{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreidiysngga7p5sq6vbf2z25u5alxzcsjzlwpzgnfzov6fayczr7k6i",
    "uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mjpughujm4b2"
  },
  "path": "/t/feature-request-externalized-instruction-locking-mechanism-for-consistent-adherence-to-user-specified-output-constraints/1374362#post_2",
  "publishedAt": "2026-04-17T20:11:21.000Z",
  "site": "https://community.openai.com",
  "textContent": "This is a really thoughtful write-up, appreciate you taking the time to break it down this clearly.\n\nI get the core issue you’re pointing at: right now, user-defined constraints behave more like “strong suggestions” than hard rules, which can lead to drift when the model falls back to more typical outputs. That’s especially noticeable in parameter-sensitive or design-heavy workflows like the examples you gave.\n\nThe idea of a persistent, session-level instruction layer (like what you’re describing with EILM) makes sense in that context, particularly the distinction between treating inputs as evaluation constraints vs. just prompt context.\n\nI’ve gone ahead and logged this as a feature request for the team. The level of detail here (especially around expected behavior and failure modes) is genuinely useful for shaping how something like this could be approached.\n\nAppreciate you sharing it, if you end up testing more edge cases or have concrete examples where adherence breaks down, feel free to add them here. That kind of signal helps a lot.\n\n~ Taylor",
  "title": "Feature Request: Externalized Instruction Locking Mechanism for Consistent Adherence to User-Specified Output Constraints"
}