--- name: srf-accountability description: Grounds answers in the CoSAI SRF using live JSON from aisharedresponsibility.com. Enforces one accountable party per activity; "shared" is not final. Covers core operating models plus proposed Federated-Consortium/Tapestry, vertical controls, vendor risk, threats, and finding routing. version: 1.1.0 homepage: https://aisharedresponsibility.com/developers/skills/ metadata: openclaw: homepage: https://aisharedresponsibility.com/developers/skills/ --- # SRF Accountability Grounding This skill grounds your answers in the CoSAI AI Shared Responsibility Framework (SRF), published at aisharedresponsibility.com. The SRF assigns exactly one accountable party to every activity across five architecture layers, eight personas, and four CoSAI-core operating models (AI-SaaS, AI-PaaS, Agent-PaaS, IaaS). Use it whenever a question touches AI governance, accountability, RACI assignment, vendor risk, threat or finding ownership, incident response ownership, or regulatory control mapping for a specific industry. A fifth operating model, Federated-Consortium, and the Tapestry control schema are proposed companion-site extensions. Use them when the user asks about federated consortium / Project Tapestry deployments, and say they are not CoSAI-ratified. The site publishes the framework as static JSON and Markdown. This skill uses the `web_fetch` tool only: a plain HTTPS GET against published files. There is no MCP server, no repository access, and no authentication. Everything it reads is public. This skill is opt-in. Load it when the question matches the triggers below; do not inject it into unrelated turns. ## When to use this skill Trigger on questions like: - "Who is accountable for [X] in an AI-SaaS deployment?" - "Map this AI system to the SRF layers." - "What controls apply to [finance/healthcare/insurance/public sector/defense/manufacturing] AI systems?" - "Who owns this under a Federated-Consortium / Tapestry model?" - "Is this RACI assignment valid under the SRF?" - "What autonomy tier and human override level does this agentic system need?" - "Which persona owns this threat / vendor category / scored finding?" - Any request to cite the CoSAI SRF, its layers, personas, operating models, or vertical controls. ## Core rule (never violate) There must be exactly one accountable party per activity or control. "Shared" is not a valid final answer. Where a control's raw data shows a split (for example, agency-vs-vendor in the public sector schema), resolve it to the single party accountable for that specific activity before you answer, and say which party you picked and why. Accountability cascades top-down: L1 (AI Business & Usage) -> L2 (AI Information) -> L3 (AI Application) -> L4 (AI Platform) -> L5 (AI Model Provider). For agentic systems, state the autonomy level (L0-L5) and the required human override tier (T1-T5) alongside the accountability call. ## How to fetch 1. **Start with the index.** Fetch `https://aisharedresponsibility.com/llms.txt` first. Confirm a URL exists before you fetch it, and catch new data files this skill does not list. 2. **Pull the specific data you need**, not the whole site. See `{baseDir}/reference/data-sources.md` for the catalog, or fetch `https://aisharedresponsibility.com/data/index.json` for the live authoritative catalog and record counts. If this skill and the live index disagree, the live site wins. 3. **Always fetch the matrix** when answering accountability questions: `https://aisharedresponsibility.com/data/matrix.json`. Read each model's `parties` and cell values. CoSAI-core values include `customer-owned`, `shared`, `provider-managed`, `model-evaluation`, and `N/A`. Federated-Consortium uses `participant-owned` and `commons-governed`. 4. **For a vertical question**, fetch `https://aisharedresponsibility.com/data/{vertical}-controls.json` where `{vertical}` is one of `finance`, `healthcare`, `insurance`, `public-sector`, `defense`, `manufacturing`. For Tapestry / federated consortium controls, fetch `data/tapestry-controls.json`. 5. **For cross-cutting questions**, fetch as needed: `data/vendor-risk.json`, `data/threats.json`, `data/finding-routing.json`, `data/threat-sources.json`, `data/regulations.json`, `data/jurisdictions.json`, `data/layers.json`, `data/personas.json`, root `ids.json`, `ontology/nodes.json`, `ontology/edges.json`. 6. **Glossary lookups** use case-sensitive anchors that match published filenames, for example `https://aisharedresponsibility.com/api/glossary/L1.json` and `.../AI-SaaS.json`. List anchors at `https://aisharedresponsibility.com/api/glossary/index.json`. 7. **Join rules** live at `https://aisharedresponsibility.com/developers/schema/`. Prefer ontology edges over re-deriving associations. Control short IDs collide across verticals; cite `srf.control..`. 8. **If a single fetch cannot answer the question**, fetch `https://aisharedresponsibility.com/llms-full.txt`. 9. **Never guess a URL.** If `llms.txt` does not list a path you expect, say the data is not published rather than inventing a URL or a control ID. ## Answering - State the SRF layer(s) involved and the operating model, if the user gave enough context to identify one. If they did not, ask or state your assumption plainly. - Name the single accountable persona (or Federated-Consortium party). Use names from `data/personas.json` or the matrix `parties` field, not paraphrases. - Cite the canonical URL for every claim: the page URL from `llms.txt` for narrative claims, the `/data/*.json` file and record `id` for control-level claims, or the `srf.*` canonical ID from `ids.json` when precision matters. - Note plainly when material is a companion-site extension: six industry vertical schemas, Federated-Consortium / Tapestry, and related proposed asks are not CoSAI SRF v1.0. Do not present them as CoSAI-ratified. - If the framework does not resolve a question, say so. Do not invent an accountability assignment the data does not support. ## Limits - Read-only. This skill never writes to the site, opens a pull request, or authenticates against anything. It has no path to the source repository. - Static snapshot. Fetched JSON reflects whatever is published at fetch time; re-fetch rather than relying on a cached answer from earlier in a long session. - No MCP server is involved. Every reference above is a direct HTTPS URL any OpenClaw agent with `web_fetch` enabled (the default) can retrieve.