Evidence attestation

Skill caioribeiroclw-pixel/pluribus/skills/evidence-attestation

Sync one reviewed context source into native AI-agent files, with privacy-safe evidence receipts that never confuse generation with runtime load.

Install

npx -y skills add caioribeiroclw-pixel/pluribus --skill evidence-attestation

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

11 stars11 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Emit privacy-safe evidence attestations for agent actions, reviews, memory answers, handoffs, scores, approvals, and tool dispatches.

SKILL.md

4.3 KB, as published. Nobody here has run it

Evidence Attestation

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?

When to use

Use this Skill for:

code review or release claims that cite tests, files, issue trails, PRs, or audit agents;
memory/RAG answers that might become edit authority;
handoffs between Claude Code, Codex, Cursor, OpenClaw, humans, or subagents;
approval gates before deploys, billing changes, customer-facing messages, or destructive tools;
benchmark/leaderboard score submissions that must not leak transcripts;
skill/registry claims where a verifier needs cited evidence.

Do not use it when the answer is informal advice and no claim will be reused as authority.

Privacy defaults

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:

stable ids, short summaries, and hashes;
source labels such as observed, official, community_claim, inferred, or human_provided;
evidence refs that a privileged reviewer can resolve, such as issue URLs, commit ids, command names, fixture paths, or artifact hashes;
explicit omissions and limits;
stale_if invalidators.

If the available evidence is weak, say so in the verdict. Do not upgrade weak evidence into authority.

Minimal attestation shape

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"]
}

Procedure

Name the claim or boundary in one sentence.
Identify the smallest evidence refs needed to support or reject it.
Replace private content with hashes, ids, buckets, source labels, or short non-sensitive summaries.
Mark each claim supported, contradicted, unknown, or not_checked.
Choose the verdict:
- 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.
Add omissions, limits, and stale_if invalidators.
Validate with the bundled checker when available.

60-second validation

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.