Skip to content

002: One model, two shutdown dates

  • Date: 2026-10-08
  • Area: Registry data

Context

Step 07 built the registry with a rule: the same model ID may not appear twice. That rule catches copy-paste mistakes. The next job was to add OpenAI's retirements.

Symptom

While collecting OpenAI's data, the same IDs kept turning up in two different sections of the page with two different shutdown dates. For example, gpt-4-1106-preview is listed for shutdown on 2026-03-26 and again for 2026-10-23. With the Step 07 rule, the data could not even be loaded.

Investigation

I checked each duplicate against the page. There were two kinds:

  1. Real reschedules. OpenAI announced a date, then later announced a new one (gpt-4-0314, gpt-4-1106-preview, code-davinci-002).
  2. Not a shutdown at all. The 2024-08-29 entry for babbage-002 and davinci-002 only stopped new fine-tuning; the models kept working. So it does not belong in a retirement registry.

Root cause

The schema matched one provider's page (Anthropic's), where every model is listed once. It did not match how retirements really happen: plans change.

Options

  • Keep only the newest date (loses information the research needs).
  • Allow several entries per ID (lookups become ambiguous).
  • One entry with the current plan plus a list of earlier plans.

Fix

The third option, recorded in ADR 0007: a superseded list, plus a shutdown_is_earliest flag for Google's "earliest possible" dates. Fine-tuning-only entries are left out.

Lesson

Read real data before fixing a schema. A validation rule that looks obviously right ("no duplicates") can encode a wrong assumption about the world, and the data is what tells you what exactly the real case is.