untactit Audit: AI-Agent Asset Ledger Claimed 763 Items, Hand Count Found About 400

TOKYO, Japan - Aug. 18, 2026 - PRLog -- untactit, a pre-launch control plane for the skills, rules, and memory AI agents run on, has published a measured audit report on how AI-agent asset ledgers drift from reality. The report is available in five languages on the company's engineering blog.

The audit compared a working asset database of 763 records -- conversations, projects, skills, instruction files, and memories collected from 21 AI platforms including Claude, ChatGPT, Gemini, Grok, Perplexity, Kimi, and Qwen -- against what actually exists on each platform. The method: calling each platform's own list API from an authenticated page, paginating to exhaustion, and comparing ID sets in both directions rather than trusting counts. The result: roughly 400 of the 763 recorded assets were real.

The report identifies three independent failure modes, each invisible on a dashboard. Ghost entries, deleted on the platform but still alive in the database, accounted for at least 173 rows. Missed items, real assets that were never collected because pagination stopped early while the interface showed success, accounted for at least 63. False positives, platform-provided default content and settings-page placeholder text recorded as user assets, accounted for at least 109. Because ghosts and false positives inflate the ledger while misses deflate it, the errors partially cancel, making the total count the one metric guaranteed to hide all three.

Notable individual findings include a platform skill list where an API page-size cap of 100 was recorded as the account's actual count (the user had installed 4), and a memory collector that missed all 8 real saved memories while storing two lines of example text from the settings screen as user data.

The report's conclusion is operational rather than tool-specific: write success says nothing about read truth, and the fix is a read-back audit as a first-class operation -- after any sync, re-read the target platform and diff IDs in both directions, on a schedule.

The full report is published in English, Japanese, Korean, Simplified Chinese, and Traditional Chinese:
https://untactit.com/the-ledger-said-763

About untactit
untactit is a control plane for the skills, rules, and memory AI agents run on: inventory, approval, deployment, and audit in one workspace. The product is currently pre-launch with a public waitlist at https://untactit.com

Contact
untactit
Email: ir@untactit.com
Web: https://untactit.com