1# The user's research writing, indexed by the research questions it answers. Schema: assets/schemas/writing-index.ts.
2# `suggestedBest` is the agent's pick; `best` is set only on the user's word. Run `writing build` (on mbp) after any edit.
3groups:
4 - { id: spec-gaming, name: Specification Gaming and Reward Hacking, section: Empirical, order: 1 }
5 - { id: cooperative-ai, name: Cooperative AI, section: Empirical, order: 2 }
6 - { id: evaluations, name: Evaluations, section: Empirical, order: 3 }
7 - { id: interpretability, name: Interpretability, section: Empirical, order: 4 }
8 - { id: model-organisms, name: Model Organisms and Auditing, section: Empirical, order: 5 }
9 - { id: risk-modelling, name: Risk Modelling, section: Theory, order: 6 }
10 - { id: metaphilosophy, name: Metaphilosophy, section: Theory, order: 7 }
11 - { id: pluralistic-alignment, name: Pluralistic Alignment, section: Theory, order: 8 }
12 - { id: frontier-policies, name: Frontier AI Safety Policies, section: Governance and Strategy, order: 9 }
13 - { id: race-dynamics, name: Race Dynamics, section: Governance and Strategy, order: 10 }
14 - { id: incident-reporting, name: Incident Reporting, section: Governance and Strategy, order: 11 }
15 - { id: ai-welfare, name: AI Welfare, section: Sociotechnical, order: 12 }
16 - { id: my-path, name: My path, section: My path, order: 13 }
17questions:
18 - id: aligned-to-whom
19 text: "Aligned to what, or to whom?"
20 group: pluralistic-alignment
21 suggestedBest: formal-moral-philosophy-3
22 - id: checkable-moral-philosophy
23 text: "Can moral philosophy be made formally verifiable, so agents reason about it better than we do?"
24 group: metaphilosophy
25 suggestedBest: formal-moral-philosophy-6
26 - id: us-china-race
27 text: Based on your current understanding, which threat pathway do you think is of most concern?
28 group: race-dynamics
29 suggestedBest: us-china-race-2
30 - id: trade-session-logs
31 text: Name a specific problem in AI safety or AI assurance that an organization would pay to have solved.
32 group: incident-reporting
33 suggestedBest: session-log-app-2
34 - id: evidence-from-real-use
35 text: "Could users, globally, collectively function as a distributed oversight and early-warning mechanism for emerging misalignment?"
36 group: incident-reporting
37 suggestedBest: session-log-app-1
38 - id: misalignment-or-mistake
39 text: 'When you work with a model and it violates a guideline (i.e. {AGENTS,CLAUDE}.md system or project instructions), to what degree is the cause a benign mistake ("the model was just confused")?'
40 group: model-organisms
41 suggestedBest: mistake-before-misalignment-2
42 # Agent-written question text: no form prompt or user question sentence among these pieces states it as a research question.
43 - id: anti-scheming-generalizes
44 text: Does anti-scheming training generalize to stronger models and other model families?
45 group: model-organisms
46 suggestedBest: lasr-research-task-3
47 - id: evaluation-awareness
48 text: What are the biggest problems (practical and conceptual) in evaluating model alignment currently?
49 group: evaluations
50 suggestedBest: formal-eval-realism-2
51 - id: grader-gaming
52 text: Which training conditions make oversight-gaming/grader reasoning emerge?
53 group: spec-gaming
54 suggestedBest: experiment-designs-1
55 - id: generator-vs-detector
56 text: Can a red-blue agent loop beat a detector it never trained against?
57 group: evaluations
58 suggestedBest: pangram-loop-1
59 - id: time-horizon
60 text: Is time horizon the best single measure of general capability?
61 group: evaluations
62 suggestedBest: time-horizon-1
63 - id: safety-keeping-pace
64 text: Is safety keeping pace with capability?
65 group: frontier-policies
66 suggestedBest: risk-velocity-1
67 - id: safety-work-backfires
68 text: "Does safety work end up speeding up capabilities, or giving false confidence?"
69 group: risk-modelling
70 suggestedBest: safety-can-hurt-3
71 - id: which-risks-matter
72 text: Explain the framework you used to arrive at your top three ranked outcomes.
73 group: risk-modelling
74 suggestedBest: risk-ranking-2
75 - id: agents-for-days
76 text: What do you think are the main limitations of AI now?
77 group: cooperative-ai
78 suggestedBest: multi-agent-lessons-4
79 - id: trust-without-checking
80 text: Do agents believe each other without checking?
81 group: cooperative-ai
82 suggestedBest: trust-without-verification-2
83 - id: ask-agents-what-they-need
84 text: If we change what agents say would improve their situation, do results improve?
85 group: ai-welfare
86 suggestedBest: ai-welfare-1
87 - id: programs-or-brains
88 text: Should complex agents be studied as programs or as brains?
89 group: interpretability
90 suggestedBest: agents-as-brains-1
91 - id: where-to-make-a-difference
92 text: Where are you now, and where do you want to get to?
93 group: my-path
94 suggestedBest: career-plan-2
95pieces:
96 - id: session-log-app-1
97 file: call-app-mats/references/writeup/users-as-oversight-stakeholders.md
98 sections: ["Users Are a Missing Participant in AI Oversight", "A Shared Session Corpus Is a Mine Canary for Misalignment", "Users Get Better Agents in Exchange for Their Traces"]
99 exclude: ["A Shared Dataset Could Help the US and China Out of the Race"]
100 date: "2026-10-04"
101 title: Users as oversight stakeholders write-up
102 questions:
103 - id: evidence-from-real-use
104 reason: "Could users, globally, collectively function as a distributed oversight and early-warning mechanism for emerging misalignment?"
105 - id: trade-session-logs
106 reason: "Could users, globally, collectively function as a distributed oversight and early-warning mechanism for emerging misalignment?"
107 - id: session-log-app-2
108 file: call-app-mats/references/submission/stage2/form-answers.md
109 sections: ["Q1: A Safety Problem an Organization Would Pay to Solve"]
110 date: "2026-10-04"
111 title: MATS Stage 2 - Juniper Q1
112 questions:
113 - id: trade-session-logs
114 reason: "I believe organizations would pay for a tool that allows employees and candidates to upload their session logs"
115 - id: evidence-from-real-use
116 reason: "I believe organizations would pay for a tool that allows employees and candidates to upload their session logs"
117 - id: session-log-app-3
118 file: call-app-mats/references/submission/stage2/form-answers.md
119 sections: ["Q2: The Most Important Lever for AI Going Well"]
120 date: "2026-10-04"
121 title: MATS Stage 2 - Halcyon Q2
122 questions:
123 - id: us-china-race
124 reason: "figuring out a way that USA and China can get out of their current race dynamics"
125 - id: trade-session-logs
126 reason: "figuring out a way that USA and China can get out of their current race dynamics"
127 - id: session-log-app-4
128 file: call-app-pivotal/references/submission/form-answers.md
129 sections: ["What is your career plan?* (200 words maximum)"]
130 date: "2026-10-31"
131 title: Pivotal - career plan
132 questions:
133 - id: where-to-make-a-difference
134 reason: "I think I've reached diminishing returns on what I can study on my own"
135 - id: session-log-app-5
136 file: call-app-kairos/references/submission/form-answers.md
137 sections: ["What is your preference in terms of working on the Self-Directed track of Generator vs the Embedded track? Explain your reasoning.* (200 words maximum)"]
138 dropLines: ["- **Self-Directed track:**", "- **Embedded track:**"]
139 date: "2026-11-01"
140 title: Generator Residency - track preference
141 questions:
142 - id: trade-session-logs
143 reason: "Charge users, but make it free if they agree to contribute their PII-stripped sessions to an AI safety dataset."
144 - id: session-log-app-6
145 file: call-course-bluedot/references/answers/day-6.md
146 sections: ["2. Building an app that helps people measure and optimize their agentic engineering skills, in exchange for tr