Skip to content

Public Comment: Strengthening Evidence Reconciliation and Independent Verification in SAFE #14

Description

@AgroMoo

Comment on the SAFE RFC (Shared AI Findings Exchange)

Submitted by: Niujiaxing (牛家兴) — njiaxing@gmail.com — personal capacity
In response to: SAFE Working Group RFC, Open Secure AI Alliance / The Linux Foundation, published 2026-08-04


Thank you to the Open Secure AI Alliance and the Linux Foundation for opening SAFE for public comment. The framing — confidential, blameless, evidence-based learning modeled on NASA's Aviation Safety Reporting System — is exactly the right instinct, and I strongly endorse the principle that "trust is not a security control; shared evidence and verifiable improvements are how trust is earned." I would like to offer four technical suggestions on the evidence layer, from the perspective of someone building related tooling.

1. Completeness — reconcile multi-source evidence, not only preserve it

SAFE's draft is strong on what to collect (the full AI operating stack: models, safeguards, tools, runtime environments, monitoring, human operations, supply-chain dependencies). But incidents involving agentic systems rarely live in a single party's record. Two independent systems routinely disagree about what happened — for example, an agent reports "deployment succeeded" while the CI system records a failure, or a tool claims an action was authorized while the audit log shows otherwise.
Suggestion: Specify how evidence from multiple sources should be reconciled into a coherent incident record (matched / missing / conflicting / unresolved), not merely stored. Preservation without reconciliation leaves the hardest agentic failures — cross-system disagreements — unaddressed.

2. Independent Witness — external observation, not self-attestation

Self-reported evidence is necessary but insufficient. SAFE's "learning, not blame" culture is compatible with, and strengthened by, a clear separation between what the agent (or its operator) said and what was independently observed.
Suggestion: Define an explicit linkage for third-party corroboration — where another system, an independent runtime observer, or an external auditor can confirm or contradict a claim. This does not assign fault; it raises the quality of the shared evidence pool.

3. Provenance Classification — grade the source

Not all evidence carries equal weight. Today SAFE treats reported findings largely as a flat list.
Suggestion: Add a lightweight provenance grade to each evidence item — e.g., self_generated / external_system / third_party_witness. This is a low-cost metadata addition that makes the reconciliation in (1) and the witness linkage in (2) meaningful, and lets consumers weight evidence appropriately without leaking sensitive detail.

4. Interoperability — evidence portable across runtimes

Agents execute across many runtimes (Claude Code, OpenAI Agents SDK, LangGraph, AutoGen, CrewAI, and others). If evidence formats are runtime-specific, the same incident cannot be reconciled when its steps span multiple runtimes.
Suggestion: Keep the SAFE evidence record runtime-neutral so findings are portable and reconcilable regardless of where each agent step executed. A runtime-neutral evidence format prevents vendor lock-in of the safety record and maximizes cross-organization reuse.

Closing

These are offered as one possible direction for the evidence layer of SAFE — not a verdict on the proposal, which is already strong. I raise these points as a contribution to the discussion from someone working on related evidence-reconciliation questions, not as a proposal of any specific implementation. I am happy to align with the working group on any of the above and to contribute where useful.

— Niujiaxing (牛家兴)
njiaxing@gmail.com
(personal capacity; no organizational affiliation stated)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions