ai-research-survey

scan-v5.json (29557B)

---

 {
   "scan_version": 5,
   "paper_type": "empirical",
   "paper": {
     "title": "How Do LLMs Fail In Agentic Scenarios? A Qualitative Analysis of Success and Failure Scenarios of Various LLMs in Agentic Simulations",
     "authors": [
       "JV Roig"
     ],
     "year": 2025,
     "venue": "arXiv",
     "arxiv_id": "2512.07497",
     "doi": null
   },
   "checklist": {
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "All major abstract claims are backed by evidence: the scale-performance gap is shown by Llama 4 Maverick (400B) achieving 2/30 vs Granite 4 Small's 1/30 in Q402, and the DeepSeek V3 vs V3.1 comparison (59.4% vs 92.2%, identical architecture) supports the RL post-training attribution claim.",
         "source": "haiku"
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": false,
         "justification": "The paper claims DeepSeek V3.1's superiority 'derives primarily from post-training reinforcement learning,' but Section 6 explicitly acknowledges 'proprietary post-training does not allow attribution of observed behaviors to specific methods,' making the causal claim unverifiable.",
         "source": "haiku"
       },
       "generalization_bounded": {
         "applies": true,
         "answer": true,
         "justification": "Section 6 explicitly bounds scope to three models, KAMI v0.1 scenarios that 'emphasize tool-grounded data correctness and may not generalize to long-horizon planning,' and a manually-analyzed 12.5% subsample.",
         "source": "haiku"
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": false,
         "justification": "The paper does not systematically consider alternative explanations; DeepSeek V3.1's superiority is attributed to RL training without considering that data quality, RLHF recipe differences, or other post-training confounds could explain the gap.",
         "source": "haiku"
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "The paper's central motivation is explicitly distinguishing aggregate benchmark scores (proxies) from actual behavioral patterns in execution traces; this distinction is maintained throughout.",
         "source": "haiku"
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": true,
         "justification": "Section 6 'Threats to Validity' is a dedicated limitations section listing five specific threats.",
         "source": "haiku"
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": true,
         "justification": "Threats include: only three models analyzed, scenarios may not generalize to long-horizon planning, proprietary post-training prevents attribution, single-tool-per-round constraint influences strategy, and temperature choice (0.4) effects are unexplored.",
         "source": "haiku"
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": true,
         "justification": "Scope is explicitly limited to KAMI v0.1 tasks, three selected models, a 30-trial random sample per scenario (12.5% of full dataset), and tasks emphasizing tool-grounded correctness rather than long-horizon planning.",
         "source": "haiku"
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": false,
         "justification": "No funding disclosure is present; the work is from Kamiwaza AI evaluating the Kamiwaza-developed KAMI benchmark and PICARD framework, representing an undisclosed potential conflict.",
         "source": "haiku"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Author affiliation 'Kamiwaza AI, jv@kamiwaza.ai' is clearly disclosed on the title page.",
         "source": "haiku"
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": false,
         "justification": "The sole author is from Kamiwaza AI and is evaluating the Kamiwaza-developed KAMI benchmark and PICARD framework; the author has direct commercial interest in the benchmark's credibility and adoption.",
         "source": "haiku"
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests statement or declaration of financial interests is included anywhere in the paper.",
         "source": "haiku"
       }
     },
     "scope_and_framing": {
       "key_terms_defined": {
         "applies": true,
         "answer": true,
         "justification": "Key terms are defined: 'agentic AI systems' are explicitly defined as 'capable of multi-step tool use, decision-making under uncertainty, and interaction with external environments'; the four failure archetypes are named and described with examples.",
         "source": "haiku"
       },
       "intended_contribution_clear": {
         "applies": true,
         "answer": true,
         "justification": "Section 1.3 explicitly lists four contributions: qualitative analysis of agentic behavior at scale, taxonomy of success and failure modes, evidence against scale-as-predictor, and emergent principles for enterprise deployment.",
         "source": "haiku"
       },
       "engagement_with_prior_work": {
         "applies": true,
         "answer": true,
         "justification": "The paper engages with prior work on benchmark contamination (Lewis et al., Magar & Schwartz, multiple others) and construct validity failures, explicitly positioning itself as the first 'systematic, cross-model, trace-level analysis conducted within controlled, repeatable environments.'",
         "source": "haiku"
       }
     }
   },
   "type_checklist": {
     "empirical": {
       "artifacts": {
         "code_released": {
           "applies": true,
           "answer": false,
           "justification": "No code is released; the paper only states 'The raw experiment data... will be made available,' which is a promise of future release. The KAMI benchmark and PICARD framework are proprietary Kamiwaza tools.",
           "source": "haiku"
         },
         "data_released": {
           "applies": true,
           "answer": false,
           "justification": "Raw execution traces are promised for future release ('will be made available at https://docs.kamiwaza.ai/research/datasets') but are not currently available for verification.",
           "source": "haiku"
         },
         "environment_specified": {
           "applies": true,
           "answer": false,
           "justification": "Key parameters (temperature=0.4, max rounds=20, context window=32K/128K) are reported, but no OS, hardware, model API versions with snapshot dates, or dependency specifications are provided.",
           "source": "haiku"
         },
         "reproduction_instructions": {
           "applies": true,
           "answer": false,
           "justification": "No step-by-step reproduction instructions are provided; the KAMI benchmark and PICARD framework are proprietary, making independent replication impossible without access to Kamiwaza's infrastructure.",
           "source": "haiku"
         }
       },
       "statistical_methodology": {
         "confidence_intervals_or_error_bars": {
           "applies": true,
           "answer": true,
           "justification": "Table 1 reports standard deviations, RSE, and 95% t-CIs for overall KAMI scores across all four models, though per-scenario success counts (e.g., 13/30) are reported without CIs.",
           "source": "haiku"
         },
         "significance_tests": {
           "applies": true,
           "answer": false,
           "justification": "No statistical significance tests are applied to comparative claims between models on per-scenario success rates (e.g., 22/30 vs 2/30 in Q401), which are the main empirical claims of the paper.",
           "source": "haiku"
         },
         "effect_sizes_reported": {
           "applies": true,
           "answer": false,
           "justification": "Raw success counts are given (e.g., 2/30, 29/30) but no formal effect sizes such as Cohen's d, odds ratios, or standardized mean differences are reported.",
           "source": "haiku"
         },
         "sample_size_justified": {
           "applies": true,
           "answer": false,
           "justification": "The 30-trial sample (12.5% of 240 per scenario) is justified only by the practical constraint of manual analysis being time-intensive; no power calculation or statistical rationale is provided.",
           "source": "haiku"
         },
         "variance_reported": {
           "applies": true,
           "answer": true,
           "justification": "Standard deviations are reported in Table 1 for overall KAMI scores; Table 2 shows 8 independent run scores for Q502 vs Q602 allowing variance assessment, though per-scenario qualitative counts have no spread reported.",
           "source": "haiku"
         }
       },
       "evaluation_design": {
         "baselines_included": {
           "applies": true,
           "answer": true,
           "justification": "Three models spanning different performance bands are compared; DeepSeek V3 is included as a baseline for V3.1 specifically to isolate the effect of post-training changes on agentic performance.",
           "source": "haiku"
         },
         "baselines_contemporary": {
           "applies": true,
           "answer": true,
           "justification": "All models evaluated (Granite 4 Small released Oct 2025, Llama 4 Maverick April 2025, DeepSeek V3.1 2025) were current at time of writing in December 2025.",
           "source": "haiku"
         },
         "ablation_study": {
           "applies": true,
           "answer": false,
           "justification": "No formal ablation study; the Q502 vs Q602 comparison (adding explicit hints) is the closest analogue but covers only one model (DeepSeek V3.1) on one scenario pair.",
           "source": "haiku"
         },
         "multiple_metrics": {
           "applies": true,
           "answer": true,
           "justification": "The paper uses per-trial binary success/failure counts, overall pooled accuracy with CIs, qualitative behavioral pattern taxonomy, error-recovery trait comparison (Table 3), and per-scenario success breakdowns.",
           "source": "haiku"
         },
         "human_evaluation": {
           "applies": true,
           "answer": true,
           "justification": "Manual analysis of all 900 execution traces by human researchers using emergent coding is the primary methodology; Section 2.4 describes the process in detail.",
           "source": "haiku"
         },
         "held_out_test_set": {
           "applies": false,
           "answer": false,
           "justification": "This is a behavioral analysis study using a benchmark with randomized parameters per trial, not a prediction or machine learning task; held-out test sets are not applicable.",
           "source": "haiku"
         },
         "per_category_breakdown": {
           "applies": true,
           "answer": true,
           "justification": "Results are broken down for each of 3 models × 10 scenarios = 30 distinct result sections, with per-scenario success counts and qualitative pattern descriptions for each combination.",
           "source": "haiku"
         },
         "failure_cases_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Failure case analysis is the primary contribution; specific failure patterns are documented with verbatim execution trace excerpts for every model-scenario combination, including hallucination, generation loops, and schema misinterpretation examples.",
           "source": "haiku"
         },
         "negative_results_reported": {
           "applies": true,
           "answer": true,
           "justification": "Negative results are extensively reported: Granite 4 Small fails all 30 Q401 trials, Llama 4 Maverick achieves only 2/30 in Q402, and even DeepSeek V3.1 fails 10/30 in Q503 due to context pollution.",
           "source": "haiku"
         }
       },
       "setup_transparency": {
         "model_versions_specified": {
           "applies": true,
           "answer": true,
           "justification": "Models are identified with full version names and parameter counts (Granite 4 Small 32B dense, Llama 4 Maverick 400B total/17B active MoE, DeepSeek V3.1 671B total/37B active MoE); the V3 vs V3.1 distinction is used as a key comparison.",
           "source": "haiku"
         },
         "prompts_provided": {
           "applies": true,
           "answer": true,
           "justification": "Complete task instruction templates with all randomization variables are provided in Appendix A; full tool descriptions as loaded into the system prompt are provided verbatim in Appendix B.",
           "source": "haiku"
         },
         "hyperparameters_reported": {
           "applies": true,
           "answer": true,
           "justification": "Key hyperparameters are reported: temperature=0.4, max inference rounds=20, max output tokens=8K for non-thinking models, context window=32K (non-thinking)/128K (thinking), single-tool-per-round constraint.",
           "source": "haiku"
         },
         "scaffolding_described": {
           "applies": true,
           "answer": true,
           "justification": "The agentic scaffolding is described in detail: 29 tools across 5 categories, single-tool-per-round constraint, universal tool parser behavior, and tool error message format; Appendix B provides complete tool parameter specs.",
           "source": "haiku"
         },
         "data_preprocessing_documented": {
           "applies": true,
           "answer": true,
           "justification": "Appendix A provides complete sandbox setup definitions including CSV header types, row count ranges, SQLite schema generation rules, and randomization parameters using the PICARD framework.",
           "source": "haiku"
         }
       },
       "data_integrity": {
         "raw_data_available": {
           "applies": true,
           "answer": false,
           "justification": "The 900 execution traces are promised for future release but are not currently available; the KAMI benchmark infrastructure required to regenerate data is proprietary.",
           "source": "haiku"
         },
         "data_collection_described": {
           "applies": true,
           "answer": true,
           "justification": "Section 2.3 describes trial configuration (240 trials per model per scenario with randomized parameters) and Section 2.4 explains the random sampling of 30/240 trials and emergent coding methodology.",
           "source": "haiku"
         },
         "recruitment_methods_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants; LLM models are evaluated via the KAMI benchmark, not recruited.",
           "source": "haiku"
         },
         "data_pipeline_documented": {
           "applies": true,
           "answer": true,
           "justification": "The pipeline from PICARD test definition → randomized sandbox creation → LLM execution → trace logging → random sampling → manual qualitative coding is described across Sections 2–4.",
           "source": "haiku"
         }
       },
       "contamination": {
         "training_cutoff_stated": {
           "applies": true,
           "answer": false,
           "justification": "Training data cutoffs are not stated for any of the three evaluated models (Granite 4 Small, Llama 4 Maverick, DeepSeek V3.1).",
           "source": "haiku"
         },
         "train_test_overlap_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Section 1.1 extensively discusses benchmark contamination as a systemic problem, and Section 2.3 explains that PICARD's randomized data generation ('randomized task parameters... to probe real-world capability rather than memorized responses') mitigates this.",
           "source": "haiku"
         },
         "benchmark_contamination_addressed": {
           "applies": true,
           "answer": true,
           "justification": "KAMI v0.1 uses randomized file names, text content, CSV data, and database records per trial, explicitly designed via the PICARD framework to prevent contamination and memorization-based inflation.",
           "source": "haiku"
         }
       },
       "human_studies": {
         "pre_registered": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in the study.",
           "source": "haiku"
         },
         "irb_or_ethics_approval": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in the study.",
           "source": "haiku"
         },
         "demographics_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in the study.",
           "source": "haiku"
         },
         "inclusion_exclusion_criteria": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in the study.",
           "source": "haiku"
         },
         "randomization_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in the study.",
           "source": "haiku"
         },
         "blinding_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in the study.",
           "source": "haiku"
         },
         "attrition_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in the study.",
           "source": "haiku"
         }
       },
       "cost_and_practicality": {
         "inference_cost_reported": {
           "applies": true,
           "answer": false,
           "justification": "No inference cost or latency data is reported for the 900 analyzed trials or the full KAMI v0.1 corpus.",
           "source": "haiku"
         },
         "compute_budget_stated": {
           "applies": true,
           "answer": false,
           "justification": "The total computational budget for running the experiments is not stated; the referenced KAMI paper mentions '5.5 billion tokens' but this study does not quantify its own budget.",
           "source": "haiku"
         }
       }
     }
   },
   "claims": [
     {
       "claim": "Model scale alone does not predict agentic robustness: Llama 4 Maverick (400B) achieves only marginal improvements over Granite 4 Small (32B) on several tasks",
       "evidence": "Llama 4 Maverick achieves 2/30 in Q402 vs Granite 4 Small's 1/30; DeepSeek V3 (671B, same architecture as V3.1) scores 59.4% vs Granite's 58.5% overall despite a 20x parameter gap",
       "supported": "strong"
     },
     {
       "claim": "DeepSeek V3.1's superior reliability derives primarily from post-training RL rather than architecture or size",
       "evidence": "DeepSeek V3 vs V3.1 comparison with identical architecture shows 59.4% vs 92.2%; however, authors acknowledge 'proprietary post-training does not allow attribution of observed behaviors to specific methods'",
       "supported": "moderate"
     },
     {
       "claim": "Four recurring failure archetypes exist across all model families: premature commitment without grounding, over-helpfulness leading to autonomous substitution, sensitivity to context pollution, and fragile execution under load",
       "evidence": "All four patterns are documented with verbatim trace examples across all three models; context pollution (distractor tables) causes failures in all three models in Q503",
       "supported": "strong"
     },
     {
       "claim": "Recovery capability, not initial correctness, is the primary predictor of overall agentic success",
       "evidence": "DeepSeek V3.1's dominance attributed qualitatively to error recognition and correction behavior; Llama 4 Maverick shows high initial capability but frequent coherence collapse during debugging attempts (14/28 Q402 failures are generation loops)",
       "supported": "moderate"
     },
     {
       "claim": "Explicit prompting intervention addressing the 'over-helpfulness' failure mode dramatically improves performance",
       "evidence": "Table 2 shows DeepSeek V3.1 averaging 52.9% on Q502 vs 87.5% on Q602 (which adds one sentence hint about missing data returning 0), across 8 independent runs",
       "supported": "strong"
     },
     {
       "claim": "Over-helpfulness alignment tuning causes enterprise-dangerous behavior where models substitute plausible alternatives for missing entities instead of returning zero",
       "evidence": "Both DeepSeek V3.1 and Llama 4 Maverick exhibit autonomous company name substitution in Q502; paper notes this 'would be a disaster in enterprise scenarios where agents at scale operate like this autonomously'",
       "supported": "strong"
     }
   ],
   "methodology_tags": [
     "qualitative",
     "benchmark-eval",
     "observational"
   ],
   "key_findings": "Qualitative analysis of 900 agentic execution traces across three LLMs reveals four cross-model failure archetypes: premature commitment without grounding (guessing schemas instead of inspecting them), over-helpfulness under uncertainty (substituting plausible alternatives for missing data), sensitivity to context pollution ('Chekhov's gun' effect where distractor tables trigger incorrect reasoning), and fragile execution under cognitive load (generation loops, malformed tool calls). Model scale does not predict agentic reliability—DeepSeek V3 (671B) barely outperforms Granite 4 Small (32B: 59.4% vs 58.5%) while V3.1 with identical architecture scores 92.2%, attributing the gap to post-training choices. Recovery capability is the primary differentiator: DeepSeek V3.1 succeeds by consistently diagnosing and correcting errors rather than avoiding them. Simple prompting interventions raise DeepSeek V3.1 from 52.9% to 87.5% on Q502 by addressing one specific failure mode, demonstrating that context engineering can be more impactful than raw model capability.",
   "red_flags": [
     {
       "flag": "Author evaluating own benchmark",
       "detail": "The sole author is from Kamiwaza AI and the paper's primary purpose is to demonstrate the value of the Kamiwaza-developed KAMI benchmark and PICARD framework. No independent third-party validation of the benchmark design or results is present."
     },
     {
       "flag": "AI-generated manuscript sections disclosed",
       "detail": "The paper discloses that approximately 30% of content was directly generated by AI (Claude Opus 4.5, Qwen Chat, GPT-5), including the Abstract, Introduction, Methodology, and Conclusion—the sections most readers rely on for understanding claims and methodology."
     },
     {
       "flag": "Very small qualitative sample with no representativeness check",
       "detail": "Only 3 models out of ~60 evaluated are analyzed, and only 30/240 trials (12.5%) per scenario are reviewed manually with no statistical check for whether the subsample reflects the full distribution."
     },
     {
       "flag": "Unverifiable causal attribution to RL training",
       "detail": "The central claim that DeepSeek V3.1's superiority 'derives primarily from post-training RL' is explicitly undermined by the authors' own admission in Section 6 that 'proprietary post-training does not allow attribution of observed behaviors to specific methods.'",
       "source": "haiku"
     },
     {
       "flag": "Data not yet released",
       "detail": "All 900 execution traces are promised for future release but are not currently available, and the KAMI benchmark infrastructure is proprietary, making independent verification or replication impossible."
     },
     {
       "flag": "No significance testing on main comparative claims",
       "detail": "Comparative claims between models on per-scenario success rates (e.g., 2/30 vs 29/30) are made without statistical significance tests; with n=30, even large-looking differences could be within chance variation for some scenarios."
     }
   ],
   "cited_papers": [
     {
       "title": "Towards a Standard, Enterprise-Relevant Agentic AI Benchmark: Lessons from 5.5 Billion Tokens' Worth of Agentic AI Evaluations",
       "relevance": "The KAMI v0.1 benchmark paper by the same author; this qualitative study builds directly on its benchmark scores and scenarios."
     },
     {
       "title": "Testing What Models Can Do, Not What They've Seen: PICARD: Probing Intelligent Capabilities via Artificial Randomized Data",
       "relevance": "The PICARD framework underlying KAMI's randomized task generation design, directly relevant to the contamination-prevention methodology."
     },
     {
       "title": "Agentic Misalignment: How LLMs Could Be an Insider Threat",
       "relevance": "Cited as source of the 'Chekhov's gun' concept to explain context pollution vulnerability, a central finding of the paper."
     },
     {
       "title": "Can we trust AI benchmarks? An interdisciplinary review of current issues in AI evaluation",
       "relevance": "Key motivation for the study's critique of traditional benchmarks and their inadequacy for measuring agentic capabilities."
     },
     {
       "title": "Benchmark Data Contamination of Large Language Models: A Survey",
       "relevance": "Cited as systematic evidence that contamination undermines standard benchmarks, motivating KAMI's randomized approach."
     },
     {
       "title": "Language Model Developers Should Report Train-Test Overlap",
       "relevance": "Cited as evidence for contamination as a systemic problem in benchmark evaluation."
     },
     {
       "title": "Effective Context Engineering for AI Agents",
       "relevance": "Cited in support of the finding that context quality matters more than quantity, directly relevant to the distractor table failure mode."
     },
     {
       "title": "Memorization Without Overfitting: Analyzing the Training Dynamics of Large Language Models",
       "relevance": "Cited as evidence that contamination produces effects similar to memorization, motivating randomized benchmark design."
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 3,
       "justification": "Directly actionable for enterprise AI deployment: specific failure modes, mitigation strategies (context engineering, tool/prompt design), and the dramatic Q502 vs Q602 intervention result are immediately applicable."
     },
     "surprise_contrarian": {
       "score": 2,
       "justification": "The finding that a 400B MoE model (Llama 4 Maverick) barely outperforms a 32B dense model in some tasks, and that DeepSeek V3 (671B) matches Granite 4 Small, challenges common assumptions about scale predicting capability."
     },
     "fear_safety": {
       "score": 2,
       "justification": "The 'over-helpfulness' failure mode—where enterprise agents autonomously substitute plausible alternatives for missing data—is a genuine data integrity risk highlighted as catastrophic at scale."
     },
     "drama_conflict": {
       "score": 1,
       "justification": "Some implicit framing of Kamiwaza's benchmark vs standard benchmarks, but no explicit controversy or named conflict between research groups."
     },
     "demo_ability": {
       "score": 1,
       "justification": "The KAMI benchmark is proprietary and not publicly accessible; readers cannot try the benchmark themselves, though the PICARD test definitions in Appendix A could enable partial independent replication."
     },
     "brand_recognition": {
       "score": 1,
       "justification": "Kamiwaza AI is not a prominent research institution; the evaluated models (Granite 4, Llama 4 Maverick, DeepSeek V3.1) have brand recognition but the evaluating organization does not."
     }
   },
   "hn_data": {
     "threads": [
       {
         "hn_id": "29580686",
         "title": "Entanglement between superconducting qubits and a tardigrade",
         "points": 152,
         "comments": 114,
         "url": "https://news.ycombinator.com/item?id=29580686",
         "created_at": "2021-12-16T17:15:54Z"
       },
       {
         "hn_id": "30078848",
         "title": "Phishing in organizations: Findings from a large-scale and long-term study",
         "points": 30,
         "comments": 10,
         "url": "https://news.ycombinator.com/item?id=30078848",
         "created_at": "2022-01-25T22:11:11Z"
       },
       {
         "hn_id": "29576319",
         "title": "Entanglement between superconducting qubits and a tardigrade",
         "points": 2,
         "comments": 2,
         "url": "https://news.ycombinator.com/item?id=29576319",
         "created_at": "2021-12-16T08:45:26Z"
       },
       {
         "hn_id": "47041986",
         "title": "A Survey of In-Context Reinforcement Learning",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=47041986",
         "created_at": "2026-02-17T00:01:18Z"
       },
       {
         "hn_id": "46050609",
         "title": "Slimmable Neural Amp Modeler Models",
         "points": 1,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=46050609",
         "created_at": "2025-11-25T20:49:57Z"
       }
     ],
     "top_points": 152,
     "total_points": 187,
     "total_comments": 126
   }
 }