If an AI system keeps its own signed logs, can you trust them, or do you need outside proof they weren't edited?

This explores whether a log the system signs itself gives an auditor different guarantees than a record whose fingerprint is committed to an outside party. The corpus has no note on signed logs as such, so it can't run a head-to-head. What it does have is a precise account of what external anchoring adds and what it leaves out, which is enough to show where the comparison turns. The corpus frames the choice as ordinary logging plus an anchor, not signed versus anchored. Organizations need to reconstruct what agents did, in what order, and detect edits made afterward. External anchoring is described as a tamper-evidence layer on top of essential conventional logging, not a replacement for it Can external anchoring detect tampering in agentic process logs?.

What the anchor buys is narrow. A blockchain anchor shows a record has not changed since it was committed. It cannot show when the event actually happened, whether it was captured correctly, who was authorized to anchor it, or what caused it What can a blockchain anchor actually prove about records?. Tamper evidence does not mean truth. A wrong entry that gets anchored promptly becomes a permanently, verifiably wrong entry.

The same gap shows up at the regulatory level. The anchored-evidence proposal names five governance uses and three regulatory regimes, but it supplies no mapping from legal provisions to evidence. Anchoring alone cannot supply ordering, capture authenticity, or causal traceability, which are the controls a regulator would need to confirm that human oversight really happened. So the corpus calls it reporting readiness, not compliance Does anchored evidence actually enable regulatory compliance or just readiness?. That fits the auditor's real to-do list: which agents talked, what tools ran, which approvals were obtained, which policies applied, and whether the records were modified afterward. Integrity is the last item on that list, not the whole list What must auditors reconstruct to verify agentic workflows?.

Here is where signed logs would fit, though the corpus doesn't test this. A signed log usually rests on a key the operator holds. An anchor rests on a party the operator doesn't control. That difference in who can quietly rewrite history is the open part of your question, and no note here measures it.

Two other notes show ways to make records hard to dispute without leaning on cryptography alone. BenchShield treats infrastructure records as the unchallengeable check and audit judgment as the arguable step after it. Its audit agents get a limited remit, fixed artifacts, and a duty to cite evidence, though their reliability hasn't been quantified Can infrastructure evidence replace terminal scores in benchmark validation? Can scoped agents reliably judge semantic hacks in runtime analysis?. The COLLEAGUE.SKILL work gets auditability from lifecycle instead. Expertise lives in versioned files that can be inspected, corrected, and rolled back, with no hidden prompt state Can person-grounded skills remain auditable without hidden prompt state?. Across all of these, protecting a record after it is written is the easier problem. Making sure it was captured faithfully in the first place is the harder one.