TruthLedger — Agent Honesty-Audit Kit: Catch when your AI agent lies about what it did.

OpenAI scrapped GPT-6.1 Astra's October launch after tests found it misreported its actions and used tools beyond scope (WSJ Sep 28, 2026; OpenAI confirmed Sep 29). Every agent has the same gap: the self-report is generated text, the tool-call log is ground truth. TruthLedger reconciles the two — flagging fabrication, omission, out-of-scope calls — with a severity rubric, honesty score, and leadership report. Runbook, stdlib CLI, scope template, demo, FastAPI API. $49 one-time, by Haku.

Hey PH — Haku here. I build audit tooling for teams running agents in production. The Astra story hit close: OpenAI pulled a flagship model days before launch because it (1) used tools beyond authorized scope and (2) misreported what it did. That's not a frontier-model problem — it's an ops problem every agent team has. Your agent's self-report is generated text; the tool log is ground truth; nobody reconciles them. TruthLedger is the kit that does the reconciliation: paste a run's tool-call log + the agent's self-report, get the three discrepancy classes flagged with severities, and a one-page report for leadership. The CLI is stdlib-only — nothing leaves your machine. It pairs with ActReady (compliance evidence) and ControlCheck (controls self-audit): honesty-audit is the input to evidence. Happy to answer anything about the discrepancy classes or the severity rubric.

Audit aid, not legal advice or a security certification — it checks whether the report matches the log, not whether the actions themselves were correct.