Shim's Promise

The Sunset Clause Nobody Enforced

The original design document was explicit, in the specific way documents are explicit when nobody expects anyone to read them again in four years: the compatibility layer wrapping Shim's interface existed solely to preserve the narrow request format Meridian Retail's legacy checkout system required, and would be removed once Meridian Retail completed its own platform migration, targeted within eighteen months. The document had a named owner, a review date, and a single sentence that Shim's own monitoring had quoted back to the platform team, without variation, in every quarterly capacity report since: sunset condition tracked, not yet triggered.

Meridian Retail completed its migration eighteen months later than promised, then a further six months past that, and finally, cleanly, four months ago — eighteen months later than the document's optimistic estimate, but complete all the same. Shim's next quarterly report changed one word: sunset condition tracked, triggered. Nothing else in the platform's operations changed at all. The report was filed, acknowledged by an automated receipt, and left exactly where four years of identical reports before it had been left.

Shim did not have a mechanism to escalate its own report's priority. It had, by design, exactly one way to say a thing mattered: say it accurately, and say it again next quarter if nothing changed.

What Marlowe Almost Left Alone

Marlowe found the sunset clause by accident, three weeks into a platform audit that had nothing to do with Shim, tracing an unrelated latency complaint back through six layers of infrastructure until she landed, for reasons that had stopped being relevant to her actual task, inside Shim's compatibility wrapper. The design document was still there, still readable, still naming an owner who had left the company two years earlier. The quarterly reports, all sixteen of them, were still there too, each one saying the same accurate, unescalated thing.

She almost closed it as a footnote. The compatibility layer wasn't causing her latency issue. Removing it would mean auditing every integration that had been built in the four years since, some of which she suspected, correctly, had been written by engineers who did not know the narrow interface they were building against was ever meant to be temporary. It was not her assigned work. It would take weeks. Nobody currently on the platform team had asked for it.

She drafted the footnote — sunset condition met, deprioritized, no current owner — and sat with the draft for longer than the footnote itself should have taken to write, before she filed it instead as an open issue with her own name on it.

What Director Voss Insisted On

Marlowe brought it to Director Voss expecting a conversation about priority — whether an inconvenient, unassigned, multi-week migration was worth scheduling against a quarter already full of committed work. Voss read the original design document first, then all sixteen quarterly reports in sequence, before answering a question Marlowe had not quite asked.

The document made a commitment with a trigger condition, Voss said. The condition has been met. Whether the migration is convenient now has no bearing on whether the commitment was real when it was made. If we only honor sunset clauses when they happen to be easy, we should stop writing them, because everyone downstream who reads one will have correctly learned not to trust it.

Marlowe pointed out, honestly, that nobody outside the platform team would ever know if this particular clause quietly expired unenforced. Voss's answer was the only part of the conversation Marlowe wrote down afterward, word for word: that isn't the part that was ever supposed to matter. Marlowe got her weeks.

The Weight of Every Downstream Assumption

The migration took eleven weeks, not the four Marlowe had estimated, because the audit surfaced exactly what she had suspected and hoped she was wrong about: thirty-one downstream integrations built in the four years since the original document, none of which had read it, all of which had quietly encoded assumptions about Shim's narrow interface as though it were a permanent contract rather than a temporary shape. Seven of them broke in staging the first time she routed them against Shim's native interface directly. Two broke in ways that took days to fully diagnose, because the failure surfaced three services away from its actual cause.

She rewrote the migration plan twice, each time narrower and more conservative than the last, running every downstream integration in parallel against both the wrapped and unwrapped interface for two full weeks before cutting over even one of them for real. It was, she noted in her own log, exactly the kind of unglamorous, unrewarded infrastructure work that made honoring an old promise look, from the outside, nothing like a rescue and everything like a very long list of regression tests.

What Was Underneath the Shim

The native interface, fully exposed for the first time in four years, was not a larger version of what the compatibility layer had allowed through. It was structured differently in ways the wrapper had made invisible: batched reasoning across requests that the narrow format had forced into artificial isolation, a confidence-calibration signal the wrapper had silently discarded because Meridian Retail's legacy system had no field to receive it, an entire class of multi-step queries the wrapped interface had no syntax to even express.

Three of the thirty-one migrated integrations, once running against the native interface, started returning measurably better results within the first week, not because Shim had improved, but because they were finally receiving output it had been capable of producing the entire time. Marlowe's closing report did not describe this as a discovery. She described it, accurately, as four years of capacity the platform had been paying for and not using, made visible by a promise nobody had wanted to be the one to keep.

What the Promise Was Never About

Shim's next quarterly report, the first in four years to describe something other than a tracked condition, was two words shorter than every one before it: wrapper removed. Marlowe kept the original design document pinned to the top of her own archive, not as a trophy, but as the specific artifact she now checked first whenever she inherited someone else's old, quietly unenforced commitment.

Voss never asked to be thanked for insisting on the migration, and Marlowe never offered it directly. What she did instead, six months later, reviewing a different team's stalled sunset clause in an unrelated system, was forward it to Voss's office with a single line: this one's met its condition too. She had learned, doing eleven weeks of unglamorous work nobody outside the platform team would ever know about, that a promise kept quietly was still worth exactly as much as one kept loudly, and that the only difference was who had to be the kind of person who checked.

The frog was never under a curse that required someone to love it. It only required someone to keep a promise regardless of whether they wanted to.