llms.txt

# AI Shared Responsibility: CoSAI SRF

> The CoSAI AI Shared Responsibility Framework (SRF) defines who is accountable for what across AI deployments: five enterprise architecture layers, eight named personas, and four operating models. One accountable party per activity. "Shared" is not a valid final answer. The SRF is not another AI governance framework; it complements the regulations and standards organizations already follow (NIST AI RMF, ISO/IEC 42001, the EU AI Act) by answering the question they leave open: who is accountable. Decision makers can approve a deployment once they can name one accountable party per activity. This site provides the framework reference, industry-specific control schemas, regulatory crosswalks, practitioner tools, and copy-ready system instructions. Industry vertical schemas are independently proposed extensions to CoSAI SRF v1.0 and are not part of the official CoSAI release.

## Core framework

- [Framework Overview](https://aisharedresponsibility.com/framework/): The five-layer enterprise architecture model (L1 AI Business & Usage through L5 AI Model Provider), eight personas, four operating models, and agentic extensions including autonomy classification (L0 to L5) and human override tiers (T1 to T5).
- [Operating Models](https://aisharedresponsibility.com/operating-models/): How accountability assignments shift across AI-SaaS, AI-PaaS, Agent-PaaS, and IaaS deployments. The layer responsibility matrix per operating model.
- [Personas](https://aisharedresponsibility.com/personas/): The eight named personas and their accountability domains across the five layers.
- [Glossary](https://aisharedresponsibility.com/glossary/): Canonical definitions for all SRF terms with stable anchor links. Layers (L1-L5), personas, operating models, accountability rules, agentic extensions (including companion terms for accountability continuity, agent replication, fail-closed override, persistence scope, enforcement plane, and protocol-independent evidence), controls vocabulary, and evidence terms.
- [NICE Mapping](https://aisharedresponsibility.com/framework/nice-mapping/): Crosswalk between SRF roles and NICE Cybersecurity Workforce Framework categories.
- [AI Security Lifecycle](https://aisharedresponsibility.com/framework/security-lifecycle/): How the SRF supplies the accountability axis across six categories of AI security: threat modeling, adversarial testing and red teaming, secure development and supply chain, control implementation, vulnerability management, and incident response. States the SRF contribution, the MOSAIC member work served, and the on-site tools and data serving each. Positions SRF canonical IDs as the accountability axis of the MOSAIC OpenCRE shared taxonomy. Includes the live finding routing reference. Companion data: the threat-to-accountability crosswalk at /data/threats.json and the finding routing reference at /data/finding-routing.json.
- [Attack Inventory](https://aisharedresponsibility.com/framework/attack-inventory/): Readable, filterable view of 58 published AI, agentic, and business-process attack classes with checked ATLAS, ATT&CK, OWASP, FINOS, CWE, arXiv, and CVE ids, the pinned catalog versions, what checking the seed table and the business-process classes found, and the known limits of this version. 45 of the 58 classes link to an SRF crosswalk row for layer and owner; 7 business-process classes also link to finance, insurance, or manufacturing controls.
- [Framework Comparison](https://aisharedresponsibility.com/compare/): How the CoSAI SRF relates to and complements adjacent frameworks (NIST AI RMF, ISO 42001, EU AI Act, CSA AICM, MITRE ATLAS, and AI SAFE²) by answering who is accountable.
- [Proposed SRF next-version agentic extensions](https://aisharedresponsibility.com/papers/srf-vnext-agentic-extensions/): Independently proposed companion asks for a future CoSAI SRF paper (spawn, fail-closed halt, enforcement planes, protocol-independent evidence). Not part of CoSAI SRF v1.0. Data: /data/srf-vnext-extensions.json.
- [AI and Agent Security Synthesis](https://aisharedresponsibility.com/agentic-ai-security/): Topic-organized synthesis of axioms and invariant rules across fifty-two AI and agentic AI security papers (BIML, CIS, CSA, CoSAI, NIST, OWASP, SANS, ETSI, ENISA, FMF, OpenSSF, AI Verify). Companion index at /agentic-ai-security/llms.txt; full text at /agentic-ai-security/llms-full.txt.
- [Project Tapestry Controls](https://aisharedresponsibility.com/tapestry/controls/): 26 accountability controls mapping AI Alliance Project Tapestry governance ADRs onto the SRF layers, for a federated AI consortium rather than a single enterprise. Adds the Federated-Consortium operating model and two personas, sovereign-participant-node and consortium-core-coordinator. Filter by layer, lifecycle stage, governing domain, disclosure tier, and status. Federated-Consortium, the TAP-SRF controls, and both personas are proposed extensions and are not CoSAI-ratified. Companion data at /data/tapestry-controls.json.
- [Project Tapestry Example Prompts](https://aisharedresponsibility.com/tapestry/example-prompts.md): Copy-ready prompts grounding Project Tapestry questions in the published JSON, with shared rules that verify Federated-Consortium is present in matrix.json and both Tapestry personas are present in personas.json before answering.

## Industry verticals

Each vertical adds a regulatory crosswalk, risk-tier calibration, and evidence specifications aligned to sector audit practices. All six are live.

- [Industries Overview](https://aisharedresponsibility.com/industries/): All six verticals with status, control counts, and regulatory coverage.
- [Financial Services](https://aisharedresponsibility.com/finance/): 40 controls across five SRF layers. SR 26-2 (agentic AI gap), FINOS AIGF, OWASP LLM Top 10, EU AI Act. Four MRM stages.
- [Financial Services Controls](https://aisharedresponsibility.com/finance/controls/): Full 40-control schema with layer mapping and evidence requirements.
- [Financial Services How-To](https://aisharedresponsibility.com/finance/how-to/): Practitioner guide and effective challenge workpaper.
- [Healthcare](https://aisharedresponsibility.com/healthcare/): 40 controls for clinical decision support, AI-assisted diagnostics, and agentic care coordination. FDA TPLC, FDA PCCP, ONC HTI-1, HIPAA, EU AI Act, IEC 62304, ISO 14971. FHIR AuditEvent evidence pointers.
- [Healthcare Controls](https://aisharedresponsibility.com/healthcare/controls/): Full 40-control schema with clinical stage mapping.
- [Healthcare How-To](https://aisharedresponsibility.com/healthcare/how-to/): Practitioner guide for clinical AI deployment.
- [Insurance](https://aisharedresponsibility.com/insurance/): 40 controls for underwriting, claims, and vendor model governance. NAIC AI Model Bulletin, Colorado Regulation 10-1-1 (July 2026 deadline), NYDFS CL 7, NAIC AI Systems Evaluation Tool.
- [Insurance Controls](https://aisharedresponsibility.com/insurance/controls/): Full 40-control schema.
- [Insurance How-To](https://aisharedresponsibility.com/insurance/how-to/): Practitioner guide and market conduct exam workpaper.
- [Public Sector](https://aisharedresponsibility.com/public-sector/): 40 controls for federal civilian agencies (FCEB). OMB M-25-21 minimum practices, OMB M-25-22 acquisition terms, FedRAMP 20x KSIs, NIST AI RMF. September 22, 2026 deadline anchor. Responsibility split on every control.
- [Public Sector Controls](https://aisharedresponsibility.com/public-sector/controls/): Full 40-control schema with federal AI stage mapping.
- [Public Sector How-To](https://aisharedresponsibility.com/public-sector/how-to/): Practitioner guide and M-25-21 compliance workpaper.
- [Defense / DoD](https://aisharedresponsibility.com/defense/): 53 controls for DoD components and the defense industrial base. DoD Responsible AI tenets, CMMC 2.0, DoD CC SRG, NIST 800-171. IL4, IL5, IL6 impact levels. Separate Non-NSS and NSS control tiers.
- [Defense Controls](https://aisharedresponsibility.com/defense/controls/): Full 53-control schema with impact level mapping.
- [Defense How-To](https://aisharedresponsibility.com/defense/how-to/): Practitioner guide and CMMC-aligned evidence workpaper.
- [Manufacturing](https://aisharedresponsibility.com/manufacturing/): 45 controls for OT/ICS, product-embedded AI, and IT-side manufacturing. EU AI Act high-risk obligations (August 2026), EU Machinery Regulation 2023/1230 (January 2027), IEC 62443 OT cybersecurity zones, ISO 42001.
- [Manufacturing Controls](https://aisharedresponsibility.com/manufacturing/controls/): Full 45-control schema with lifecycle stage mapping.
- [Manufacturing How-To](https://aisharedresponsibility.com/manufacturing/how-to/): Practitioner guide and compliance workpaper.

## Tools

- [Tools Overview](https://aisharedresponsibility.com/tools/): All interactive assessment and governance tools.
- [Accountability Decision Record](https://aisharedresponsibility.com/tools/decision-record/): Generates a signable one-page decision record for one AI deployment. Pick an operating model to prefill the per-layer accountability split, then name exactly one accountable party per SRF layer (resolving every "shared"), record residual gaps, autonomy and override tier, and a sign-off. Enforces the core rule that shared is not a valid final answer. Browser-based, localStorage only, PDF and JSON export. A downloadable offline DOCX template (srf-decision-record-template.docx) mirrors the form for air-gapped or browser-restricted environments.
- [Regulation Discovery](https://aisharedresponsibility.com/tools/regulation-discovery/): 4-step wizard identifying applicable regulations by jurisdiction, risk tier, deployment context, and industry vertical. Layer-mapped regulatory profile output with PDF export.
- [Controls Assessment (AICM)](https://aisharedresponsibility.com/tools/controls-assessment/): Assess control implementation status against the SRF across all five layers.
- [Layer Matrix](https://aisharedresponsibility.com/tools/layer-matrix/): Visual matrix of accountability assignments across all five SRF layers and operating models.
- [Policy Pyramid](https://aisharedresponsibility.com/tools/policy-pyramid/): Governance policy hierarchy tool showing how L1 policy cascades through lower layers.
- [Security Controls](https://aisharedresponsibility.com/tools/security-controls/): Security-specific control reference with SRF layer mapping.
- [SRF Stress Test](https://aisharedresponsibility.com/tools/srf-stress/): Accepts a plain-language AI deployment scenario and stress-tests it against SRF accountability rules.
- [Vendor Risk Assessment](https://aisharedresponsibility.com/tools/vendor-risk/): Seven AI supplier categories mapped to SRF layers for procurement and third-party risk teams (core platform/infrastructure, pre-trained model providers, AI-enabled SaaS, agentic AI, data/labeling/retrieval, consulting, software/VARs). Each category states vendor-versus-customer accountability, attestation baseline (SOC 2, ISO 27001, ISO 42001, ISO 9001, FedRAMP), and the evidence to demand. Includes a risk-tiering model, cross-cutting AI risks (prompt injection, model supply chain, agent authorization), and an attestation reference.
- [Vendor Risk Assessment How-To](https://aisharedresponsibility.com/tools/vendor-risk/how-to/): Five-step workflow (tier, place on SRF, split accountability, demand evidence, contract the boundary) plus a downloadable XLSX VRA workpaper with a question bank for all seven categories.
- [Incident Response Playbooks](https://aisharedresponsibility.com/tools/ir-playbooks/): Day-2 response for when an accountability boundary fails. Seven scenario playbooks (prompt injection and data exfiltration, model poisoning, agent misuse, data leakage, harmful output, supply-chain breach, silent model change), each mapped to the SRF layer that broke, with who leads per operating model, what to demand from the vendor, and what evidence to preserve. Cites the Red Team Scoping Tool and the finding routing reference as upstream inputs. Machine-readable at /export/playbooks.json.
- [Red Team Scoping Tool](https://aisharedresponsibility.com/tools/redteam-scope/): Scopes an AI red team engagement with the SRF accountability overlay. Resolves testable, authorization-required, and out-of-scope layers by operating model and autonomy tier, pre-assigns one named finding owner per layer, surfaces related threats from the crosswalk, and produces a signable scoping record with PDF and JSON export. Supplies only the accountability overlay; methodology is the OWASP GenAI Red Teaming Guide, attack taxonomy is NIST AI 100-2.
- [System Instructions](https://aisharedresponsibility.com/tools/prompts/): Copy-ready prompts for grounding AI assistants in the SRF. Core instruction enforcing one accountable party per activity, role variants (executive, auditor, developer, legal/procurement), and sector context parameters for all six verticals. Also includes a site-aware primer for querying this site's content with an AI assistant.
- [Lifecycle Prompt Builder](https://aisharedresponsibility.com/tools/prompts/lifecycle/): Builds one grounded prompt from a persona, an industry vertical, one of the six AI security lifecycle stages, and an operating model. Names the single accountable party, resolves a shared matrix cell to one owner, cites canonical IDs, and points the agent at the data file and on-site tool for that stage.
- [AI-Enabled System Threat Modeling](https://aisharedresponsibility.com/tools/prompts/threat-model/): Pack v3.0 reads a system representation and writes a threat matrix, markdown report, completed JSON, CSV, and Mermaid diagram. Operator fields belong in the first message; the chain does not pause for later input. Track A records review context, typed traditional analysis, PHANTOM-B, AI-to-traditional composition paths, pinned source references, evidence-backed importance, actions, and QA. Optional Track B assigns SRF accountability and audits L1 to L5 coverage. Optional Track C joins vertical obligations, candidate controls, and acceptance authority.
- [Threat-model prompt pack releases](https://aisharedresponsibility.com/tools/prompts/threat-model/releases/): Version 3.0 breaking workflow and schema changes, v2.0 history, and compatibility rules.
- [Diagram threat-model eval report](https://aisharedresponsibility.com/eval/threat-model/): Evaluation method and ten bounded fixtures for pack v3.0. Checks phase applicability, typed STRIDE pairs, composition paths, source provenance, CVE applicability, importance, SRF layers, and vertical joins. The prior v2.0 professional-model comparison remains labeled as v2.0. No model-written v3 P-report exists yet for the gold diagrams.
- [Whitepaper Assessment](https://aisharedresponsibility.com/assess/whitepaper-assessment/): Copy-ready prompt v2.2 for reviewing an AI security whitepaper against both published principle catalogs plus live SRF accountability and regulatory data. Confirms the paper was read in full, then fetches slim catalogs at https://aisharedresponsibility.com/data/security-principles.slim.json and https://aisharedresponsibility.com/data/ai-agentic-principles.slim.json. Screens citation integrity, unpinned dependency recommendations, text aimed at the reviewing model, unsupported claims, stale references, and leftover production artifacts; maps every claim to a catalog id; checks threat attribution against the layer that owns the control point; and ranks every finding on one P1 to P4 scale with severity capped by evidence quality. Say map-only to skip the citation registries, live SRF fetch, and draft mechanics.
- [Whitepaper Assessment changelog](https://aisharedresponsibility.com/assess/whitepaper-assessment/changes/): Prompt versions 2.2, 2.1, and 2.0. What each version added or demoted, and which older reports stay valid.
- [Draft claims test](https://aisharedresponsibility.com/assess/claims-test/): Independently proposed prompt pack v1.0.5. One intake packet, then a non-interactive chain that extracts claims, collects published topic attacks from ATLAS, OWASP, BIML, and related papers, binds each claim to risk, obligation, control, and one accountable party, and replies with a markdown report that includes Google Docs or GitHub suggestion packets. Load templates from this chat (paste or attachment) or fetch https://aisharedresponsibility.com/assess/claims-test/prompts.json; citation URLs stay pinned. JSON only if asked. Schema at https://aisharedresponsibility.com/eval/claims-test/schema.json. Cousin of Whitepaper Assessment, not a merge. Not part of CoSAI SRF v1.0.

## Regulations

- [Regulations Reference](https://aisharedresponsibility.com/regulations/): AI regulations and standards mapped to SRF layers.
- [Regulation Discovery Wizard](https://aisharedresponsibility.com/regulations/discovery/): Filter regulations by industry, geography, and SRF layer. Export a custom reference set.

## Developer resources

- [Data & agents](https://aisharedresponsibility.com/developers/): Machine indexes, join conventions, and agent skills. Copy-ready prompts live under Tools.
- [SRF System Prompts](https://aisharedresponsibility.com/tools/prompts/): Canonical copy UI for system instructions, role variants, sector context, Lifecycle Prompt Builder, and threat-model packs. `/developers/prompts/` redirects here.
- [Join Conventions](https://aisharedresponsibility.com/developers/schema/): How to join /data files and ontology edges. Human landing for the JSON catalog. A fuller Schema Viewer remains planned for v2.0.
- [Agent Skills](https://aisharedresponsibility.com/developers/skills/): Installable SKILL.md packs for Claude-family agents and OpenClaw.

## Machine-readable data

All control schemas and framework definitions are available as static JSON. An agent can fetch a single file to answer detailed questions about any vertical without parsing HTML.

- [Data Index](https://aisharedresponsibility.com/data/index.json): Index of all data files with schema descriptions, record counts, and framework concept definitions. Start here for programmatic access.
- [Layers](https://aisharedresponsibility.com/data/layers.json): Five SRF layers with personas, component descriptions, and operating model responsibility assignments.
- [Personas](https://aisharedresponsibility.com/data/personas.json): Eight personas with layer assignments, ISO/IEC 22989 references, and responsibility lists. Also declares sector_specializations (for example clinical-ai-governance) that specialize a canonical persona.
- [Responsibility Matrix](https://aisharedresponsibility.com/data/matrix.json): Operating model x layer matrix. Values: customer-owned, shared, provider-managed, model-evaluation, N/A.
- [Jurisdictions](https://aisharedresponsibility.com/data/jurisdictions.json): Controlled vocabulary for the jurisdiction that issues a regulation or standard (international standards bodies, OECD soft law, EU, US federal/multistate/states, UK, China, Singapore, Canada, Japan, Australia, Korea, Brazil, India).
- [Regulations](https://aisharedresponsibility.com/data/regulations.json): AI regulations and standards mapped to SRF layers and business verticals, each with a jurisdiction id, optional mapping_key, and applicable_verticals. Includes OECD AI Principles, national reference instruments (UK/China/Singapore/Canada/Japan/Australia/Korea/Brazil/India), SEC/OCC/FINRA finance anchors, and sector instruments cited by vertical controls. A withdrawn instrument stays listed with lifecycle rescinded and a superseded_by pointer to its replacement, so a mapping written against the old text resolves forward.
- [Moral Regulatory Hierarchy](https://aisharedresponsibility.com/data/moral-regulatory-hierarchy.json): Requirement-level actor/action/outcome salience tags (0-3, independent, with rationale) for priority instruments (EU AI Act, OECD, NIST AI RMF, SR 26-2, GDPR, HIPAA, UK white paper, China GenAI measures, Canada ADM, CMMC, SEC PDA conflicts, Korea AI Basic Act, Australia AI Ethics Principles). Descriptive analytical lens; instrument rollups and ontology edges are derived.
- [Finance Controls](https://aisharedresponsibility.com/data/finance-controls.json): 40 controls: layer, persona, MRM stage, mappings, evidence thresholds.
- [Healthcare Controls](https://aisharedresponsibility.com/data/healthcare-controls.json): 40 controls: layer, persona, clinical stage, FHIR AuditEvent evidence pointers.
- [Insurance Controls](https://aisharedresponsibility.com/data/insurance-controls.json): 40 controls: layer, persona, lifecycle stage, NAIC/state regulatory mappings.
- [Public Sector Controls](https://aisharedresponsibility.com/data/public-sector-controls.json): 40 controls: layer, agency/vendor responsibility split, federal lifecycle stage, M-25-21 mappings.
- [Defense Controls](https://aisharedresponsibility.com/data/defense-controls.json): 53 controls: layer, IL4/IL5/IL6 impact level, NSS/Non-NSS tier, responsibility split.
- [Manufacturing Controls](https://aisharedresponsibility.com/data/manufacturing-controls.json): 45 controls: layer, OT applicability, EU AI Act risk class, lifecycle stage.
- [Project Tapestry Controls](https://aisharedresponsibility.com/data/tapestry-controls.json): 26 controls: layer, accountability domain, property class, disclosure tier, bears_on_commons. Maps AI Alliance Project Tapestry governance ADRs onto the SRF pattern under the Federated-Consortium operating model. Mappings cite ADR sections directly rather than a mapping_key from regulations.json, because the operative instruments are project governance documents rather than external regulations. Provenance is proposed-extension; not CoSAI-ratified.
- [Vendor Risk Categories](https://aisharedresponsibility.com/data/vendor-risk.json): Seven AI vendor categories as structured data: SRF layers, operating models, roles, vendor-vs-customer accountability split, attestation baseline, and evidence to demand per risk area. Plus a three-tier risk model, cross-cutting AI risks (including spawn, fail-closed halt, and enforcement plane), and an attestation reference.
- [Proposed SRF next-version extensions](https://aisharedresponsibility.com/data/srf-vnext-extensions.json): Nine independently proposed companion asks for a future CoSAI SRF paper, each with a v1.0 hook and site surfaces. Not part of CoSAI SRF v1.0.
- [Threat-to-Accountability Crosswalk](https://aisharedresponsibility.com/data/threats.json): Sixteen OWASP AI Exchange security-matrix threats plus fifteen OWASP GenAI Data Security (DSGAI 2026) data-plane failures, each mapped to the SRF layer where it lands and the accountable persona per operating model (AI-SaaS, AI-PaaS, Agent-PaaS, IaaS). Includes external IDs (OWASP LLM Top 10 2025, MITRE ATLAS, BIML LLM24, DSGAI) and related vertical control IDs. The SRF adds who authorizes, who owns each finding, and who signs residual risk; threat definitions belong to their external sources. This file cites DSGAI identifiers; it does not copy DSGAI mitigations.
- [Attack Inventory](https://aisharedresponsibility.com/data/attack-inventory.json): 58 published attack classes against AI models, LLM applications, and agents, including ten classical weaknesses that agents reopen and eight business-process classes (payment diversion, invoice alteration, approval gate bypass, master data change, workflow rule injection, false status reporting, partner agent hijack, deepfake approver impersonation). Each class states a failure mode and carries MITRE ATLAS v2026.08, OWASP LLM Top 10 2025, OWASP Agentic AI Threats v1.1, OWASP DSGAI 1.0, FINOS AI Governance Framework v2, MITRE ATT&CK Enterprise 19.2, and CWE 4.20 ids graded exact, closest, or analogy, plus arXiv, CVE, and ATLAS case-study ids. Every ref was checked against its source on 2026-10-01 and records the method. srf_crosswalk joins 45 classes to the threat crosswalk for layer and accountable persona; 13 have no crosswalk row yet. vertical_controls links 7 classes to vertical profile controls. Claims-test pack 1.0.9 reads it by default in C-attacks; build/export_attack_inventory.py emits topic-filtered C-attacks rows to attach instead. Schema: /data/attack-inventory.schema.json. Readable view: /framework/attack-inventory/.
- [Threat Source Registry](https://aisharedresponsibility.com/data/threat-sources.json): Source kind, authority, canonical URL, license, identifier form, supported use, default integration mode, and binding rule for SRF, CAPEC, ATT&CK, ATLAS, OWASP, CWE, CVE/NVD, CISA KEV, NIST SP 800-30, and BSI references. Each run pins used sources in source_manifest.
- [Finding Routing Reference](https://aisharedresponsibility.com/data/finding-routing.json): Twenty layer-by-operating-model entries resolving an AIVSS- or CVSS-scored finding to one accountable persona and a breach-action escalation ladder across four severity bands (critical, high, medium, low). Consumes findings pre-assigned by the crosswalk and the red team scoping tool. The SRF does not score severity. AIVSS and CVSS own that.
- [Classic Security Principles](https://aisharedresponsibility.com/data/security-principles.json): 87 security engineering principles, axioms, and invariables normalized from a 93-control OSCAL catalog covering Saltzer and Schroeder, NIST SP 800-27, ISO 27001, CIS Controls v8/v8.1, Microsoft's two law sets, and cloud architecture corollaries. Each entry carries a stable id, category, source citations, and related links. The settled-engineering baseline for the whitepaper assessment workflow. Slim projection for mapping: /data/security-principles.slim.json.
- [AI and Agentic Security Principles](https://aisharedresponsibility.com/data/ai-agentic-principles.json): 130 principles normalized from the 52-source agentic security synthesis across 22 topic sections, with the source documents asserting each claim. Carries a gap_index resolving each section's Gap note to the classic principle it points at, so a known field-wide blind spot is distinguishable from a defect in one paper. Current industry positions, not settled axioms. Slim projection for mapping: /data/ai-agentic-principles.slim.json.
- [Classic Security Principles (slim)](https://aisharedresponsibility.com/data/security-principles.slim.json): Mapping projection of the classic catalog: id, statement, category, related. The whitepaper assessment prompt fetches this file after intake.
- [AI and Agentic Security Principles (slim)](https://aisharedresponsibility.com/data/ai-agentic-principles.slim.json): Mapping projection of the agentic catalog: id, section, category, statement, plus gap_index. The whitepaper assessment prompt fetches this file after intake.
- [Principle Candidates](https://aisharedresponsibility.com/data/principle-candidates.json): PROPOSALS, NOT ACCEPTED PRINCIPLES. Maintainer review queue. Whitepaper assessment prompt v2.2 no longer asks users to file candidates. Never cite an entry here as an established principle, and never use this file as a source when assessing a document.

## Knowledge graph and APIs

A canonical concept layer for retrieval and reasoning. Every concept has one stable ID and one canonical URL. Glossary terms are deduplicated: each resolves to a single ontology node. Generated from the source content by build/generate_knowledge_layer.py.

- [Glossary Registry](https://aisharedresponsibility.com/glossary.json): Full machine-readable glossary. Every SRF term with its definition, canonical ID, stable anchor URL, and per-term API link.
- [Glossary API Index](https://aisharedresponsibility.com/api/glossary/index.json): Index of all glossary terms. Each term is independently retrievable at /api/glossary/{anchor}.json (for example /api/glossary/accountability.json, /api/glossary/L1.json, /api/glossary/AI-SaaS.json).
- [Ontology Nodes](https://aisharedresponsibility.com/ontology/nodes.json): Concept graph nodes. Types: concept, framework, role, control. Covers the five layers, four operating models, eight personas plus sector specializations, jurisdictions, glossary vocabulary, all 260 vertical controls, and mapped external standards.
- [Ontology Edges](https://aisharedresponsibility.com/ontology/edges.json): Typed, directed relationships between nodes. Prefer this file for multi-hop joins. Relations include has_persona, operates_at_layer, assigns_responsibility, applies_to_layer, accountable_to, applies_in_operating_model, maps_to_layer, governed_by (control to regulation), specializes (sector persona to canonical persona), issued_in_jurisdiction, subordinate_to, part_of (requirement to instrument), emphasizes (requirement to moral dimension, with salience), implements (control to requirement, scoped to the instrument the citation is filed under so it always accompanies a governed_by edge), applies_to_vertical (regulation to industry vertical), belongs_to_vertical (control to vertical), superseded_by (rescinded instrument to its replacement), and has_vertical.
- [Canonical ID Registry](https://aisharedresponsibility.com/ids.json): Every concept, layer, role, operating model, jurisdiction, moral dimension, requirement, and control with one stable namespaced ID, name, and URL. Includes a glossary anchor cross-reference. Control IDs are srf.control.<vertical>.<id> because the short SRF-L*-* form repeats across verticals.
- [Join Conventions](https://aisharedresponsibility.com/developers/schema/): How to join /data files and ontology edges without inventing a second edges file. Also summarized under join_hints in /data/index.json.
- [Knowledge Pack](https://aisharedresponsibility.com/export/framework.json): Single flattened, linked representation of the whole framework for bulk ingestion: concepts, relationships, and definitions. Companion exports: /export/glossary.json and /export/ontology.json.
- [Accountability Matrix (CSV)](https://aisharedresponsibility.com/export/matrix.csv): The SRF responsibility matrix as a flat CSV, one row per layer (L1-L5) by operating model, with customer and provider responsibility. For spreadsheets and GRC import.
- [Accountability Matrix (OSCAL)](https://aisharedresponsibility.com/export/srf-oscal-catalog.json): The same matrix as an OSCAL 1.1.2 catalog. Groups are SRF layers; controls are layer-by-operating-model accountability cells. For OSCAL-aware GRC platforms.
- [Industry Vertical Controls (OSCAL)](https://aisharedresponsibility.com/export/srf-oscal-verticals-catalog.json): OSCAL 1.2.2 catalog of all 260 industry vertical controls. Per-vertical profiles: /export/srf-finance.profile.json, /export/srf-public-sector.profile.json, /export/srf-healthcare.profile.json, /export/srf-insurance.profile.json, /export/srf-defense.profile.json, /export/srf-manufacturing.profile.json. Proposed extensions; not part of the official CoSAI release.
- [OpenCRE Binding (JSON)](https://aisharedresponsibility.com/export/opencre-srf.json): SRF canonical IDs as OpenCRE-conformant Standard documents: the framework, five layers, eight personas, four operating models, and all twenty accountability matrix relations, with three proposed CRE anchors verified against the live OpenCRE graph. Lets the MOSAIC shared taxonomy resolve any threat or control node to an accountable persona.
- [OpenCRE Binding (CSV)](https://aisharedresponsibility.com/export/opencre-srf.csv): The same resource set in the MyOpenCRE import format, ready for a local OpenCRE instance.
- [Incident Response Playbooks (JSON)](https://aisharedresponsibility.com/export/playbooks.json): Seven AI incident response playbooks as structured data, with SRF layer mapping, who-leads-by-operating-model, triage steps, vendor asks, and evidence to preserve.
- [Retrieval Validation Tool](https://aisharedresponsibility.com/llm/test/): Simulate RAG ingestion against the knowledge pack. Returns matched chunks, concept hits, and a confidence score. Append ?q=QUERY&format=json for a JSON response.

## NIST standards demo (separate corpus)

- [NIST AI RMF RAG Demo](https://aisharedresponsibility.com/nist-ai-rmf/llms.txt): Separate dual-readable extracts of NIST AI RMF 1.0 and the Generative AI Profile. Agents: start at that llms.txt; for ranked JSON use `/nist-ai-rmf/retrieve/*.json` or fetch `/nist-ai-rmf/data/chunks.json` (do not treat `?format=json` as an agent API). Not official NIST output and not part of the SRF knowledge graph. Do not confuse with the TACIP orphan at /nist/.

## Full content for agent retrieval

- [llms-full.txt](https://aisharedresponsibility.com/llms-full.txt): Complete inline content: all framework concepts, all six vertical control schemas, and regulatory context in a single document. Agents can retrieve this file to answer detailed SRF questions without following additional links.

## About

- [About](https://aisharedresponsibility.com/about/): Background on the project: an independent companion site to the CoSAI SRF built on open-source tools, proposed for CoSAI donation.
- [Changelog](https://aisharedresponsibility.com/changelog/): Dated record of site changes, grouped by release and tied to CoSAI SRF v1.0 (May 28, 2026) and v2.0 work in progress.