rfc: add coordination, legal carve-outs, and near-miss tiering

[alvin-chang]

Summary

This PR adds three sections to rfc-safe-proposal.md addressing critical gaps in the Reporting Compact that make it unenforceable in its current form.

Additive only. +80 lines, -0 lines, no rewrites. All existing content (Reporting Compact, Notification Timelines, Evidence Preservation, 8-layer Review Framework, Disclosure Model, Lessons -> Controls) preserved verbatim.

What's added

1. Coordination with Existing Channels

A new section making explicit that SAFE defers to, complements, and cross-references existing disclosure channels:

• CVE / CNA - SAFE encourages CVE publication; "public safety report" may reference the CVE
• FIRST - alignment with CSIRT classifications, CVSS/EPSS
• CERT/CC and national CSIRTs - preserved for law-enforcement and cross-border cases
• AIID (Partnership on AI) - cross-referenced for public safety reports
• OECD AI Observatory - aggregate anonymized findings may be shared
• CVD best practice - SAFE timeline operates within, not in place of, CVD
• Regulatory obligations - explicit pointer to Legal carve-outs section

The Reporting Compact's intake form should ask whether the incident has been reported elsewhere.

2. Legal & Sovereignty Carve-outs

Seven numbered carve-outs applying to every SAFE timeline, trigger, and reporting requirement:

1. Law overrides compact - GDPR Art. 33 (72h), SEC 4-business-day, NIS2 (24h/72h), HIPAA, PCI-DSS, state breach laws
2. Contract overrides compact - customer/supplier notification obligations
3. Privilege and confidentiality - attorney-client, work-product, trade secrets
4. Active investigation - law-enforcement/regulator/internal forensic delay
5. Cross-border data transfer - GDPR/UK GDPR/PIPL/LGPD/APPI mechanisms
6. No waiver of rights - membership does not waive sovereign or statutory rights
7. Liability bounded - SAFE is a peer-learning network, not a regulator

These are explicitly stated to be "part of the compact, not exceptions to it."

3. Near-Miss Definition & Severity Tiers

Defines "near miss" with both inclusion criteria and exclusion list. Tiers into:

• T1 - near miss of T1 trigger (sandbox escape, unauthorized third-party access, mass data exposure). Mandatory report (30-day preliminary analysis).
• T2 - near miss of T2 trigger (sensitive data access without exfil, persistent probing stopped). Mandatory report (quarterly aggregate).
• T3 - low-severity policy violation, classifier-blocked output. Optional.
• T4 - pre-tool-call prompt rejection. Not reportable.

Includes "aggregate disclosure" sub-section: SAFE publishes quarterly aggregate of T1+T2 near-miss counts by category, member-identifying info stripped.

Also adds three corresponding bullets to the Guiding Principles section linking to these new sections.

Why this matters

In its current form, the Reporting Compact imposes contractually binding notification timelines that conflict with GDPR Art. 33, SEC cyber disclosure rules, and NIS2 - making SAFE membership potentially illegal for any member subject to those laws. No member (Cisco, CrowdStrike, Hugging Face, NVIDIA, Red Hat) would sign a compact that puts them out of compliance with their regulatory obligations.

Adding these three sections does not weaken the compact; it makes it enforceable by removing the legal conflict.

Testing

No code changes. Markdown only. Visual review of rendered output recommended.

Signed-off-by

Alvin Chang 1977968+alvin-chang@users.noreply.github.com