Eggs — decomposition review

What actually ran

Backend: Lyra’s scoped public Responses endpoint, model alias lyra-chatgpt-pro, requested effort instant. Three separate durable jobs reused the app’s specialist instructions, input templates, schemas, and assembly function. Schema guidance was supplied in the prompt; this was not OpenRouter structured-output mode.

Dimension scout: proposed seven axes and their alternatives.
Trace specialist: mapped exact words to those axes.
Context and retrieval specialist: drafted evidence fields, mismatch risks, a claim template, and interview questions. No answers to the interview were supplied.

Comparison limit: this is a live backend run plus a local assembly before/after comparison. It is not a completed run of the deployed web UI, nor a blinded comparison of two models.

proposed dimensions

preserved alternatives

completed provider jobs

Hosted site access

The saved site URL responds with HTTP 401 without authentication. After approved ChatGPT sign-in, the old deployment returned “You don’t have access to this site.” Management access to that old project is unavailable. A new deployment is being prepared under the current account.

Review the interpretation map

Every choice starts unresolved here. These are alternative scopes, not claims that the options are healthy or harmful. You can record a provisional focus and download your review; nothing is sent to a server. Some options overlap and should be split in a later revision.

What the app discarded

The raw model output was broader than the assembled artifact. The local fix retains all five permitted alternatives and reserves at least one exact cue for each valid dimension.

Object	Before local fix	After local fix

Validation: all three regression tests reproduced the loss before the fix; all 14 targeted tests passed after it. The changed assembly has not been deployed.

Codex review: what still needs work

This is an attributed methodological assessment of the live output, not source evidence about eggs.

Stop here before contextualizing

The generated interview starts by asking for personal risk factors. That is premature if the goal is a general, shareable evidence map. Establish the purpose first; then ask only context that changes scope or evidence matching.

The next stage will need your priorities, intended dose, realistic replacement foods, and any relevant population constraints. None have been assumed.

Inspect the model’s proposed interview

Inspect the generated claim template

This is a provisional template, not an approved claim. List order should not decide the active research question.

Take the artifact with you

This HTML file works offline and can be shared as a file. The JSON export includes the original and preserved decompositions, the audit, the Codex review, and your selections. It contains no endpoint credentials or personal medical profile.

Inspect complete underlying records