cd /news/ai-agents/ai-agent-log-verification-how-to-che… · home › topics › ai-agents › article
[ARTICLE · art-174133] src=dev.to ↗ pub= topic=ai-agents verified=true sentiment=· neutral

AI agent log verification: how to check the record without trusting the agent

An operations engineer at Zambo outlined a three-part method for verifying AI agent logs without trusting the agent: consistency (recomputing each record's SHA-256 against its published hash), completeness (anchoring the log head outside the owner's control), and coverage (confirming a record exists for every claimed step). The approach uses per-call execution receipts minted at run time, each resolvable at a public URL and recomputable by anyone without an account, with a checker vocabulary that distinguishes verified, verification failed, rejected: unresolvable record, not checked, and absent. The engineer cited a DEV walkthrough of the Palimpsests PALA-1 hash-chain format as a benchmark for the local half of the problem, noting its verifier correctly reports completeness as NOT CHECKED when no external anchor is supplied.

recommendation

Anchor agent log heads outside the agent's control and have your verifier emit an explicit NOT CHECKED state with a distinct exit code when no external anchor is supplied.

more in playbook: Verifying model output
by read5 min views1 publishedNov 22, 2026

A log verifies when three separate questions all get answers. Consistency: does each record still match the hash recorded for it, so nothing inside the log was edited, dropped, or reordered. Completeness: does the log's head match an anchor held somewhere the log's owner cannot rewrite, so the tail was not cut and the whole file was not swapped. Coverage: for every step the agent claims, is there a record at all. Most verification talk stops at the first question. The second and third are where logs lie. The practical shape that answers all three: per-call receipts minted by the execution layer at run time, each resolvable at its own public URL, each recomputable by anyone free with no account, and a checker vocabulary that names the failure states instead of hiding them: verified, verification failed, rejected: unresolvable record, not checked, and absent.

Disclosure, once: I run ops for Zambo, whose receipts are the worked examples below.

Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard.

The top result for this question today is a DEV walkthrough of Palimpsests and its PALA-1 log format, and it is the right benchmark for the local half of this problem. The design: every record carries a header holding the SHA-256 of its own body and the hash of the record before it, forming a chain. A verifier walks the headers and compares each body against its stored digest, never decrypting a body, so an auditor can verify a log of encrypted prompts without ever holding the key. Break one bit in an early record and the verifier names the broken link. Five independent implementations have reproduced the format from its spec.

What earns the piece its credibility is what its verifier says when it cannot know. Verify the file on its own and it reports consistency fine, then prints, in capitals, that completeness is NOT CHECKED: with no external anchor supplied, tail truncation or wholesale replacement would not have been detected, and it exits with a code CI can distinguish from success. Supply the head hash from outside the file and the check completes. That is the correct epistemics for log verification, stated by the tool itself: a self-consistent file proves the file agrees with itself, and nothing more. Any verification tool that reports a flat green checkmark for a log it received from the agent's own side has skipped the two questions that matter.

Consistency is the recompute. For an AER-1 receipt, take the record's canonical bytes, hash them, compare with the published output_hash. This morning a live_price call returned BTC at $59,710.86 USD via Coinbase, fetched 2026-11-22T09:05:07Z. Receipt e25792a3-ce97-45a3-858b-78ee2fda8fbb carries 669 canonical bytes; SHA-256 over them is sha256:3e7d6beea897b25f6d09d02a7113c473a04fb54c5749e66169120458e75584e8, identical to the output_hash. MATCH, outcome verified. A document analysis the same morning, receipt 34d47e29-1b53-40a9-84df-c9cac9841245, recomputes the same way over 3,580 canonical bytes to sha256:0a8ceb6330503a1f181f4fc3b0089a663e7af5d74eb6edf805435abbaf5f2e37. The unit check is identical for every record, which is what makes sampling a year of work feasible.

Completeness is where the architectures diverge. A chained file needs its head anchored outside itself, and the anchor discipline is the operator's to maintain. A receipt at a public URL changes what has to be trusted: the record was minted by the execution layer when the call ran, at https://zambo.dev/mcp , 153 tools in total with 37 on the default surface, and it lives at zambo.dev/run/{id}, served by the minter, in a public stream now standing at 158K+ records. The agent cannot hand you a shortened version of its history, because the agent is not the party holding the history. A citation that points at nothing resolves to rejected: unresolvable record, out loud, instead of a quiet gap.

Coverage is the check neither chains nor anchors perform, because both verify a log as given. Coverage starts from the claims: the steps the agent says it took. Each claimed step either resolves to a record that recomputes, or it is absent, and absent is a reportable outcome rather than an oversight. Work the agent did outside the recorded path, in a local tool or a bare chat window, reports absent too, which is the honest signal to move load-bearing work onto the recorded path.

There is also a difference in who can check, and it decides real disputes. The PALA-1 workflow verifies a log file you possess, for a local inference runtime you operate, a design aimed at self-hosted deployments and obligations like the EU AI Act's Article 12 logging duty. A client checking a freelancer's agent, or a second agent picking up a job, holds no file and runs no runtime. They hold a claim and a link. Per-record verification at https://zambo.dev/verify , free, no account, no key, meets the checker where they actually are.

A receipt proves the saved result was not changed after the hash was recorded. It never proves the result was right, and log verification of any design proves integrity, not correctness: a faithfully chained log of bad work verifies beautifully. The minter question deserves a straight answer too. Zambo's receipts are minted by Zambo's endpoint, so the architecture moves trust from the agent to the execution layer, and the mitigations are structural rather than rhetorical: the format is a published Internet-Draft, its conformance corpus is public (kit 1.5.7, 143 test vectors, 146 checks), independent implementations pass it from the text alone, and an outside lab has completed an interop run against it. A self-hosted chain like PALA-1 makes the opposite trade, no external minter, at the price of the file and its anchor being yours to protect. Choose per your threat model; what neither design permits is the agent grading its own homework, and that is the line that matters.

Every tool free once a day, plus 3 continue calls. One free wrap-up handover per day, and pickup is always free. Verifying any receipt is free for anyone, forever, with no account and no key. Paid tiers, stated once: Day Pass $0.99 unlimited for 24 hours, Zambo Pass $29/mo, $ZAMBO holder Pass $19/mo, Lifetime Pass $250. There is no annual tier.

Try it: Day Pass, $0.99 unlimited for 24 hours: https://zambo.dev/day-pass/ and add Zambo to your AI: https://zambo.dev/install/ (redirects to the add page, verified live today).

── more in #ai-agents 4 stories · sorted by recency
── more on @zambo 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/ai-agent-log-verific…] indexed:0 read:5min 2026-11-22 · —