Essay 06

What Does a Conditional Approval Require?

A late report. Conflicting numbers. A safety signal that pauses the work. Regulated Gate Decision Support asks how those findings should change permission to proceed.

First repository record ·

Research context

Reference implementation · Six illustrative records

Essay draft · Based on practitioner accounts in Syner-G’s IND webinar. Timestamps refer to the supplied automated transcript; accounts are paraphrased. The RGDS example is separate and illustrative.

A report is marked final, but the numbers in its summary do not match the numbers inside. Which should the team use?

In Syner-G’s webinar on preparing submissions for investigational drugs, medical writer Kasturi Puram describes encountering this problem (38:51). The supporting data gave her a reason to suspect which value was correct. She still asked the team to confirm it.

Her account stops short of explaining how the discrepancy was resolved. But it reveals a practical problem: finding an error is only the beginning. Someone must establish the correct information and update the documents that depend on it. If the error affects a decision, that decision may need to be reconsidered too.

The problem becomes especially important when approval comes with conditions: “You may proceed, provided this is resolved.” What work may continue? What must wait? Who decides that the condition has been met?

Regulated Gate Decision Support, or RGDS, provides a way to record those answers. Its purpose is to make the permission, its conditions and the responsibility for resolving them available for examination. Whether people actually honor those conditions remains a separate question.

A task asks someone to do something. A condition defines what permission depends on. If the two are treated interchangeably, work can continue while the condition sits on a task list.

The last report changes the work

An Investigational New Drug application, or IND, brings together evidence for a proposed clinical investigation. FDA’s overview of the application describes its nonclinical safety evidence, manufacturing information and clinical protocols. For the people who may enter a study, errors in that connected account can matter beyond the document where they first appear.

Puram describes drafting with placeholders while reports are still arriving, then carrying information between summaries, the investigator’s brochure and the protocol (07:38). This lets useful work proceed before every source is final. It also creates a dependency: when the source changes, the people preparing each dependent document need to know.

The deadline can obscure that dependency. Considering a planning example in which a key report on toxic effects would arrive only four days before submission, regulatory specialist Drew Barlow says the team would challenge whether the timetable was realistic (43:43). He distinguishes that situation from a late manufacturing result or a measurement of how a drug holds up over time. The time needed depends on what the evidence could change. He is discussing planning situations, not reporting a four-day turnaround achieved by a particular project.

A team leader can ask which work may continue while that report is pending and which decision must wait. Those permissions can differ. Drafting a section, finalizing its account of the evidence, submitting an application and beginning a study are separate actions. An internal approval should state the action it authorizes; regulatory requirements remain separate.

Regulated Gate Decision Support records a phase-gate decision: a point where an authorized team decides whether work can proceed, must stop or needs more evidence. Its reference implementation preserves the question, alternatives, evidence gaps, accepted risks, responsible people and follow-up obligations. Applied to the scheduling problem, it would ask what the pending report could change and who must examine it before the next action is authorized.

A reason to reopen the decision

Puram also recalls a project where a possible safety concern in a study of toxic effects in dogs led the team to pause IND writing and seek regulatory input before submitting the application. A briefing document was prepared, and the resulting input was incorporated into the IND (47:49). The transcript does not identify the compound, the signal or the eventual regulatory outcome.

Here, an emerging concern changed the course of the work. To reconstruct that decision, a later reviewer would need more than the revised schedule: what prompted the pause, who authorized it, what question required clarification and what the response changed. The webinar supplies a practitioner’s account of the intervention. It is not a record of RGDS use.

The same discipline applies outside drug development. A supplier check may uncover a discrepancy, or a late test may alter a planned software release. A condition has practical force when the person responsible for the next action can withhold permission and knows what evidence would justify reconsideration.

The repository’s AI-assisted readiness example describes a decision about a non-clinical evidence package for an Investigational New Drug submission. A report on toxic effects awaits final sign-off, and data on how the drug holds up over a longer period is pending. The authored record chooses a conditional go. It is an illustration of the record structure, not evidence that an actual submission could appropriately proceed on these terms.

Two obligations make the choice more specific. A scientific adequacy owner must provide the final signed toxicity report before filing. A manufacturing and stability readiness owner must supply the longer-duration dataset and update the dependent summaries and tables. Each condition identifies an owner, a due date and the evidence needed to close it.

Both follow-up actions remain marked open in the example. That matters more than the reassuring presence of a recorded approval. The record tells a reviewer what is still owed; it does not establish that either obligation was fulfilled. Its historical example dates likewise do not turn those entries into current operational deadlines.

The example also preserves an assumption: the audited draft report will not change materially when signed. Puram’s account of inconsistent numbers shows why final status alone cannot settle that question. The comparison is this essay’s analysis; the authored RGDS example is separate from the webinar accounts. A reviewer would still need to examine changes and decide whether the original rationale holds.

Reading the illustrative record

Permission recorded.
Conditions still open.

  1. A conditional go

    The example records human approval subject to two obligations.

  2. Evidence to close

    A signed report and additional stability data, with dependent records updated.

  3. Judgment still required

    A changed report may require reconsidering the original rationale.

  4. Outcome unestablished

    Both follow-up actions remain open. Completion is not demonstrated.

Conceptual reading of RGDS-DEC-0006. The authored example demonstrates a record structure, not the appropriateness of an actual submission or fulfillment of its conditions.

A valid record is a start

RGDS includes software that checks whether a decision record contains required information and whether selected entries fit together. A conditional approval, for example, needs stated conditions. These checks can expose omissions that a reader might miss.

A complete record does not establish that a condition was fulfilled. A field containing a person’s name cannot authenticate their participation. A complete-looking evidence reference cannot establish the quality of the source. Even a record that passes all configured checks may rest on an unsupported assumption or describe an action that never occurred.

This boundary makes the validator useful without asking it to do the reviewer’s work. Software can check that a closure requirement is present. A qualified person must determine whether the supplied evidence meets it. The institution must then ensure that this determination reaches the process whose permission depends on it. These are connected responsibilities, each requiring its own evidence.

Assistance without transferred authority

The AI assistance policy permits bounded work such as drafting, comparison and extraction, while reserving decisions, approval and risk acceptance to people. Regulated Gate Decision Support can also be used without AI. When AI contributes, the record is meant to identify the contribution and its review rather than let generated prose disappear into the institution’s account.

Consider an AI assistant summarizing the kind of report Puram describes. This extension is hypothetical; her account does not attribute the discrepancy to AI. A model could carry one conflicting value into a fluent summary without making the disagreement visible. The reviewer would need the source passages, the authority to reject the summary and a way to prevent dependent drafts from continuing to use it. Recording that someone reviewed the output would leave those practical questions unanswered.

The related AI Assistance Governance work asks whether material decision claims can be reconstructed from inspectable non-AI sources. Its assessment distinguishes checking the record’s structure from examining the substantive basis for its reasoning. Keeping those questions separate prevents a technically valid document from being mistaken for an independently supported decision.

Preserving that distinction also means preserving history. Reassessing a decision without relying on model output should not erase the fact that AI contributed to the original process. The original disclosure, the reviewer’s changes and the sources supporting the final rationale each tell a different part of the story.

A projection is an invitation to test

The Regulated Gate Decision Support independent study explores ten questions about decision reconstruction, AI accountability, implementation and possible institutional benefits. It provides the earlier research context for the work, including literature synthesis and modeled scenarios. “Independent study” identifies the author’s research; it does not establish independent validation of the implementation.

The study’s estimates are proposals to examine, not measured benefits. The current reference implementation does not establish time or cost savings, fewer submission deficiencies or regulator acceptance.

Technical checks and projected savings

The decision-log schema checks whether information has the required form. Additional semantic checks examine selected relationships between fields. For example, a conditional go needs conditions, and disclosed AI use requires review information under the implementation’s rules. These checks can expose omissions that a reader might otherwise miss.

The documented local review reports that all six illustrative records passed strict validation and all nine regression tests passed. Earlier checks exposed a gap between structural and semantic validation: removing conditions or human-review information could pass one entry point while failing the stricter checks. The v2.0.1 correction unified the entry points’ semantic checks and enabled date-format checking. The six canonical examples and schema contract were preserved. A validation pass still leaves the truth of the record and fulfillment of its conditions to substantive review.

Its quantitative projections need to be read on that basis. One scenario compares a 45-day decision cycle with a proposed 22-day cycle across fifteen decisions. The arithmetic gives 345 summed decision-days. It does not show that a program finished 345 calendar days earlier. Decisions may overlap, waiting time may differ from staff effort, and a faster decision may leave the overall completion date unchanged.

The study also discusses financial value and possible regulatory developments. Those estimates and forecasts should remain distinguishable from observed outcomes and current requirements. The current reference implementation does not establish savings, reduced deficiencies or regulator acceptance. Linking the study makes its assumptions available for examination without importing every projection as a result.

The reviewed core release is v2.0.1; the earlier study identifies itself separately as v1.4. Their versions and evidence boundaries remain distinct.

The core evaluation plan is explicitly prospective. It asks whether the approach improves decision readiness, retrieval, evidence quality and governance execution. A useful study could compare how accurately reviewers reconstruct matched decisions with and without the structured records, recording omissions, disagreement and time. Speed would matter alongside accuracy, not as a substitute for it.

A separate inquiry would need to follow conditions into operations. Did a hold prevent the next action? Did the responsible person receive the evidence in time? Was an adverse finding allowed to change the decision? A record that is easy to retrieve could still describe a process in which nobody had practical power to intervene.

For people approving the work

A conditional approval should give the next person enough information to act. Whether the setting is a drug submission, a school purchasing a service or a business preparing a release, the approving team needs to answer five questions:

  • What may proceed? Name the action being authorized and the work that must wait.
  • What is still required? State the evidence needed, who will supply it and when it is due.
  • Who accepts completion? Identify the person with the expertise and authority to decide whether the evidence satisfies the condition.
  • What would reopen the decision? Explain which new findings require a pause or reconsideration, and who can order it.
  • Who needs the correction? Identify the people, documents and actions that relied on the earlier information, then verify that the change reached them.

These are questions for evaluating practice. The example records do not demonstrate that an institution has carried out these responsibilities. A named owner and a completed field are starting points for examination.

The approval must remain answerable

Return to the inconsistent report. Puram describes recognizing the discrepancy and asking for confirmation. Following that concern through an institution would require evidence of the answer, the corrected source and the judgment about its consequences. A revised number might leave a decision intact or change its basis. The record should let a later reviewer distinguish those possibilities.

If work has already proceeded, the inquiry extends downstream. Which documents relied on the old finding? Which teams received the earlier permission? What can be corrected, and what consequence remains? Closing the original task cannot establish that dependent records and actions were corrected with it. Evidence of delivery, revision and execution is needed to trace that propagation.

An institution does not finish its responsibility when it records a conditional approval. It has made a commitment about what must happen next. Regulated Gate Decision Support gives that commitment a structure that others can examine. Whether it keeps its force depends on people being able to recognize a reason to pause, exercise authority and establish what changed as a result.

AI use & disclosure

Disclosure updated November 3, 2026

RGDS specifies how AI assistance should be recorded, including its purpose, sources, limitations and human disposition. Its six decision records are illustrative. Example 0006 names a “ChatGPT-style assistant” with an unpinned version; its review entries model a workflow and do not independently establish an actual institutional review. The responsible researcher retains accountability for adopted work. Schema and semantic checks examine required fields and specified relationships. The inspected records do not provide a retrospective task-level inventory of AI assistance in developing the research, prose, code or tests.

Codex (OpenAI) also assisted this website essay’s drafting, revision and source-record inspection. That editorial work does not establish independent human verification of every underlying source.

Project AI-use and verification record ↗

Program AI-use statement