Embedded Files
System cards, model evaluations, risk frameworks, and the public form of frontier-AI safety claims
Saman Samadi works on AI safety documentation and public accountability, with a focus on system cards, model-release documentation, evaluation evidence, risk frameworks, safety claims, mitigation language, residual risk, and the public documents through which frontier-AI systems become answerable.
This work develops from a longer research practice concerned with form, evidence, interpretation, public language, and the conditions under which complex material can be read, tested, and held open to judgement. In AI safety documentation, those concerns acquire a new technical and institutional setting.
The project focuses on the documentary forms through which AI laboratories, governance bodies, and safety teams make claims about frontier models. Its aim is practical: to contribute to documents that remain accurate, readable, internally coherent, proportionate to their evidence, and accountable to public scrutiny.
Portfolio Artifacts
Portfolio Artifacts
This portfolio gathers public-facing work in AI safety documentation, model-release analysis, governance documentation, evaluation-to-claim reasoning, and technical rewriting. The artifacts are designed to show how safety claims, risk frameworks, evaluation evidence, mitigations, residual risk, and uncertainty are carried into documents that can be read, questioned, compared, and revised.
Research Articles and Reports
Research Articles and Reports
System Cards and the Public Form of AI Accountability
System Cards and the Public Form of AI Accountability
Article / PDF | 26 February 2026.
Examines system cards and model-release documents as public accountability artifacts, with attention to evaluation evidence, safety claims, mitigation language, residual risk, and deployment reasoning.
When an Evaluation Becomes a Claim
When an Evaluation Becomes a Claim
Article / PDF | 1 March 2026.
Analyses how evaluation results move into public safety language, focusing on benchmark interpretation, red-team findings, uncertainty, mitigation, residual risk, and the risk of overstatement when technical evidence becomes public prose.
Reading Tables in AI Safety Reports
Reading Tables in AI Safety Reports
Article / PDF / checklist | 18 March 2026.
Examines how tables, scorecards, benchmark results, effect sizes, thresholds, and attack-success rates support public safety claims in AI system cards and risk reports, with case studies from Anthropic, OpenAI, and Google DeepMind.
From Risk Framework to Public Claim
From Risk Framework to Public Claim
Article / PDF | 10 June 2026.
Studies how responsible-scaling policies, preparedness frameworks, AI risk-management documents, and public governance materials organise the movement from risk classification to institutional responsibility and public accountability.
How Frontier AI Labs Present Safety Claims
How Frontier AI Labs Present Safety Claims
Comparative report / PDF | 3 June 2026.
Reviews frontier AI system cards, model cards, safety reports, and release documents across major AI labs, examining how capability claims, safety claims, evaluation evidence, mitigations, residual risk, uncertainty, and deployment reasoning are made public.
Frameworks and Tools
Frameworks and Tools
A Safety-Claim Audit Framework for Frontier AI Documentation
A Safety-Claim Audit Framework for Frontier AI Documentation
Framework / PDF / templates | 17 June 2026.
Provides a practical method for reviewing safety claims in system cards, model cards, safety reports, preparedness documents, and release artifacts, with attention to claim type, evidence source, evaluation method, uncertainty, mitigation, residual risk, terminology, audience risk, and accountability.
Focus: Safety-claim review; audit methods; evidence traceability; terminology discipline.
Links: Website | PDF | Compact template | Expanded template
Practical Documentation Samples
Practical Documentation Samples
Rewriting Threshold Classification under Uncertainty
Rewriting Threshold Classification under Uncertainty
Technical rewriting sample | 27 May 2026.
Demonstrates how a dense Anthropic risk-report passage on ASL-3 safeguards, automated evaluations, uplift trials, and proxy-task uncertainty can be rewritten for governance, public, and internal documentation audiences without overstating the evidence.
Documentation Skills Demonstrated
Documentation Skills Demonstrated
Claim and Evidence
Claim and Evidence
Evidence-to-claim reasoning
Safety-claim review
Mitigation and residual-risk language
Threshold and safeguard classification
Evaluation and Governance
Evaluation and Governance
System-card and model-release document analysis
Evaluation and benchmark communication
Risk-framework comparison
Documentation Practice
Documentation Practice
Terminology consistency
Audience-sensitive technical rewriting
Public-facing research documentation
Areas of Focus
Areas of Focus
AI safety documentation and external artifacts
AI safety documentation and external artifacts
System cards, model-release documents, safety reports, research summaries, and public-facing artifacts that translate internal research and evaluation evidence into accountable documents.
Responsible AI governance and risk documentation
Responsible AI governance and risk documentation
Risk frameworks, preparedness policies, responsible-scaling documents, assurance methods, governance reports, and the procedural language through which AI safety commitments become legible.
Model-evaluation communication
Model-evaluation communication
Evaluation summaries, red-team findings, benchmark interpretation, mitigation reporting, residual-risk language, table clarity, and the movement from technical result to public claim.
Selected Previous Research
Selected Previous Research
Becoming Persian Music: A Poststructuralist Approach to Composition
PhD thesis, University of Cambridge, 2023.
DOI: https://doi.org/10.17863/CAM.102136
Composing Otherwise: Actualising Difference in Persian Musical Ontology
Journal of the Royal Musical Association, vol. 151, no. 2, forthcoming Autumn 2026.
A Comparative Study on Two Performances of Karlheinz Stockhausen’s Klavierstück X
Perspectives of New Music, vol. 61, no. 2, Summer 2023.
DOI: https://doi.org/10.1353/pnm.2023.a950410
AI Has Entered Composition at the Weakest Point: Form
Substack, 13 May 2026.
https://samansamadi.substack.com/ai-composition-form
Full research record: Research Publications and Presentations
Contact
Contact
For research, writing, documentation, and AI safety documentation enquiries, please email ss2728@cantab.ac.uk.
Last updated: June 2026
Artwork: Jan Davidsz. de Heem (1628); Still Life with Books and a Violin.
Page updated
Google Sites
Report abuse