Drafting note: this post was written with AI assistance, under the verification discipline it describes. The claims — and responsibility for them — are mine.
Epistemic status: a practitioner's argument, not a benchmark. I would value being shown wrong on the empirical claims below.
In my first post here I described a register of how AI-assisted legal work goes wrong — defects in how one checks, not in what the machine says. This is a follow-up, on a question the legal-tech industry is now asking loudly: what becomes legal's "system of record"?
The moment is real. Over recent months OpenAI, Anthropic, and Microsoft moved into dedicated legal offerings, and on 25 August 2026 Google Cloud launched Gemini Enterprise for Legal. The common pitch is a governed container: the trusted place where AI-assisted legal work will live.
I want to correct something I have said before. It is no longer true that the vendors are silent on what must fill the container. Google's launch foregrounds exactly the right words — grounding in primary legal authority, citation verification, audit logging. So the gap is not silence. The gap is subtler, and it is the whole point.
A vendor can assert that an output was verified. That is not the same as the output carrying a trace that I — the human who signs it — can re-test myself. The first is a feature I am asked to trust. The second is a property I can exercise. A system of record is only as trustworthy as the second.
Against any container, theirs or mine, I hold the work to five axes. I keep them in a file I call KEYCHAIN:
I examine KEYCHAIN against the AI-memory landscape, and lay out these five axes, in a short paper: Examining the Silence of the Memory Landscape (SSRN).
None of this is exotic. It is ordinary scholarly and forensic hygiene. What is new is only the need to hold machine-assisted work to it without exception.
In that paper I read the memory-product landscape as a dated census — where it stood in June 2026 — and, against each product's own description, none answered all five for the user rather than the vendor. I claim only stated absence, not that they could never build these. The field moves fast, so anyone reading should re-check the primary sources. They store; they do not chain to an origin I can walk back. They retrieve; they do not anchor a claim to a source I can open.
Why raise this here. Law is a useful stress test for AI-assisted knowledge work: the stakes are high, the checking is adversarial, and primary-source discipline is already native to the craft. But the underlying property — whether a human can examine, re-test, and stand behind machine-assisted work — is not specific to law. It reads to me as a general condition for trusting AI in any high-stakes domain, and closer to a governance question than a product feature.
So, two asks. First, correction: if a tool does answer the five axes for the user, tell me — and tell me how you tested it. Second, contact from anyone working on provenance, verifiability, or evaluability of AI-assisted work in a domain other than law. I want to know whether the five hold, break, or need a sixth once they leave my field.
— Tandry Laksana◆, Justa Causa Law Firm, Jakarta
Sine vestigio nulla probatio — without a trace, nothing can be tested.
