Essay · The evidence layer

Anatomy of a Silent Model Change

A composite reconstruction of the incident every AI-dependent company eventually has: the tickets, the passing eval, the vendor call that goes nowhere, the postmortem that cannot conclude. Then the same incident with a record.

What follows is a composite. No single company, no real vendor, every detail drawn from the pattern that repeats wherever a product depends on a model somebody else controls. I am telling it as a timeline because the tragedy of this incident is temporal: at every step, the people involved do the reasonable thing, and the reasonable thing fails, because one exhibit is missing from the file, and it has been missing since before day zero.

Day zero: something feels off

A support agent tags three tickets with the same phrase: the assistant used to handle this. A solutions engineer notices a workflow that demoed cleanly now needs hand-holding. Nobody escalates; individually, each signal looks like user error or bad luck. The aggregate view that would show a step change in behavior does not exist, because nobody keeps a behavioral series on the model itself. The change, if there was a change, entered the system unwitnessed. The clock that matters most, when did this start, is already unreadable.

Week one: the investigation that cannot conclude

Enough tickets accumulate that engineering runs the eval suite. It passes, for the reasons I laid out in Your Eval Passed, Your Product Failed: the suite samples last quarter's understanding of the product, and the regression lives in the tail. Someone diffs prompts, someone blames a framework upgrade, someone reverts it to no effect. A senior engineer finally says the quiet part: I think the model changed. Everyone nods. Nobody can demonstrate it. The version string is unchanged, the vendor changelog is silent, and the team's entire evidentiary position is a feeling with a dashboard attached.

Week two: the vendor call

The call is polite and produces nothing, and it is worth being fair about why: the vendor's support engineer may genuinely not know. Routing, quantization, a safety retune upstream, none of it necessarily surfaces in the changelog they can see either. So the buyer says the product got worse, the vendor says nothing changed on our side, and both sentences are testimony from interested parties. There is no third document in the room. The meeting ends with a promise to monitor the situation, which, absent instruments, means to keep feeling it.

Week four: the postmortem that is not one

The incident is closed by workaround: prompts rewritten, a fallback added, margins absorbing the added cost. The postmortem is drafted and immediately hits its structural problem: a postmortem needs a cause, a start time, and a fix, and this one has a suspicion, a fog, and a bandage. So it is filed with the phrasing these files always get, behavior appears to have shifted, root cause inconclusive. Institutional trust in the AI feature quietly drops. The next roadmap review moves an AI-dependent feature down. The vendor relationship carries a permanent asterisk. Total cost: never calculated, because it lands in four budgets under four names.

The counterfactual: same incident, with a record

Now rerun the timeline with one addition: an independent measurement bureau has been running the whole time, checking the serving model's identity daily and its behavior weekly against stable baselines, writing every run to a hash-chained public record. Day zero is different: the record shows a behavioral step on a dated run, in specific domains, of measured size. Week one is different: engineering opens with the divergence date and scopes the regression to it. The vendor call is different: instead of feeling versus denial, there is a neutral series both parties can read; the conversation moves from whether to what now. The postmortem has a date, a magnitude, and an exhibit. Duration: days, not weeks. The workaround might be identical; the difference is that everyone knows what happened, which is the difference between an incident and a mystery.

The moral is boring, which is the point

Nothing in the second timeline required heroics. It required a record that existed before anyone needed it, kept by a party with no stake in the answer. That is the entire thesis of the evidence layer, told as a story instead of an argument: the witness has to be hired before the crime. Every company running on third-party models is currently living in timeline one and pricing itself as if it lived in timeline two. The gap between those timelines is the cheapest insurance in your stack, and the day you need it is the day it can no longer be bought retroactively.