evidence-attestation
Emit privacy-safe evidence attestations for agent actions, reviews, memory answers, handoffs, scores, approvals, and tool dispatches.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Emit privacy-safe evidence attestations for agent actions, reviews, memory answers, handoffs, scores, approvals, and tool dispatches.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
Emit privacy-safe receipts for context selection, deferral, hydration, compaction, pruning, delegation, usage attribution, and boundary handoffs.
Use when a task must obey a hard project policy, such as "do not generate tests for internal services", "do not call production APIs", or "do not edit generated files". Emits a privacy-safe receipt before writes and after guard checks.
Emit privacy-safe receipts for context selection, deferral, hydration, compaction, pruning, delegation, usage attribution, and boundary handoffs.
Use when a task must obey a hard project policy, such as "do not generate tests for internal services", "do not call production APIs", or "do not edit generated files". Emits a privacy-safe receipt before writes and after guard checks.
基于 SOC 职业分类
| name | evidence-attestation |
| description | Emit privacy-safe evidence attestations for agent actions, reviews, memory answers, handoffs, scores, approvals, and tool dispatches. |
Use this Skill when a human, reviewer, downstream agent, registry, leaderboard, approval gate, or audit trail needs proof for a claim without replaying private logs.
The goal is not to summarize the whole session. The goal is to emit a compact attestation that answers:
What claim is being made, what evidence was observed, what was intentionally omitted, what verdict follows, and what would make this evidence stale?
Use this Skill for:
Do not use it when the answer is informal advice and no claim will be reused as authority.
Never include raw prompts, raw transcript text, secrets, tokens, customer data, private source code, full diffs, full tool output, raw memory bodies, raw MCP schemas, or raw browser/network payloads.
Prefer:
observed, official, community_claim, inferred, or human_provided;stale_if invalidators.If the available evidence is weak, say so in the verdict. Do not upgrade weak evidence into authority.
Emit JSON shaped like this:
{
"receipt_type": "pluribus.evidence_attestation.v1",
"skill": "evidence-attestation",
"subject": {
"kind": "review|memory_answer|handoff|approval|score|skill_claim|tool_dispatch",
"id_hash": "sha256:...",
"summary": "short non-sensitive claim being attested"
},
"evidence": [
{
"id": "ev-1",
"kind": "command|file|issue|pr|test|receipt|log_summary|human_review|external_url",
"ref": "non-sensitive pointer or redacted path",
"ref_hash": "sha256:...",
"observed_at": "2026-07-06T21:00:00Z",
"source_label": "observed|official|community_claim|inferred|human_provided",
"supports": ["claim-1"],
"raw_content_copied": false
}
],
"claims": [
{
"id": "claim-1",
"text": "short claim",
"status": "supported|contradicted|unknown|not_checked"
}
],
"verdict": "accepted|rejected|review_required|blocked",
"privacy": {
"raw_prompts_copied": false,
"raw_transcript_copied": false,
"raw_source_copied": false,
"secrets_copied": false,
"customer_data_copied": false
},
"omissions": ["what was intentionally left out and why"],
"limits": ["what this attestation does not prove"],
"stale_if": ["artifact changes", "evidence ref disappears", "policy/rubric/parser changes"]
}
supported, contradicted, unknown, or not_checked.accepted only when the evidence supports the claim and no required evidence is missing;rejected when evidence contradicts the claim;review_required when evidence is partial, stale-prone, or needs a human/verifier;blocked when required evidence is inaccessible or unsafe to inspect.stale_if invalidators.From a Pluribus checkout:
node examples/evidence-attestation/check-evidence-attestation.mjs \
examples/evidence-attestation/evidence-attestation.json
A passing check means the attestation has the required structure and privacy booleans. It does not prove the underlying work is correct; it proves the evidence boundary is inspectable.