Her Silent Data

From a gap in women's data to a study design.

Open the app · Watch the demo (3 min) · How it works

A platform for researchers who want to help close the data gap on women, but don't know which data to collect first. The name is the subject: data that exists in women's lives and was never recorded, so medicine cannot hear it.

What a record answers

1. What problem do women face?
2. What data is needed to solve it?
3. Has that data already been collected?
4. Has anyone collected it since the gap was named?
5. If not: which women should it come from, what should be found out from them, and in what form?

Most projects on this subject stop at "there is a gap." This one cites who named the gap, searches the live web to see whether it has been closed since, and — on request — drafts the study that would close it.

The rule the register rests on

Nothing is invented and stored. A record carries what its source said and nothing more. Where a source itself specified how and from whom to collect, that is recorded and cited; otherwise the field is null, which is the normal case. The validator rejects a guidance block with no source, so the rule is enforced in code rather than promised here.

What the platform generates, it labels and keeps out of the register. Pressing Show research design reveals a possible study design — who to recruit, what to measure, the breakdowns, the instrument — marked AI-generated study design — unverified with the date. It is an answer, not a finding: it lives in research-designs.js, never in data.js, and deleting that file leaves the app fully working.

It says Show and not Generate because the design was generated during the intake run, not when you click. The app is static and holds no API key (SPEC.md §13), so a button promising generation would be the one place the product overstated itself.

A design can be carried out of the app. Copy this design puts it on the clipboard as plain text, together with the data need, the gap claim, its source and date, and what the check found. The provenance travels with it on purpose: a design pasted somewhere without the claim it answers is orphaned model output, which is the thing this whole project exists not to produce.

Every record carries a check, and states its limit. A gap claim is dated, and a live web search records whether the data has been collected since. Where the check finds partial coverage the card names the dataset and says what it does not cover; where it finds the data in full, the candidate is not a gap and is not merged at all. No search proves a negative, so a missing record says on its face that this was a search of published sources, not proof, and invites a correction from anyone who knows of data it missed.

No record has status collected. A register of gaps that lists data which exists is listing the wrong thing.

How records arrive

intake/ holds the pipeline: discover, filter, extract, verify, validate, review — and then, only for records a person has approved and merged, design. Drafting a study for a candidate nobody admitted was money spent on an answer to a record that did not exist, so the design step is behind the human decision and refuses any id that is not already in data.js.

Discovery is plain code and costs nothing. Filtering is a yes/no read of an abstract, so it runs on the cheapest model that can do it; extraction, the check and the design run on a stronger one. Validation runs no model at all. The one step that is not automated is the one that decides: nothing reaches the register until a person has read the claim against its source and admitted it.

It is not part of the app — index.html loads nothing from it — and the app still opens from file:// with no server and no dependencies. See intake/README.md.

Status: the register holds 24 records across 6 areas — 8 where the data is missing, 16 where it is only partly covered — each with a dated check, and 21 carrying a generated study design. Three designs were withheld by the safety rules rather than repaired.

Run it

Open index.html — double-click the file, or visit https://minimarina.github.io/her-silent-data/.

There is no build step, no install and no server. Plain HTML, CSS and JavaScript, no dependencies — four screens and one seed file do not need a framework.

Data

Every problem cites a published source, and every data need cites the review or guideline that named the gap — systematic reviews, narrative reviews and guidelines published within the last twelve months. Every claim links to its source on the record itself; the rules every record must satisfy are in SPEC.md §9.

Every record has the same shape and arrived the same way. Five hand-sourced problems seeded this project and were removed on 18 Sep: they predated the rule above, carried an unsourced specification, and could not be brought into the current shape, so a reader would have seen two generations of record. They are in the git history.

Third-party assets

Two.

• Female body silhouette — OpenClipart #71126, dedicated to the public domain under CC0. https://freesvg.org/female-body-silhouette
• The GitHub mark in the footer — GitHub's trademark, used unaltered and in a single colour to link to GitHub, which is what their logo guidelines permit. Not covered by this repository's licence.

The silhouette's single <path> is pasted inline into index.html so the app still opens from file:// with no server and no network. The path itself is unaltered; only its fill and stroke are set, from the palette in SPEC.md §7.1.

Everything else — the layout, the markers, the ring, the data and all the code — is original work.

Licence

The code is MIT. The register — the selection, descriptions, verification notes and generated designs — is CC BY 4.0. The quoted sentences are not ours to license: each record reproduces one sentence from the paper it cites, to report what that paper claimed, and those stay the copyright of their publishers and authors. intake/validate.mjs rejects a quotation longer than 300 characters. See LICENSE.

Not medical advice

This is a research tool, written for researchers. Nothing in it is medical advice, and the study designs are generated, unreviewed model output.

Accessibility

Built to the rules in SPEC.md §8.2: every status carries a colour, a shape and a word, so it survives greyscale; the whole flow is keyboard reachable with a visible focus ring; body text starts at 16px and steps to 18px above 820px and 20px above 1180px, with captions at 14px and field labels at 13px, every one measured at 4.5:1 contrast or better against its own background; no horizontal scroll at 360px; prefers-reduced-motion is honoured.

The type scale is on :root, so it also follows the reader's own browser font-size setting rather than overriding it.

Hackathon

Built for the Elevate Women Global Hackathon 2026. All work in this repository was done during the Hackathon Period (Sep 14–20, 2026). No pre-existing code is used. Development was AI-assisted throughout, using Claude Code; the intake run calls Claude Haiku 4.5 and Claude Sonnet 5. The published site itself calls no model.