Your Model Retires. Your Records Don't.
The model your product answers with will be retired within months. The record that answer becomes will be kept for years, and is now discoverable in litigation. Surviving that gap means making the record carry its own sources, timestamp, and human decision, so that it does not depend on the model that produced it.
Three clocks, none of them synchronized
Every AI-assisted system in a regulated industry runs on three clocks, and almost nobody has checked whether they agree.
The model clock is the fastest. Anthropic commits to at least 60 days' notice before retiring a publicly released model; OpenAI's published practice is at least six months for generally available models and as little as two weeks for previews. After the date, every vendor says the same thing in different words: requests fail. And a model can die on two different days — Claude Sonnet 4 was retired on Anthropic's own API on June 15, 2026, while the same model carries an October 14, 2026 end-of-life on Amazon Bedrock. The lifecycle belongs to the platform you call, not to the model you think you are using.
The record clock is much slower. SEC Rule 17a-4 and FINRA Rule 4511 did not change when firms started routing work through models: an AI-generated item used in a customer communication or a recommendation is a record on the same terms it would be if a person had written it. Client files and employment records run on their own multi-year schedules.
The rule clock runs on a third schedule — the argument in The Deadline Moved. The Requirement Didn't. Obligations shift dates on a few months' notice while the underlying duty stays put.
Nothing forces these into alignment. The default outcome is a record that outlives, by years, the only system that could explain it.
The discovery question stopped being hypothetical
Last week, coverage of the Watson Grinding explosion litigation reported that more than 350 pages of ChatGPT interaction history from a retained engineering expert had been produced in civil discovery — including a prompt directing the model to show 3M bearing zero fault. The logs came out through ordinary discovery, not a breach, and from a third party working on the organization's behalf rather than an employee.
Read it as an architecture fact rather than a scandal: AI interaction history is part of the discoverable record, the perimeter includes parties you contract with, and the prompt is as producible as the output. Your AI layer generates records whether or not you treat it that way.
An answer whose only witness is a retired endpoint is not a record. It is an assertion with a date on it.
Why "we kept the logs" is weaker than it sounds
Two things get offered as the fix, and each covers less ground than it appears to.
Preserved weights are not a callable system. Anthropic has published commitments on model deprecation and preservation, including long-term preservation of weights. Good policy — and not the same thing as an endpoint you can query in 2032 to reproduce what your product told a customer today. Pinning a version has the same shape: it converts an unbounded risk into a dated one, and the date still arrives.
A transcript is not a reconstruction. Storing the prompt and the output preserves what was said, not why the answer was right — which sources it rested on, what the rule was on that date, or who judged it fit to send. Those are what an auditor or opposing counsel actually asks about, and they are exactly what a transcript omits.
Four properties of a model-independent record
The fix is not to slow the models down. It is to stop letting the model be the thing holding the record together. Four properties do that:
It carries its sources, not just its conclusion. Stored alongside the authority it rests on — citation, jurisdiction, and effective date as fields — the answer is re-derivable by a person, from the source, with no model in the loop.
Its timestamp is fixed at the moment, not asserted afterward. A date your own system supplies later is a claim. A date fixed independently at the time is evidence.
A human decision is recorded where one was required. Not "AI-assisted" in a metadata field, but a named person, a disposition, and a time, captured before the work product left the building.
Its provenance is structured. Which model, which version, which inputs, which tools it was permitted to touch — as fields, not reconstructed from a log file two years later.
None of these is an AI feature. They are the ordinary properties of a durable record, which is the point: what makes a record survive model turnover is what makes it survive an audit — the same root as the billing argument in You Can't Bill for an Outcome You Can't Prove.
What this looks like in shipped systems
LegiOS holds the first property: every obligation carries its citation, jurisdiction, effective date, and deadline as fields, which makes the compliance calendar re-derivable rather than remembered. Its scope is bounded too — when a question exceeds what software should decide, it escalates to an attorney rather than producing a confident answer nobody can later defend. Casebound holds the third, by putting the decision in the schema: a matter cannot open without a cleared conflicts check, a draft from the Lexwright engine cannot reach a client without a recorded attorney disposition, and every mutation is audited, including from an admin account.
LeoLog takes the timestamp property to its conclusion: each entry is hashed with SHA-256 and encrypted on the device, then anchored on-chain via Base and OpenTimestamps, with only the fingerprint ever leaving. The proof bundle is checkable by anyone at verify.leolog.io with no account, and would still check out if LeoLog were gone. LeafIQ makes the same move physically, keeping routes, orders, deliveries, and AR on one data model. And provenance is the standing problem at MagenticOS, where an agent's claim to have finished something must be backed by an artifact rather than a report — the argument in What "Completed" Means.
None of this was built as a response to model deprecation, and we are not claiming a model-lifecycle feature we haven't shipped. The observation is narrower: a system whose records are independent of any particular model treats a retirement date as a migration ticket; a system whose records are entangled with the model treats the same date as an evidentiary problem.
The test
One question, for anything you are building or buying: if the model behind this feature were switched off tomorrow, what would still be true about the records it produced?
If the sources are stored, the timestamps fixed, the human decisions named, and the provenance structured, you have records, and the model is a dependency like any other. If the answer is "we have the transcripts," you have a pile of assertions whose only remaining witness has been decommissioned on a schedule you never saw.
Models will keep getting better and keep getting retired, and both are fine. What cannot churn on a vendor's calendar is the record of what your system did and why — the same instinct behind fixing a research process before the decision rather than narrating it afterward, which is the standing discipline on the MacroSavant Substack. Building for it is most of what we mean by an operating system for an industry.
We publish here as we learn. Subscribe by RSS, or write to hello@mutagenic.io.