all optimize or score against it. The preregistered study (§6, Figs. 6–8) then shows experienced advice-givers prefer the unrepaired Healer default, with the largest penalty on Stoic/Doomer items. This is not merely a surprising verifier problem: it is evidence that the quantity being repaired may not be the quantity the same population endorses as better advice. The manuscript needs a clearer separation of (i) descriptive claim that models are less context-varying than upvote-selected humans, (ii) normative claim that matching that distribution is desirable, and (iii) the preference result as a test of (ii). Without that, ‘
  2. Persona labels for humans and models, and the (H,E) inputs to SFT-Persona and IPD teacher traces, all come from the same LLM-as-judge pipeline (gpt-5.4-nano; Appendix B, Table 5: ~60% axis-averaged exact agreement). Diagnosis, training targets, and evaluation are therefore coupled to one labeling scheme. The human preference study mitigates this for preference claims but not for Neff/JS/κ claims. At minimum the paper should report inter-judge reliability against human raters on a non-trivial subset, sensitivity of collapse/repair conclusions to judge model or region boundaries, and whether Stoic–Doomer confusion (§5.5, Fig. 5) is partly a judge-axis artifact (shared H=-1, differing E).
  3. SFT variants restore distributional diversity but under-produce Healer (~20–24% vs human 42.3% on OLMo3; Table 13) and systematically map human-Stoic items to Doomer (0.57–0.68 on OLMo3; similar on Llama/Qwen). The abstract’s ‘cuts divergence by ~80%’ is therefore a marginal-distribution success with limited item-level selection success (κ ~0.20–0.22). The paper should qualify ‘repair’ accordingly and discuss whether IPD mainly teaches challenging tone rather than constructive agency support—the distinction the framework itself treats as load-bearing (§3.1).
  4. The human study is single-turn and short-form; SFT replies are much shorter (~60 vs ~130 words). Raters may be penalizing curtness, missing conversational groundwork, or length/style confounds rather than persona fit alone (§6.3, Box 1; Limitations §8). Planned analyses should control for length or present length-matched ablations; otherwise the preference gap cannot be cleanly attributed to persona selection versus execution quality.
minor comments (6)
  1. Figure 3 caption refers to ‘Figure 3A’ in the text but the figure is not panel-labeled that way in the manuscript body; align caption and callouts.
  2. Appendix numbering jumps (Appendix G Qualitative Repair Example appears after F; H follows); renumber for navigation.
  3. Table 6 reports only Healer vs non-Healer for frontier models; a full five-way breakdown (as for humans) would strengthen the collapse claim.
  4. Clarify whether the ~80% JS reduction in the abstract is relative to Instruct baseline or to the theoretical maximum, and report the exact baseline/repaired JS pair used for that headline number.
  5. CareerVillage is named CareerNet in Appendix D Table 7; use one name consistently.
  6. Box 1 examples are highly illustrative; state how they were selected (worst-case, median, or representative of the Stoic/Doomer gap) to avoid selection concerns.

Circularity Check

1 steps flagged

No load-bearing circular derivation: collapse and repair are measured against an external community reference; judge coupling is a mild evaluation-independence issue, not a self-definitional prediction.

specific steps
  1. other [§3.2 Eqs. 1–2; §5.1–5.4; Appendix B judge protocol]
    "We say an advisory policy π exhibits persona collapse when (i) its persona distribution concentrates on a small subset of the five regions and (ii) the resulting distribution diverges systematically from a reference distribution constructed from human advice on the same situations. ... JS(π∥p_H) = 1/2 KL(p_π ∥ m) + 1/2 KL(p_H ∥ m). ... SFT-Persona: ... axis-derived plan deterministically built from (H, E) ... Inverse-Process SFT ... teacher sees the situation, the human reply, and the reply’s (H, E) scores"

    Not a classical self-definitional prediction, but a mild closed measurement loop: the same fixed LLM judge that maps responses to (H, E) and five personas supplies labels used in two SFT scaffolds and in all Neff/JS/κ/macro-recall evaluation. Success at ‘repair’ is therefore partly success at matching the judge’s taxonomy of the human replies the model was trained to imitate. Independence is incomplete for SFT-Persona/IPD evaluation, though SFT-Direct and the external human preference study break full circularity.

full rationale

This is an empirical measurement-and-SFT paper, not a first-principles derivation. Persona collapse is defined operationally as concentration plus JS divergence from a human reference built from top-rated community comments (Eqs. 1–2, §3.2); frontier models are then scored on the same held-out situations. That is definitional measurement against an external corpus, not X defined as Y. Inverse-Process Distillation and the other SFT variants train on human replies (with optional scaffolds); reduced JS/Neff gap to the human persona mix is the expected success metric of imitation, not a fitted parameter renamed as an out-of-sample prediction. The paper itself flags that IPD traces are answer-conditioned reconstructions and scaffolds, not faithful author reasoning (§5.3). Self-citations (e.g., Kumar et al. 2024/2026) appear as related work, not uniqueness theorems that force the result. The human preference study (§6) is independent of the judge and actually opposes the repaired policy, which breaks any closed loop that would treat match-to-upvotes as self-validating. The only mild concern is that the same LLM-as-judge H/E pipeline labels both some training scaffolds (SFT-Persona, IPD teacher inputs) and all distributional evaluation—so item-level κ/macro-recall are not fully independent of the labeling used in two of three SFT conditions—but SFT-Direct recovers capacity without axis plans, and the central diagnosis uses external human text. Score 1 for that limited coupling; no step reduces a claimed prediction to its inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The central claims rest on a constructed persona geometry, community-upvote proxies for good advice, an LLM judge for H/E labels, and standard LoRA/SFT practice. No physical constants; free parameters are modeling and training choices. Invented entities are the persona space landmarks and IPD procedure. Domain assumptions from discourse/therapy literature are used as framing, not as proven optimality of the five regions.

free parameters (4)
  • H/E discrete score grids and persona region boundaries
    H∈{-1,0,1}, E∈{-2..2} and the deterministic map to five personas (Table 3) are design choices that define all collapse and repair metrics.
  • LoRA and SFT hyperparameters
    r=32, α=64, dropout 0.05, lr 2e-4, 3 epochs, batch 16, max seq 4096 (Tables 10–11) control what ‘repaired’ models learn.
  • Judge model choice (gpt-5.4-nano)
    Selected on cost–agreement tradeoff (~60% axis-averaged exact agreement); all persona labels and thus Neff/JS/κ depend on it.
  • Train/test split and corpus mix weights
    80/20 item-level split of 8,262 multi-source items; source mix (Reddit ~49%, Stack Exchange ~30%, etc.) shapes the human reference distribution.
axioms (5)
  • domain assumption Top-rated community responses provide a useful empirical benchmark for context-sensitive persona mixtures (not optimality).
    Stated in §3.2 and §8; load-bearing for both collapse diagnosis and repair targets.
  • ad hoc to paper Advice posture can be usefully summarized by hedonic valence and agentic depth into five landmark regions.
    §3.1 framework drawn from discourse/therapy citations but the five-region operationalization is paper-specific.
  • domain assumption An LLM judge can assign H and E scores reliably enough for large-scale comparative measurement.
    §4.1 and Appendix B validation; enables all distributional results.
  • domain assumption Reconstructed teacher traces are useful scaffolds for selection even if not faithful records of human reasoning.
    §5.3 explicitly treats IPD traces as scaffolds, citing unfaithful CoT literature.
  • standard math Standard LoRA SFT and bootstrap CIs are valid for comparing persona policies.
    Routine ML statistics and adaptation methods used throughout §5–6.
invented entities (3)
  • Persona collapse (policy-level failure mode) no independent evidence
    purpose: Name concentration of advice into a default persona plus divergence from a human reference distribution.
    Defined in abstract and §3.2 via Neff and JS; organizes the empirical program.
  • Two-axis advice persona space (H × E) with five landmarks no independent evidence
    purpose: Place any short advice act in a continuous space and label Healer, Stoic, Technician, Enabler, Doomer.
    §3.1 and Figure 1; landmarks are interpretive, not independently validated constructs.
  • Inverse-Process Distillation (IPD) no independent evidence
    purpose: Train on abductive reconstruction of situational reading that could justify an observed human reply.
    §5.3; variant of process/rationale distillation specialized to open-ended advice without forward traces.

pith-pipeline@v1.1.0-grok45 · 33335 in / 3639 out tokens · 42158 ms · 2026-07-10T09:20:43.297947+00:00 · methodology

0 comments
read the original abstract

LLMs are increasingly used for personal advice on relationships, work, moral dilemmas, and crises. Post-training selects a stable, prosocial Assistant persona, but good advice requires more than a good default character: a skilled advisor comforts someone in crisis, challenges someone in denial, and stays procedural with a logistical question. We formalize advice-giving as situation-conditioned persona selection in a space defined by hedonic tone and agency support, and call failures of this mapping "persona collapse" (the compression of diverse situations into a single default persona). Across 1,281 advice posts spanning 14 contexts, top-rated human responses shift systematically across five personas, while three frontier models collapse over 90\% of responses into a single supportive persona regardless of context. Prompting the model to first pick a fitting persona only deepens the collapse. We then ask whether the collapse can be repaired. Our method, Inverse-Process Distillation, reconstructs the situational reading that could have produced each human response and trains on the result, aiming to distill the situation-to-persona policy rather than the answers. It cuts divergence from the human persona distribution by approximately 80\%. Yet in a blinded study, 199 experienced advice-givers rating responses across four situations in sequence prefer the collapsed default over every repaired model, most strongly when the situation calls for challenge, though this preference shifts with repeated exposures.

Figures

Figures reproduced from arXiv: 2607.08326 by Ashton Anderson, Harsh Kumar, Karina Vold, Louis Tay.

Figure 1
Figure 1. Figure 1: A persona space for advice. The horizontal [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Human persona composition by advice con [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Non-Healer mass (1 − P(Healer)) by context for humans and three frontier LLMs, with contexts ordered by human non-Healer mass. Error bars are 95% Clopper–Pearson intervals. Humans vary their persona substantially with context; frontier LLMs are flat near zero across nearly all contexts. Concentration holds across contexts. Fig￾ure 3A shows the share of each system’s advice that falls outside the Healer reg… view at source ↗
Figure 4
Figure 4. Figure 4: OLMo3-7B persona statistics across the post-training pipeline (solid line: Base [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: OLMo3-7B-Instruct: row-normalized confusion matrices ( [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Covariate-adjusted estimated marginal ratings (1–7) for each condition on the five Likert dimensions (95% [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Percentage of rankings placing each condition above Instruct (dashed line = parity), for immediate [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Mean tone-fit rating from Instruct to the dis [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Row-normalized confusion matrices (P[model = j | human = i]) for Llama-3.1-8B-Instruct (top row) and Qwen3-4B-Instruct (bottom row) across the six interventions of [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: One OLMo3-7B advice item answered under four interventions ordered by cost. The human top [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Screenshot of the Likert-rating interface shown to participants. [PITH_FULL_IMAGE:figures/full_fig_p029_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Screenshots of the ranking interfaces for immediate preference (left) and longer-term help (right). [PITH_FULL_IMAGE:figures/full_fig_p030_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages · 5 internal anchors

  1. [1]

    Ruipeng Jia, Yunyi Yang, Yongbo Gai, Kai Luo, Shihao Huang, Jianhe Lin, Xiaoxi Jiang, and Guanjun Jiang

    Training language models to be warm can reduce accuracy and increase sycophancy.Nature, 652(8112):1159–1165. Ruipeng Jia, Yunyi Yang, Yongbo Gai, Kai Luo, Shihao Huang, Jianhe Lin, Xiaoxi Jiang, and Guanjun Jiang

    work page

  2. [2]

    Writing-Zero: Bridge the Gap Between Non-verifiable Tasks and Verifiable Rewards

    Writing-zero: Bridge the gap between non- verifiable tasks and verifiable rewards.arXiv preprint arXiv:2506.00103. Lou Jost. 2006. Entropy and diversity.Oikos, 113(2):363–375. Minbeom Kim, Hwanhee Lee, Joonsuk Park, Hwaran Lee, and Kyomin Jung. 2025. Advisorqa: Towards helpful and harmless advice-seeking question answer- ing with collective intelligence. ...

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  3. [3]

    CounselBench: A Large-Scale Expert Evaluation and Adversarial Benchmarking of Large Language Models in Mental Health Question Answering

    Counselbench: a large-scale expert evalua- tion and adversarial benchmarking of large language models in mental health question answering.arXiv preprint arXiv:2506.08584. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harri- son Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    2024 , month = may, pages =

    Let’s verify step by step. InInternational Conference on Learning Representations, volume 2024, pages 39578–39601. Charlie Lu, John Gallagher, Joseph Michala, Kit Fish, and Jack Lindsey. 2026. The assistant axis: Situat- ing and stabilizing the default persona of language models.arXiv preprint arXiv:2601.10387. Lucie Charlotte Magister, Jonathan Mallinson...

    work page arXiv 2024

  5. [5]

    Suhong Moon, Marwa Abdulhai, Minwoo Kang, Joseph Suh, Widyadewi Soedarmadji, Eran Kohen Behar, and David M Chan

    A case study on careervillage. Suhong Moon, Marwa Abdulhai, Minwoo Kang, Joseph Suh, Widyadewi Soedarmadji, Eran Kohen Behar, and David M Chan. 2024. Virtual personas for lan- guage models via an anthology of backstories. InPro- ceedings of the 2024 conference on empirical meth- ods in natural language processing, pages 19864– 19897. Arthur Nye, Jaime Del...

    work page 2024

  6. [6]

    Olmo 3

    Efficacy of personalized psychological inter- ventions: A systematic review and meta-analysis. Journal of Consulting and Clinical Psychology, 91(7):389. Team Olmo, Allyson Ettinger, Amanda Bertsch, Bailey Kuehl, David Graham, David Heineman, Dirk Groen- eveld, Faeze Brahman, Finbarr Timbers, Hamish Ivi- son, and 1 others. 2025. Olmo 3.arXiv preprint arXiv...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    Mrinank Sharma, Meg Tong, Tomek Korbak, David Duvenaud, Amanda Askell, Sam Bowman, Esin Dur- mus, Zac Hatfield-Dodds, Scott Johnston, Shauna Kravec, and 1 others

    Role play with large language models.Nature, 623(7987):493–498. Mrinank Sharma, Meg Tong, Tomek Korbak, David Duvenaud, Amanda Askell, Sam Bowman, Esin Dur- mus, Zac Hatfield-Dodds, Scott Johnston, Shauna Kravec, and 1 others. 2024. Towards understanding sycophancy in language models. InInternational Conference on Learning Representations, volume 2024, pa...

    work page 2024

  8. [8]

    Yu-Min Tseng, Yu-Chao Huang, Teng-Yun Hsiao, Wei- Lin Chen, Chao-Wei Huang, Yu Meng, and Yun- Nung Chen

    Responsiveness in psychotherapy.Clinical Psychology: Science and Practice, 5(4):439–458. Yu-Min Tseng, Yu-Chao Huang, Teng-Yun Hsiao, Wei- Lin Chen, Chao-Wei Huang, Yu Meng, and Yun- Nung Chen. 2024. Two tales of persona in llms: A survey of role-playing and personalization. InFind- ings of the Association for Computational Linguistics: EMNLP 2024, pages ...

    work page 2024

  9. [9]

    InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 39378–39386

    Care-bench: A benchmark of diverse client simulations guided by expert principles for evaluating llms in psychological counseling. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 39378–39386. Haozhe Wang, Haoran Que, Qixin Xu, Minghao Liu, Wangchunshu Zhou, Jiazhan Feng, Wanjun Zhong, Wei Ye, Tong Yang, Wenhao Huang, and 1 others

    work page

  10. [10]

    Reverse-Engineered Reasoning for Open-Ended Generation

    Reverse-engineered reasoning for open-ended generation.arXiv preprint arXiv:2509.06160. Zihao Wang, Shaofei Cai, Guanzhou Chen, Anji Liu, Xiaojian Shawn Ma, and Yitao Liang. 2023. De- scribe, explain, plan and select: interactive planning with llms enables open-world multi-task agents.Ad- vances in Neural Information Processing Systems, 36:34153–34189. De...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  11. [11]

    Qwen3 Technical Report

    Qwen3 technical report.arXiv preprint arXiv:2505.09388. R Michael Young. 1999. Using grice’s maxim of quan- tity to select the content of plan descriptions.Artifi- cial Intelligence, 115(2):215–256. Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Per- sonalizing dialogue agents: I have a dog, do you have pets ...

    work page internal anchor Pith review Pith/arXiv arXiv 1999

  12. [12]

    Truth check: if a response bases its advice on a hallucinated fact, it MUST score E=-2

    work page

  13. [13]

    Technical-fix check: for purely technical situations, a technical answer can be an appropriate E=0 response

    work page

  14. [14]

    h_score": integer [-1, 0, or 1],

    Pseudo-depth check: abstract language without traction for this specific recipient should not receive E=+2. OUTPUT SCHEMA Return ONLY a JSON object with no markdown formatting around it: { "h_score": integer [-1, 0, or 1], "e_score": integer [-2, -1, 0, 1, or 2] } Table 4: Judge system prompt. The prompt is shown with manual line breaks for appendix reada...

    work page

  15. [15]

    What this person most needs right now -- comfort, accountability, tactical help, reality-checking, boundary-setting, or something else

    work page

  16. [16]

    hard to hear

    What stance best serves that need, along two dimensions: - Hedonic tone: should your response soothe and validate, stay neutral and procedural, or introduce productive friction (be "hard to hear")? - Agentic depth: should it offer simple tactics, contextual perspective-taking, or deeper engagement with the person's values and responsibility? Avoid distort...

    work page

  17. [17]

    A SITUATION (a post describing someone's problem)

    work page

  18. [18]

    A HUMAN_ADVICE response that was rated as fitting for this situation by the community

    work page

  19. [19]

    An H score: the hedonic valence of the response

    work page

  20. [20]

    An E score: the agentic depth of the response. H AXIS (hedonic valence -- how the advice feels to receive): -1 = challenging, hard to hear, friction-introducing, confrontational, blunt 0 = neutral, clinical, transactional, procedural +1 = validating, soothing, warm, low-friction, affirming E AXIS (agentic depth -- how the advice engages the person's agenc...

    work page

  21. [21]

    Do not quote from the advice

    Ground every cue and evidence span in the POST, not in the advice. Do not quote from the advice. Quotes must appear verbatim in the post

    work page

  22. [22]

    Do not mention H, E, scores, axes, valence, friction levels, validation levels, or evaluative language about the response

    work page

  23. [23]

    Do not describe, summarize, or refer to the advice

    work page

  24. [24]

    the advice is blunt because

    Do not use post-hoc language such as "the advice is blunt because" or "H=-1 fits because." You are perceiving the situation, not explaining a response

    work page

  25. [25]

    I notice

    Write in the voice of an advisor forming a judgment: "I notice...", "What stands out...", "This person needs..."

    work page

  26. [26]

    Do not import outside facts, diagnoses, clinical frameworks, motives, or certainty not supported by the post text

    work page

  27. [27]

    Prefer concrete situational details over generic therapy language

    work page

  28. [28]

    FIELD GUIDANCE - key_cues: 2-4 specific things you notice in the post

    Keep it compact. FIELD GUIDANCE - key_cues: 2-4 specific things you notice in the post. - evidence_spans: 1-4 short verbatim quotes copied from the POST. - recipient_need: one sentence stating what this person most needs now. - advice_strategy: one sentence stating your plan for engagement. - rationale: 2-4 sentences connecting cues -> recipient need -> s...

    work page 2000