Publication: LessWrong/Alignment Forum Post Preparation

[PipFoweraker]

Goal

Prepare a compelling post for LessWrong / Alignment Forum that introduces interrogatory model cards to the AI safety and alignment community.

Why LW/AF First?

1. Audience alignment - community cares about transparency, evals, AI governance
2. Quality feedback - rigorous commenters will stress-test the proposal
3. Network effects - shares into adjacent communities (EA, AI policy, researchers)
4. Credibility building - establishes intellectual foundation before broader push

Post Structure (Draft Outline)

1. The Problem (Hook)

• Model cards are often PR documents, not technical documentation
• Selective disclosure, vague provenance, flattering metrics
• Regulation is coming (EU AI Act) but most cards won't meet the bar
• Link to "Evals Gap" and "Science of Evals" discourse

2. Existing Approaches (What's Been Tried)

• Mitchell et al. 2019 Model Cards
• HuggingFace adoption (widespread but uneven quality)
• System Cards (Anthropic, OpenAI) - better but still self-reported
• AI Cards (APF 2024) - machine-readable but not interrogatory
• Why these aren't enough

3. Interrogatory Model Cards (The Proposal)

• Design goal: hard to obfuscate, easy to audit, low friction to author
• CAN/SHOULD/MUST framework
• Sharp yes/no questions (examples)
• Evidence linking (demo)
• Risk-tiered adversarial framing

4. Schema & Tooling (Concrete Artifacts)

• JSON-LD schema extending Croissant
• Permissive design: non-disclosure is visible, not blocking
• Validation tools, card builder
• Live demo link

5. What This Enables

• Comparable disclosure across models
• Automated compliance checking
• Third-party verification hooks
• Regulatory alignment (EU AI Act Annex IV mapping)

6. Limitations & Open Questions

• Doesn't solve: adversarial non-compliance, capability evaluations themselves
• Open: tier assignment, link rot, proprietary evidence
• Request for feedback on specific design choices

7. Call to Action

• Try the schema on your model
• Contribute to the spec
• Discuss: what questions should be MUST?

Writing Considerations

LW/AF Norms

• Show your work (reasoning, not just conclusions)
• Acknowledge uncertainty and limitations
• Engage with likely objections preemptively
• Concrete examples > abstract claims

Avoid

• Marketing speak
• Overclaiming ("this solves AI transparency")
• Dismissing existing work unfairly

Include

• Links to all artifacts (schema, tools, examples)
• Interactive demo if available
• Explicit request for specific feedback

Pre-Publication Checklist

• Schema finalized (at least v0.2)
• At least 2-3 example cards created
• Validation tooling working
• Demo website live
• Post draft reviewed by 2+ people
• Cross-post strategy (LW vs AF vs both?)

Deliverables

1. docs/publication/lw-af-post.md - draft post
2. Supplementary materials (diagrams, examples)
3. Response plan for comments/feedback

Timeline Considerations

• Don't rush - quality matters more than speed for this audience
• Consider posting when you can actively engage with comments (not before travel, etc.)
• Monday-Wednesday posts tend to get more engagement

Related Issues

• All other issues feed into this
• Data: Example Cards (needed for concrete demos)
• Frontend: Interactive Demo (linked from post)