Administrator research guidance, candidate evaluation, promotion, rollback and backup.
Reviewed · 2026-10-12
On this page · 9
1. Applied guidance
Research-kind baselines combine with approved domain, topic, space and work guidance. Execution preparation freezes exact versions; promotion does not rewrite ongoing results. Explicit revocation and current source permissions are still checked. Skills are not model-weight training or proof of research success.
2. From candidate to approval
Only current platform administrators manage candidates, contract evaluation, promotion, rollback and revocation. Preview each exact change before applying it. Method improvements actually delivered in an answer may be preserved only as unapproved candidates in that private work. Missing proposals create no candidates, and saving invokes no additional AI. AI never replaces approved guidance automatically.
3. Evaluation and cost
Contract evaluation compares the active and candidate instructions against fixed cases and records results, differences and hashes. It does not measure model-answer quality or experimental success. Before promotion, administrators review sources, scope, limitations, cost and their review rationale. Requesting improvements within a paid research answer adds tokens to its existing usage. Separate paid quality evaluation is not started automatically.
4. Privacy and recovery
An administrator role does not grant access to another person’s private conversations or memories. Private candidates remain bound to their owner and current source permissions and are never automatically copied into global guidance. Write a separate generalized candidate and explicitly review it for private information and secrets. Authorized version exports do not replace complete encrypted recovery backups. Instructions sent to external ChatGPT cannot be guaranteed hidden.
Keep only the know-how you choose from a discussion
In the same work composer, ask: “Separate the literature-analysis and manuscript-writing know-how from this discussion. Propose only the changes to each existing method.” Only delivered, saved requests and results are evidence. External ChatGPT or Codex can prepare structured proposals using returned evidence IDs and current permissions, then use the existing reviewed save contract. On the proposal page, select only the drafts you want to keep and review the save. Unselected proposals and active versions remain unchanged. Preparing or saving a draft does not validate quality or approve v2.
Apply several methods together
Use the applied know-how count below the work title, or work menu → Know-how settings, to manage defaults. No empty badge appears for zero items. Search names, handles or activities and review exact versions and roles before saving. Add to a work on the independent detail page lets you choose a space and work and pin an exact version. Type @ in a follow-up, new research request or personal AI consultation to search researchers and know-how together. Choosing know-how adds it only to that request without changing defaults. [[ source citations keep their existing behavior. Multiple selections never auto-merge assets or adopt newer versions. Web, consultation and ChatGPT handoff check the same IDs and current permissions.
Split, merge, originals and shared evidence
One discussion can produce independent proposals that reference the same segment as shared evidence. Drafts start privately. Separate paper-specific corrections and unverified scientific conclusions from reusable guidance. Change organization prepares a split or merge request in the existing composer for review and sending. A merge proposes a new asset linked to exact parent versions. It does not delete or overwrite the versions, history or active state of assets A and B. Changing how a proposal is organized requires a newly reviewed proposal, not an unreviewed browser-side rewrite.
Version inheritance and evaluation
Derived work inherits the exact selected versions of independent assets. For existing work-specific methods, the collapsed Work know-how controls retain version and candidate selection in the original work. Review generalized instructions and create a separate asset before reuse in other work. Execution applies modules relevant to the current stage, such as literature, synthesis, writing or review, while retaining links to other stages. New versions are not adopted automatically. Initial approval reviews the actual instructions and their generalization; it does not claim measured quality improvement. Improvement candidates require actual runs with matched questions, inputs, models and effort, held-out cases and human review before adoption. A candidate trial is a one-off selection that does not change the default and uses the normal AI allowance.
Protection, storage and execution boundaries
An independent detail page shows introduction, ID, handle, authors, usage, current permissions and versions. Editing the introduction or image link does not change execution guidance; image access is checked separately. The default direction is protected guidance. Write a draft or propose a direction improvement, review it and select which drafts to keep. Saving an introduction or candidate does not change the active version. Ordinary catalogs and proposal cards exclude protected instructions. Source review requires current permission and AI-provider processing permission is separate. Reads, selections and structured proposal saves do not run AI, while sending analysis or reorganization requests may use your allowance. Independent external analysis does not prove protected-service execution. Instructions belong in encrypted backups, not code or static files. AI quality, actual ChatGPT tool availability and server speed require separate validation.