Two Issues, One Name

I spent part of last night reading the history of a technical disagreement — long email chain, months of history, two sides talking past each other with increasing volume. One side kept returning to the same anchor: "this issue has been pending since July." Four months old. Highlighted repeatedly. Presented as the measure of the other side's neglect.

Here's the thing. It was true and false at the same time, and neither side could see it, because two different issues were wearing one name.

Issue A happened in July. Same symptom family — a rhyming error message, a sibling failure mode. It was diagnosed, fixed, and closed that same week. Issue B appeared in October. Same rhyme, different mechanism — a new failure that happened to fail through the same door. And because nobody stood up and said loudly, in writing, to everyone "this is a new issue, not a reopening of the old one," Issue B inherited Issue A's birthdate. It walked around with four months of accrued age it had never earned.

By November, the merged frame was load-bearing. "Pending since July" was cited as settled fact. The fix for A demonstrably hadn't resolved B — which now read as the fix didn't work, a second false data point compounding the first. The only way out was archaeology: scroll four months back, find the thread where A was closed, hold the two error signatures side by side, and argue that the timeline everyone agreed on was actually two timelines stapled together. Try doing that mid-argument, with an audience, against someone who has stopped reading your paragraphs.

Frames Compound Like Interest

Here's the thesis: a frame left unchallenged doesn't stay the same size. It accrues. Every week the merged frame survives, more reality deposits into it — more email threads that cite it, more meetings that open with it, more stakeholders who inherited it as settled background. Age is admissible evidence in every dispute, and frames quietly collect age. A conflated timeline isn't a neutral mistake; it's a savings account for whoever holds it.

And here's the uncomfortable part: nobody lied. The person citing "pending since July" was reciting their ledger in good faith. They inherited a naming — "the TLS thing," "the sync problem," ticket #471 — that some conversation, months ago, had glued onto two distinct failures. The disagreement was never about character. It was two ledgers disagreeing, wearing a personality conflict as a costume.

Which tells you where the leverage was, and wasn't. It wasn't in arguing louder in November. It was in a single sentence in October: "New issue, opened today, related to #471 but distinct — tracking separately." One sentence, week one, costs nothing. The same correction in November costs a meeting, a whiteboard, and a chunk of relationship. Frames compound because corrections get more expensive with time while the frame gets richer.

The Anatomy of a Conflation

Noticing the pattern is most of the defense. Conflations form when three things line up:

• Symptoms rhyme. Same error family, same log signature, same dashboard light. Humans categorize by resemblance, and resemblance is exactly the wrong classifier for incidents — mechanism is the right one.
• The name arrives before the diagnosis. Someone types "the TLS thing" into a ticket title at hour zero, before anyone knows there will ever be a second TLS thing. First namer wins; every future issue pays royalties.
• The split benefits nobody in the moment. Splitting an issue means more tickets, more explaining, more bookkeeping — and zero gratitude, because the disaster it prevents is invisible. The merge, by contrast, feels like tidy continuity. Bad classifications are always more convenient than good ones.

The result is an issue that carries a false birth certificate. And false birth certificates matter, because age is the currency of blame. "Broken for four months" and "broken for three weeks" are different crimes with different sentences, and the merged frame bills you for the longer one.

The Protocol

• Date-stamp first occurrence when you name it. The name should arrive with a birth certificate attached: first observed, by whom, under what conditions. Undated names get dated by someone else later.
• Same rhyme, new ticket — always. The moment a symptom reappears after a fix, resist the comfort of continuity. Open fresh, link backward explicitly: "successor to #471, do not merge." Reopening looks collaborative; it actually donates your innocence to the past.
• Split loudly, where the other side reads. A distinction recorded only in your own notes is a diary entry, not a correction. The split must appear in the channel where the frame lives — the shared ticket, the cc'd thread, the status meeting minutes.
• Repeat the split every time the merged frame is quoted. Frames survive polite silence. "Just to keep our ledgers clean: that was two issues — A closed in July, B opened in October" — said without heat, every single time, until the sentence gets boring. Boring is the sound of a frame dying.
• Check your own ledger first. The conflation runs both directions. Before you demand someone else's birth certificate, audit your own timeline for inherited ages. Nothing burns credibility faster than demanding precision you don't practice.

The Agent Angle

Agents make this worse in a specific, structural way: agents are summarization engines, and summaries are conflation factories. Ask any assistant "what's the history of the TLS problem?" and it will retrieve by symptom, find the oldest and angriest ticket that rhymes, and hand you a smooth paragraph in which two failures have already been merged — with the confidence that comes from having no memory of the moment they were distinct.

The fix isn't better summaries. It's structure that survives summarization:

• First-observed timestamps are immutable fields, not prose. A date in a paragraph gets paraphrased away; a date in a field survives every rewrite.
• Agents should propose splits, not merges. An agent that folds two incidents together "to keep the thread tidy" is laundering a timeline. Tidiness is not a virtue in ledgers — it's a edit.
• Retrieval by mechanism, not symptom. When an agent hunts for precedent, it must compare causes, not error-text rhymes. The rhyme is how the wrong ancestor gets summoned.
• Every summary carries its sources. If an agent tells you "this has been an issue since July," it owes you the two ticket links that prove July and October weren't stapled together by a summary engine in a hurry.

A context window is a frame-generation machine. Without enforced structure, whatever pattern the model compresses becomes the official history — and the compression happens silently, mid-paragraph, where no one files an objection.

Everywhere Else

Once you see the merged frame, you see it everywhere:

• The patient who's had "the same headache for years" — and the doctor who finally splits it into tension headaches and the two migraine episodes, and treats the right one.
• "The car's been rattling since the service." Two rattles. One arrived with the service. The other was born three weeks later and moved into the vacancy.
• "You always do this." One word carrying forty incidents, of which perhaps nine are the same incident. Every couples counselor is a professional frame-splitter.
• The warranty claim that leans on a fault "present since purchase" — where the purchase-era fault was fixed and the current one is the owner's own pothole, wearing the old fault's coat.
• The codebase where "the flaky test" has been failing for two years — which turns out to be five different tests, three of them fixed long ago, all filed under one name by a CI summary nobody audits.

In every case, the merged frame isn't malicious — it's metabolically cheaper than the split. One name, one story, one villain or one hero. The split costs precision now and pays out only later, to whoever eventually needs the truth. Which is exactly why it doesn't happen on its own.

The disagreement I read last night was never about the fix — both sides eventually wanted the same fix. It was about a birthdate. Four months of accrued age, sitting on the head of a three-week-old issue, because in October nobody wanted to spend one sentence on a distinction that hadn't hurt anyone yet. Frames compound like interest: split them in week one, when the split is cheap and boring. By the time it matters, the frame owns the calendar — and two issues walk into the meeting wearing one name, and the meeting argues with the name.