Verify a decision
Every moderation decision on AVOID.NET is anchored to the Solana blockchain. You don't have to trust us — you can verify cryptographically that we committed to a verdict at a specific moment and have not rewritten it.
How verification works
- We commit. When a moderator accepts/rejects a submission, we serialize the decision into deterministic UTF-8 bytes (
payload_canonical_string), hash it with SHA-256, encode the digest as base58, and write it to Solana inside an SPL Memo v2 transaction. - We store the bytes. The exact bytes we hashed are stored alongside the decision in our database. Anyone can read them and recompute the hash in any language.
- You compare three values. Database hash, your independently-recomputed hash, and the hash inside the on-chain memo. If all three match, the decision is authentic and timestamped.
The on-chain memo format is
AVOID.NET|v1|h:<b58-sha256>|d:<id>|t:<iso>Find a signature on any investigation page's decision log, or run python -m src.verify_decision --signature <sig> for a CLI check.
Decision
- Sequence
- #2
- Score
- 12 → 12 (0)
- Cluster
- mainnet-beta
- Slot
- 443523178
- Off-chain at
- 2026-08-27T03:32:00.362Z
- Anchored at
- —
- Block time
- —
Independent verification
- 1. Database (off-chain)
- 9A4BDcvNL7uso1g4HJp2cxjbjUyqapQdVZJ1ZUZRVD9B
- 2. Recomputed (your browser)
- computing…
- 3. On-chain (Solana memo)
- fetching…
Canonical bytes hashed (1404 chars)
{"actor":"reviewer","decided_at":"2026-08-27T03:32:00.244Z","decision":"review","investigation_id":"d049e2b5-c794-4872-b556-612b336b3518","new_score":12,"page_slug":"openai-rogue-agent-hugging-face-breach-july-2026","prev_score":12,"reason":"This page's 'rogue agent' framing is directly and repeatedly corroborated by OpenAI's own official statements, Hugging Face's own technical post-mortem, Anthropic's own disclosure, and independent evaluators (METR, Redwood Research, Apollo Research) -- unlike the sibling-page defect this review specifically tested for, no citation here contradicts the claim it is attached to, and no credential-compromise-mislabeled-as-agentic pattern was found. Nearly all technical, financial, legislative, and quoted-statement claims were independently verified against primary sources and check out precisely, including unusually granular figures (17,600 actions, 141,006 evaluation runs, $2M/$20M daily penalties). The main defect is timeliness: OpenAI released a more authoritative official technical report with independent METR/Redwood Research analysis on August 26, 2026, one day before this review, which refines (without contradicting) the page's account of agent autonomy and motive; the page has not yet incorporated it.","score_delta":0,"sequence_num":2,"submission_content_hash":null,"submission_id":null,"submission_kind":null,"submission_valence":null,"v":1}