Governance frameworks now require agentic AI systems to prove what they did, but none of them tells a builder how. This paper systematizes the evidence an accountable agent must produce into five principles—Provenance, Contemporaneity, Reproducibility, Disclosure, and Independence—derived jointly from the Singapore IMDA Model AI Governance Framework for Agentic AI, the EU AI Act, the NIST AI Risk Management Framework, and ISO/IEC 42001. It contributes a multi-framework crosswalk mapping each principle to all four regimes at once, isolating framework churn from a builder's core implementation; a measured, minimal, from-scratch reference implementation (the Incident Ledger, an append-only, hash-chained SQLite evidence log) costing a mean of 0.47 ms per append, 4.0 MB of storage per 1,000 logged inferences, and a hash-chain overhead under 0.1 ms against a plain-insert baseline (within run-to-run variance); and an evidentiary maturity lens, the Accountability Self-Check, grading a system's evidence D0–D4 against what would actually survive an adversarial challenge. Tamper-evident logging, deterministic replay, and evidence-maturity grading are cited prior art, not claims of this paper; its contribution is the systematization across a fragmented literature, the crosswalk, the measurement, and the evidentiary lens, demonstrated end to end on a constructed case study evolved from an unaccountable to an independently verifiable refund-issuing agent. A central finding follows from putting the cost next to the maturity ladder: for the deployment class measured here — a single-writer inference log committing durably on every append — the cost of accountability inverts the usual intuition. The technical levels are the cheap ones — tamper-evidence adds well under a tenth of a millisecond per action — while the expensive step is organizational, the standing independent verification that no amount of engineering alone can supply. That last step is one this paper has not itself taken: every claim here is self-assessed by a single author, which by the paper's own scale makes it a D3 artifact.
Abhijeet Verma (Fri,) studied this question.