The Research Data Director / Prototype gallery

RDA 27th Plenary Meeting, London, 6–8 October 2026

Companion page

Prototype gallery

Annotated screenshots of the Research Data Director as it was built, from the first friendly interface to the named persona team, and the first biomedical crosswalk studio

These are research prototypes. They were built in July and August 2026 to test the blueprint, not to run in production. They use synthetic or illustrative cases, every output is a draft for human review, and interfaces change from version to version. Click any screenshot to enlarge it.

Versions 1–5: the interface and the workspace

The early versions established the conversational interface and the working screen that later versions keep. Versions 1–4 simulate their outputs; from version 5, the agents reason live through the Anthropic API. The case shown is an illustrative riparian ecology dataset.

Welcome screen: I am your Research Data Director, let me introduce you to the team, with three starting options

Entry

The welcome screen

  1. The Director speaks first, in the first person, and introduces the team rather than a form.
  2. Three ways in: ask a question, get help with a dataset, or take a guided look around.
  3. A standing promise: the researcher decides how much to see, when to act, and when to hand work over.
  4. Live and Demo modes and a Guide sit at top right; the footer repeats the principle that a human decides what matters.
Guide panel explaining how a case moves and where a person decides

Orientation

The guide: how this works

  1. “In one breath”: a plain-language summary of what the team does and why.
  2. “Built to be economical”: each specialist runs on the lightest model that does its job well.
  3. How a case moves, in six steps, from conversation to a signed audit trail.
  4. Where a person decides: a table of acts that always pause for a human, such as creating a permanent address or assigning a licence.
Workspace with the team panel, the Thinking Window narrating the work, a question for the researcher, and amber decision cards

Working screen

The workspace and the Thinking Window

  1. The team panel lists each specialist and what it is doing now.
  2. “Watching the team think”: the Thinking Window narrates each step in plain language as it happens.
  3. A question for the researcher, with three explicit choices: here, two plant names reclassified since fieldwork.
  4. Amber decision cards stop the work on sensitive items (landowner names in field notes; nesting sites of a listed bird species) and name the office each goes to.
  5. Shared notes count missing fields, items waiting for a person, and files read; output tabs run along the bottom.
The Fact Checker confirms identifiers and flags a dead dataset link with a proposed fix

Fact Checker

Verification in parallel

  1. The Fact Checker and the quality reviewer work in parallel, independently.
  2. Identifiers are confirmed against registries: the researcher, the institution, and the grant.
  3. A dead dataset link is flagged with a proposed fix rather than silently dropped.
  4. The researcher chooses what happens next: supply the link, remove the citation, or leave it flagged.
Finished dataset description with schema badges and a confidence rating on each field

Output: metadata record

The description, with sources and confidence

  1. The record is written in the field’s own standard, with a common discovery layer for a permanent address.
  2. Badges show the target schema, the discovery layer, an imagery catalog, and the intended archive.
  3. Every field shows where its value came from and how confident the team is.
  4. The location is marked low confidence: blurred to about a kilometre, pending approval.
Data dictionary with one row per variable: name, type, units, description, and missing-value code

Output: data dictionary

The data dictionary

  1. One row per variable: name, type, units, description, and missing-value code.
  2. It makes the columns themselves reusable, not just the dataset findable.
  3. Available to download as CSV.
Team panel listing six agents with model-tier badges and live status

Team status

The team panel

  1. Each agent carries a badge showing its model tier, so the cost of the work is visible.
  2. Live status for each role: routing to a person, brief ready, description drafted, one item flagged, checklist running, listening.
  3. The Recorder is always listening; it is never dispatched.

Version 11: the named persona team

Version 11 introduces the named personas specified in the Persona Appendix, a demo mode that runs without input, and a case file of formal reports.

Version 11 home screen with Watch the demo, three entry points, and six named team members

Entry

The front door

  1. “Watch the demo”: the full workflow runs in about 90 seconds with no input needed.
  2. Three entry points: share data, ask a question, or meet the team.
  3. Each specialist can be addressed directly by name: Tomas, Miriam, Priya, Daniel, Hannah, and Kenji.
  4. A standing assurance: nothing is shared or published without the researcher’s go-ahead.
The team working notes panel, showing timestamped entries from each named persona

Thinking Window (opt-in)

The team’s working notes

  1. The opt-in Thinking Window, now attributed to named personas.
  2. The Director reads the full intake and plans the sequence; sensitive fields are held until a person decides.
  3. The Recorder notes each human decision as it is made.
  4. The Metadata Librarian grounds terms in real vocabularies and flags what it cannot map rather than forcing it.
Case file listing a formal report from each specialist and a completed deposit message

Deliverables

The case file: reports from the team

  1. A formal note from each specialist, each written with the others in view: schema recommendation, case plan, description and crosswalk, verification report, FAIR assessment, provenance and audit trail.
  2. Every note can be read or downloaded; the whole set downloads at once.
  3. The case closes as a full deposit: findable, trustable, reusable.
  4. An optional narration script supports the voice experiments of version 12.

A variant application: the Medical Research Data Director (August 2026)

A very early prototype. The biomedical studio is a demonstration of metadata and crosswalk governance for research. It uses synthetic or de-identified data only, it is not for clinical diagnosis or treatment, and every AI suggestion requires human review. Figures shown in it are illustrative.
Medical research metadata prototype overview with the workflow and illustrative evidence panels

Studio overview

Overview

  1. A standing research-prototype notice heads every page.
  2. The left navigation follows the workflow, from a guided tour through crosswalks, review, and download.
  3. Summary counters: registered datasets, crosswalk assertions, standards represented, and provenance coverage.
  4. Illustrative panels show complementary blindness: the same cohort seen by different instruments.
  5. Four steps: describe, connect, check, reuse.
Crosswalk table mapping original fields to shared meanings and standard codes, with review status

Crosswalks

The crosswalk table

  1. A translation table: each original field or value is mapped to a shared meaning and a standard code.
  2. Each mapping carries evidence, a confidence level, and a review status: approved, in review, draft, or deprecated.
  3. Mappings are proposed for human adjudication, never applied silently.
Research and background page linking the argument, the blueprint, and the working prototype

Context

Research and background

  1. The central idea: discovery can depend on correspondence as much as on measurement.
  2. Instruments create measurements; metadata creates the connections that make them a coherent research record.
  3. The argument and the architecture: the Diagnostic Synthesis paper, the biomedical blueprint, and the working prototype.