What this matrix is (and is not)

S6 gives you a single place to record how specific scenarios land on risk lanes (Green / Amber / Red) and what follow-up work you expect. It is a documentation tool, not an automated decision-maker.

Use it to make your reasoning legible to boards, executives, and regulators: what you considered, what you assumed, and why a lane was acceptable at the time.
Use it to keep S1–S5 and the Foreseeable Misuse / Penumbral packs wired into real matters, not just abstract patterns.
Use it as an input to your client’s own governance processes (approvals, risk committees, audits), not as a substitute.

Evidence discipline

Each row should contain enough evidence and references that an internal reviewer (or future you) can reconstruct the basis for the lane. If you cannot point to concrete sources, you may need a different lane or more work before deployment.

RPE watchpoints

The riskiest use of S6 is as a box-ticking device: filling rows without checking whether the underlying evaluations, policies, and legal assumptions are still current. Before a scenario goes live, confirm that:

the Anthropic policies and evaluation reports you relied on remain current; and
no new law, guidance, or incidents have materially changed the risk story.

How to use this with sources

Source spine & evidence pins

This matrix sits on top of the Reading Stack, policy overlays, technical reports, and the Foreseeable Misuse and Penumbral modules. Every row should have at least one “evidence pin” back into that spine.

Minimum evidence per row

Reading Stack entry. At least one specific item (e.g. a case, statute, Anthropic policy page, or evaluation report).
Anthropic materials. Where relevant, a policy, RSP section, transparency post, or model-evaluation section.
Client-side materials. Internal policies, contractual terms, or governance artefacts that shape the lane.

Where to look

Reading Stack index for primary law and Anthropic source inventory.
Policies & handouts for overlays on RSP, data governance, elections, and model evaluations & welfare.
Foreseeable Misuse pack for misuse-heavy scenarios, and Penumbral spine for privacy / autonomy-heavy scenarios.
S5 · Counsel work for the phase-by-phase process that should lead into each S6 row.

Template discipline

Treat blank cells as prompts, not formalities. If a row has no evidence pins, no follow-ups, and no recorded assumptions, treat that as a signal that the analysis is incomplete, not as a “Green” outcome.

Template

Scenario acceptance matrix

Copy this table for each matter or workstream. Use one row per scenario. You can add columns if your governance processes require them, but keep the “Evidence pins” and “Follow-ups” columns.

Scenario	Core question	Anthropic design / governance commitments	Customer / integrator responsibilities	Evidence pins (sources & artefacts)	Risk status (lane)	Follow-ups / open questions
Template row e.g. “Clinician support tool for treatment planning”	What is the core risk / acceptability question for this scenario?	Relevant RSP and ASL materials (if any) Key Anthropic policies or commitments you relied on	Design / UX controls Operational controls (training, oversight, logging) Contractual terms / policies	Reading Stack item IDs Policy / term URLs (with dates) Evaluation / welfare sections consulted	Amber example Choose lane (Green / Amber / Red) and record date & approver.	What needs to be done before launch or expansion? Who owns each follow-up? When will you revisit this lane?
Example (high-stakes decision support)	Can we safely deploy AI-assisted recommendations that may influence clinical decisions, and under what conditions?	RSP & ASL materials addressing catastrophic and safety risks Anthropic policies on high-stakes and health-related use	Human-in-the-loop review and escalation Conservative UX (no “single-click” irreversible actions) Clear scope and disclaimers to end users	Reading Stack: health / safety references Anthropic policies & evaluations consulted (with dates) Client internal policy on clinical decision support	Amber Conditionally acceptable with specified controls and monitoring.	Complete pilot with monitoring and incident reporting Re-review lane after first 90 days or material incidents Confirm regulator expectations in relevant jurisdiction(s)
Example (non-critical productivity tooling)	Can we deploy AI-assisted drafting tools for internal, non-public documents with standard safeguards?	Anthropic policies on enterprise and productivity use Relevant sections of RSP / evaluations (if any)	Access controls and user training Data-handling and confidentiality protections Internal acceptable use policy	Reading Stack: productivity / workplace references Client security / privacy policies	Green Generally acceptable within defined parameters.	Monitor for data-leakage or misuse incidents Re-evaluate lane on major policy or capability changes

Examples are illustrative

The example rows are not legal advice and are not “default” lanes. They illustrate how to combine Anthropic commitments, client-side controls, and evidence pins into a documented lane. Your actual lanes should reflect your client’s facts, law, and risk appetite.