Governed Intake Systems
Ethics review, AI risk assessment, and innovation-fund intake: the same question in three vocabularies: is this allowed, and can the agency still explain why a year later?
The pattern underneath
These started as three unrelated requests. An ethics checklist for researchers. A funding pipeline for employee ideas. A risk assessment for proposed AI systems. Different sponsors, different years, different vocabulary.
They are the same system. Each takes an unstructured intention, I want to study this, fund this, build this, and runs it through the same governed pipeline: published criteria, proportionate review, a named sign-off, a durable record. The domain changes; the machinery does not.
The three systems
The agency's ethics-in-research checklist, built as an application rather than a document. Four parts follow the shape of the work: who is asking and who approves, study design and data acquisition, analysis and interpretation, and how results are communicated. The analysis part asks what the model does, what it assumes, whether those assumptions hold, and whether a methodologist reviewed the method.
- Conditional follow-ups: answering one way opens the questions that answer must justify
- Named approver captured alongside the respondent, so a review has an owner
- Responses aggregated for reporting, so patterns across submissions are visible
- Disclosure here is what triggers a full algorithmic assessment downstream
The second stage of the ethics review rather than a parallel process, required only when what was disclosed upstream shows automated decision-making. It is the most mechanised of the three: an eight-dimension assessment feeds a scoring service, a DMN table classifies risk on thresholds that tighten as a system nears production, and a BPMN process routes every submission to a manager, anything production-bound to a director as well.
- 87 scored questions reduced to a risk ratio by a service, not a form
- Risk thresholds in a decision table, editable by policy owners without a deploy
- Two-tier approval with a real edit-request cycle back to the author
The same machinery pointed at a friendlier problem: employees submitting ideas against posted challenges, and the agency deciding which to fund. The form saves partial work, so a long proposal can be finished another day, and an applicant can see where their submission stands instead of waiting for an email. Later releases added reporting, because a queue of adjudicated submissions is itself data.
- Challenge and fund entities modelled explicitly, so intake criteria are data rather than page copy
- Draft-save on a long form, so a multi-day proposal survives a closed browser instead of being lost
- Applicant-visible status, so a submitter can see where their proposal stands without chasing an email
What repeats across all three
Each system fixes what will be asked before anyone submits. A decision is reviewable after the fact because every submission was measured against the same published set, whatever the reviewer thought at the time.
None of them asks every question of every submission. Ethical Checkpoint opens follow-ups conditionally; AIAA routes on declared lifecycle phase and skips scoring for exploratory work. The clearest case sits between the systems: clearing the ethics review ends the process for most projects and starts a much heavier one for the few that automate decisions.
Every one of them ends at a person, identified in advance, who is accountable for the outcome. Scores and checklists shape what the reviewer sees and how hard they have to look. None of them approves anything.
Submissions persist whether or not they were approved. That turns a review process into an inventory of what was proposed, what was allowed, and on whose authority. It is the artifact an institution needs when the decision is later questioned.
Why this transfers
Governed autonomy is the same problem with a faster actor. Whether the thing being reviewed is a research study, a funding proposal, or an autonomous system taking an action, the institution still needs the same four things. What changes is throughput, and throughput is what breaks a review process that was only ever a document.