NIH Data Sharing Index Challenge · Phase 2 Finalist

SHAREscore

SHARE · S-Index · S-Impact

Make data the currency of science.

76M

datasets scored

140K

researchers ranked

live extensions

repositories

Built, live, and self-funded, today. · sharescore.org

sharescore.org · ConductScience Foundation · conductscience.com

About: open, shareable, FAIR data by default

ConductScience.com builds scientific tools

ConductScience.org maintains the standards

Shuhan He, MD

Founder, ConductScience

Faculty, Department of Emergency Medicine & Laboratory of Computer Science Mass General Hospital · Harvard Medical School

conductscience.com

1,200+

institutions served

280

verifiable methodology citations across 184 journals

NIH Replication
Prize Winner

2026

Cited in the methods of

CellNeuronNature CommunicationsScience AdvanceseLifeCell ReportsiScience+ 177 more

sharescore.org · ConductScience Foundation · conductscience.com

The vision

SHAREscore · core to the ConductScience mission

Already building the instruments & software that produce open, FAIR data at the source · SHAREscore = the missing layer that measures and rewards how well it's shared · turns a mandate into momentum.

Why this is the mission, not a side project

Instrumentation outputs FAIR, open data with provenance receipts at the source
Already ship ~2,900 instruments producing open, FAIR data at the source, at scale
SHAREscore = how we measure and reward it · sharing finally counts

sharescore.org · ConductScience Foundation · conductscience.com

How it works

Three linked metrics, built from standards the field already agrees on · SHARE grades the dataset · S-Index ranks the researcher · S-Impact weights it by real reuse.

SHARE grades the dataset

Every dataset scored 0–100 from 25 machine-checked signals · quality becomes visible, automated, comparable.

the dataset's score

S-Index ranks the researcher

max(k): k datasets each scoring ≥ k · the h-index for data · producing high-quality data finally counts.

S-Index

the h-index for data quality

S-Impact weights it by reuse

Citations earned by a researcher's S-Index core · tracks real reuse.

S-Impact

the reuse measure

sharescore.org · ConductScience Foundation · conductscience.com

Design principles

First principle: FAIR, open data deserves a FAIR, open standard · not one any company or consortium can own or gatekeep.

Measure what the researcher controls

Scored: documenting, formatting, licensing, linking at upload

Scored: facts a computer can verify

Not scored: citation luck, field size, topic popularity

Every signal earns its place with evidence

Scored: signals that empirically predict reuse across 76M+ objects

Not scored: anything a committee likes but data doesn't back

Rule: future changes must prove predictive validity

Governed so no one can capture it

Council: external-majority 6:1, owns the methodology

Open: signals, code, pledges fully forkable

Recourse: disagree and you can fork it

Grows and adapts, permissionlessly

Anyone submits: any community member can score a dataset — or a whole repository — not just its owner

Self-onboard: a repository joins with a one-PR pledge · no approval queue

New object types: code, software, protocols score on the same 25 signals, no rule change

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 5, Design principles

These were our design principles coming into this competition.

First: FAIR, open data deserves a FAIR, open standard, not one any company or consortium can own or gatekeep.

Then four axioms. One, measure what the researcher controls: SHARE scores what a depositor actually does at upload, the documenting, formatting, licensing, and linking a computer can verify, so credit reflects effort and stewardship, not citation luck, field size, or topic popularity, things outside their control.

Two, every signal earns its place with evidence: each of the 25 is validated against real reuse across 76 million objects, and any future change must prove improved predictive validity, so it's grounded in data, not committee opinion.

Three, governed so no one can capture it: an external council shepherds the standard, but the methodology, signals, code, and pledges are fully open and forkable, and if you disagree with the governance you can fork it, like Linux, so no company or consortium can own it or block a rival.

Four, grows and adapts permissionlessly: any community member can submit a dataset, or a whole repository, for scoring, not just its owner; any repository self-onboards with a one-PR pledge, no approval queue; and new object types like code, software, and protocols score on the same 25 signals with no rule change. It grows itself instead of ossifying into a rulebook that can't keep up with the science.

Where the 25 signals come from

Grounded in four standards chosen on explicit criteria, authoritative · cross-domain · machine-readable · adopted at scale

DataCite 4.6

2024 · the DOI standard

authoritative registry every research DOI already uses

Dublin Core

DCMI · ISO 15836

the cross-domain metadata foundation, ISO-ratified

schema.org/Dataset

web · Google

what powers Google Dataset Search, adoption at web scale

RDA FAIR Maturity

2020 · the community ref

the community's own yardstick for measuring FAIRness

64candidate fields

↓Standards convergence · in ≥ 2 of the 4 standards

↓FAIR mapping · maps to a FAIR sub-principle

↓Measurability · a computer can check it anywhere

↓Discriminative generalizability · separates good sharing from bad

25signals · 5×5

Five per dimension · not designed, what survived the funnel.

sharescore.org · ConductScience Foundation · conductscience.com

SHARE grades the dataset

SHARE = signals present / 25 × 100 · five dimensions, five machine-checked signals each · one big letter per dimension:

SStewardship

keywords

contributors

subjects

geo_context

temporal_context

HHarmonization

description

methods

references

contributor_pids

org_pids

AAccess

access_status

license_clarity

license_permissive

no_embargo

open_format

RReuse

discovery

access_events

citations

derivatives

community

EEngagement

related_pubs

related_data

funding

versioning

standards_ref

Every signal a binary check, present or absent, computed from public metadata. Each of the 25 worth 4 points: 25 × 4 = 100-point scale.

sharescore.org · ConductScience Foundation · conductscience.com

SHARE maps to FAIR at the signal level

All 25 signals map to FAIR, each under the letter it supports · also a map of data quality.

FFindable8

S1 · Geographic context F2

S2 · Temporal context F2

S3 · Contributor roles F2

H2 · Contributor PID F1

H5 · Description quality F2

S4 · Subject classification F2

R1 · Discovery / views outcome

R5 · Community engagement outcome

F3 · metadata name the data's ID archive-set

F4 · indexed in a searchable resource archive-set

AAccessible3

A1 · Open-access status A1

A4 · No embargo A1

R2 · Access / downloads outcome

A2 · metadata persist after data archive-set

Accessibility is mostly the archive's job (hosting, resolvable IDs), so few signals are depositor-controlled — which is exactly what keeps the score fair. License & format signals also count under Reusable and Interoperable.

IInteroperable7

S4 · Subject classification I2 · controlled vocab

H2 · Contributor PID I2

H3 · Organization PID I2

A5 · Open format I1

H4 · Bibliographic refs I3 · qualified

E1 · Related publication I3 · qualified

E2 · Related data I3 · qualified

RReusable10

A2 · License clarity R1.1

A3 · License permissive R1.1

H1 · Methods documentation R1.2

S3 · Contributor roles R1.2 · provenance

S5 · Ethical transparency R1.2

E3 · Funding information R1.2

E4 · Version tracking R1.2

E5 · Community-standard ref R1.3

R3 · Formal citations outcome

R4 · Derivative works outcome

All 15 FAIR sub-principles, bucketed. We score the 12 a researcher controls; the 3 greyed ones (F3, F4, A2) are set by the archive · scoring them would grade the repository, not the dataset. Violet = reuse outcomes, held out to validate the score.

sharescore.org · ConductScience Foundation · conductscience.com

From SHARE, two career metrics

Every dataset gets a SHARE score · from a researcher's scored datasets, two metrics follow.

every dataset
scored 0–100

S-Index · the researcher

How many high-scoring datasets

max(k): k datasets each scoring ≥ k on SHARE · the h-index, for data

e.g. 90 72 55 41 12 6 · 4 · 2 → S-Index = 6

S-Impact · the reuse

How much that core is cited

Σ citations of the S-Index core · empty deposits earn nothing

S-Impact = total citations of those 6 · 90 72 55 41 12 6

sharescore.org · ConductScience Foundation · conductscience.com

Section 3

The proof

Eight questions, answered on 183,872 datasets · and NIH's own registry. Next slide: how we tested it.

01Does SHARE at deposit predict reuse?

05Does it hold across fields?

02Does it hold every year?

06Does the S-Index reward quality?

03Does structure beat volume?

07Does S-Impact track influence?

04Is there a real dose-response?

08How much room to improve?

sharescore.org · ConductScience Foundation · conductscience.com

The proof · how we tested it

The method, in one slide.

What we did

Cohort. 183,872 Zenodo datasets deposited 2017–2022.

Predictor. SHARE at deposit, S + H + A + E only — the Reuse (R) bucket is held out.

Outcome (what we predict). Whether the data gets reused after deposit — a “derivative” (someone builds a new dataset on it) or a citation. Measured later than the score, so it can’t leak in. 407 of the 183,872 (0.22%) gained a derivative.

Model. Logistic regression, odds ratio per +10 points, adjusted for year, domain, and license.

Why it's clean

The reuse signals never enter the score.

We score the data at deposit, then watch what actually gets reused later. Because reuse is never part of that score, it can't secretly predict itself — no circularity, no leakage.

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 11, The method

One method slide before the results, so the panel trusts what follows. What we did: the cohort is 183,872 Zenodo datasets deposited 2017 to 2022, of which 407, about a fifth of a percent, were later reused, meaning someone published a new dataset built on them, what we call a derivative. The predictor is the SHARE score at deposit using only the S, H, A, and E buckets; the Reuse bucket, R, is held out. The outcome, meaning the thing we are predicting, is whether the dataset gets reused after deposit: we count derivatives (someone publishes a new dataset built on it) and citations, both measured later than the score, which is what keeps the outcome from leaking into the predictor. The model is a logistic regression, odds ratio per plus ten points, adjusted for deposit year, domain, and license. Why it's clean, in one line: we score the data at deposit, then watch what gets reused later, and the reuse signals never enter that score, so it cannot predict itself. No circularity, no leakage.

If a judge asks why the Reuse bucket exists in SHARE at all when we hold it out here: we hold R out only for this validation, to prove the deposit-time signals predict reuse without circularity. In the live SHARE score, R is included, because once a dataset has actually been reused, that realized reuse is genuine information about its quality. We separate them on purpose, drop R to make the prediction claim clean, keep R in the operational score to reflect complete quality, and the whole S-Impact researcher metric is built on that realized-reuse signal.

The proof · question 1

Does SHARE at deposit predict reuse?

5.73×

the odds of being reused

for every +10 points of SHARE score at deposit · 95% CI 4.97–6.61 · p < 0.001

183,872 Zenodo datasets deposited 2017–2022 · the outcome (a later derivative) is read years after the score

Answer

YES.

Better-documented data at deposit is far more likely to be reused later.

Citations move the same way: 3.02× per +10

p < 0.001 · Zenodo 2016 cohort, 8-year follow-up

Corroborated on NIH's own registry: ClinicalTrials.gov intent-to-share deposits score +17.2 points.

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 12, Q1: quality predicts reuse

Question 1: does deposit-time metadata quality, the SHARE score, predict later reuse? Answer: yes, decisively. Each +10 SHARE points multiplies the odds of a derivative dataset by 5.73× (95% CI 4.97–6.61, p<0.001) and the odds of citation by 3.02× (p<0.001, from the Zenodo 2016 cohort with 8 years of follow-up). Say "derivative" plainly: it means someone published a new dataset built on the original. In plain language: better-documented data at deposit is far more likely to be reused later, and because reuse is read after deposit, nothing leaks. NIH's own ClinicalTrials.gov registry corroborates it on a separate corpus: deposits that declare intent to share score 17.2 points higher.

Say this explicitly so no one tries to compound it: 5.73× is a per-+10 logistic odds ratio, a local rate, it does not multiply across the whole 0-to-80 range. The assumption-free way to read the full-range effect is the observed dose-response, a 32-fold gradient from the bottom to the top SHARE bin, which is the number I lead with if anyone challenges the odds ratio.

Methods, if asked, otherwise on the appendix slide: the predictor is measured at deposit and the outcome later, no leakage. Cohort, 183,872 Zenodo datasets deposited 2017–2022, of which 407 (0.22%) gained a derivative. Predictor is deposit-time SHARE using S+H+A+E only, the Reuse bucket held out. Model, logistic regression, odds ratio per +10 points, adjusted for year, domain, and license, and it holds within S-Index bins so it is not just that prolific authors deposit more. The 3.02× citation figure comes from a separate 2016 Zenodo cohort, n=48,771, 8 years of follow-up. The full robustness battery and the pre-registered author/lab-covariate and second-repository replication are on appendix slide 39.

The researcher-metric proofs now live in this section: the S-Index (rewards quality, not volume) and S-Impact (tracks influence, 1.60× senior authorship) are their own question slides later in the proof.

The proof · question 2

Is the effect consistent year over year?

Derivative odds ratio (per +10 SHARE points), by deposit year

Derivative odds ratio stable across deposit years 2017-2022

Answer

Yes — every year.

Near 5.5× per +10 SHARE points every deposit year tested, 2019 through 2021. Temporally stable, not one unusual year.

sharescore.org · ConductScience Foundation · conductscience.com

The proof · question 3

Does organizing the 25 signals beat just counting how many of the 25 are present?

0.930 AUC

five dimensions, each with its own weight

vs 0.848 if you sum all 25 into one score, or 0.846 for a plain count of how many are present. Letting the five dimensions carry separate weights is the whole +0.08 gain. AUC = how reliably the score separates reused from un-reused data (0.5 = chance, 1.0 = perfect)

Answer

Structure wins.

Counting asks how many of the 25 boxes are ticked. Structure asks which kinds — the five dimensions entered as separate signals. Two datasets with the same count but different profiles score differently, and that difference is what predicts reuse.

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 14, Q2: structure predicts reuse

Question 3: does the structure of the metadata predict reuse better than just counting how many fields are filled? Answer: structure wins. Here is the simplest way to say it. Take the same 183,872 datasets and try three ways to score them. One, a plain count of how many of the 25 fields are present: AUC 0.846. Two, sum all 25 into a single 0-to-100 number: 0.848, essentially no better. Three, keep the five dimensions separate and let each one carry its own weight: 0.930. That jump, from 0.848 to 0.930, is the whole point of the framework. AUC, in plain language, is just how reliably the score separates reused datasets from un-reused ones, 0.5 is a coin flip and 1.0 is perfect. The reason it matters: two datasets can tick the same number of boxes but have completely different profiles, twelve signals all in one dimension versus twelve spread across all five, and the structured score tells them apart while a count cannot. Adding the reuse outcomes on top only reaches 0.934, so the deposit-time structure is doing the work.

If a judge asks whether there is a single p-value for 0.930 versus 0.846: be honest, it is not one significance test, it is the comparison of nested logistic models on the same cohort, and the load-bearing point is that going from one summed score to five separately-weighted dimensions is what buys the gain, not extra noise-fitting. If they want it formalized, a DeLong test on the paired ROC curves would do it. NIH's own data corroborates descriptively: across ClinicalTrials.gov, trials declaring intent to share score 17.2 points higher, though that is corroboration, not a second independent validation, because intent-to-share is itself one of our Access signals. The load-bearing evidence is the leakage-controlled Zenodo analysis.

Is there a real dose-response?

New datasets built on the original (“derivatives”) per 1,000 deposits, by SHARE bandper 1,000 so bands of different sizes compare fairly · derivatives counted at the 2026 harvest, 4–9 years after deposit

Answer

Yes — a smooth curve.

Reuse climbs across every SHARE band, 32× more derivatives at the top than the bottom — a real dose-response that rises the whole way up.

A derivative is a new dataset someone built on the original — the strongest, hardest-to-fake form of reuse, well beyond a view or a download. That's why we measure it, not raw traffic.

sharescore.org · ConductScience Foundation · conductscience.com

The proof · question 5

Does it hold across fields?

Derivative odds ratio, ecology vs the rest

Derivative odds ratio by field: ecology vs other fields

Answer

Yes — every field.

Ecology is the strongest field (17.3×) but only 5,983 datasets — 3% of the cohort. Drop that 3%, and the other 177,889 (97%) still show 4.9×. Field-general.

We singled out ecology because its data-rich deposits are reused most — the toughest place to rule out a one-field artifact. Removing that small, strongest slice barely dents the effect, and the 4.9× still rests on 177,889 datasets, large and well-powered.

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 16, Stress test 3: field-general

Third check: is it an artifact of one data-rich field like ecology? No, it holds in every field. Why ecology specifically: ecology deposits are unusually data-rich and get reused the most, so it's the single hardest place to rule out a one-field artifact, if the effect were going to be carried by one field, it would be this one. It's strongest there at 17.3×, but exclude ecology entirely and the other 177,889 datasets still show 4.9×. On the natural follow-up, does 17.3 versus 4.9 mean the other fields are under-tested: no. The 4.9× is estimated on 177,889 datasets, a large, well-powered sample, so it's a real, robust effect; the gap just means the effect size is larger in data-rich ecology than elsewhere, not that other fields lack data. Bottom line across all three checks: the relationship is real, dose-dependent, temporally stable, and field-general. The full robustness battery, and the pre-registered author/lab-covariate and second-repository replication, are on the appendix slide.

The proof · question 6

Does the S-Index reward quality, or just volume?

S-Index = max(k)

datasets sorted by SHARE · k is where the bars still clear the k×k line · solid bars count, faded ones don't

2,517

most datasets one
depositor has

highest S-Index
observed · avg SHARE 53

Answer

Quality.

The S-Index is max(k): the largest k where a researcher has k datasets each scoring at least k. It's the h-index logic applied to data quality, so 2,000 sloppy deposits can't inflate it.

Complementary, not redundant: pre-registered to correlate with the H-Index; the reuse-weighted S-Impact then shows it tracks real influence — next slide.

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 17, S-Index rewards quality not volume

Question 6 is the researcher metric we propose, the S-Index. Does it reward quality or just volume? Answer: quality. The S-Index is max(k), the largest k where a researcher has k datasets each scoring at least k on SHARE, exactly the h-index logic applied to data quality. The proof it is not a volume count: the most prolific depositor in the corpus has 2,517 datasets, but the highest S-Index anyone reaches is 56, because piling on low-scoring deposits cannot raise max(k), each dataset has to clear its own bar. It is pre-registered to correlate with the H-Index, so it is complementary to citation impact, not redundant, and the very next slide, S-Impact, shows the reuse-weighted version tracks real scientific influence with the 1.60x senior-authorship result.

If asked what that top researcher's actual scores are: the highest-S-Index researcher reaches 56 on roughly 882 datasets with an average SHARE around 53, while the most prolific depositors sit near an average SHARE in the low 40s with an S-Index well below the maximum. So volume and quality are genuinely decoupled, the person with the most datasets is not the person with the highest S-Index. These per-researcher figures are from the live platform, not the frozen submission, so cite them as current, not as submitted numbers.

The proof · question 7

Does S-Impact track real influence?

S-Impact = total citations of a researcher's S-Index core datasets

1.60×

the odds of senior authorship

on highly-cited papers · for researchers with S-Impact ≥ 5 · p = 0.003

Answer

Yes.

S-Impact counts how often a researcher's best data is actually cited. The independent check: those with S-Impact ≥ 5 are 1.60× more likely to be senior authors on highly-cited papers.

Senior authorship is an outside marker of scientific leadership — nothing to do with our score — so the link isn't circular. The people whose data gets reused are the ones leading influential work.

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 18, S-Impact tracks influence

Question 7 validates the researcher-level metric. First, what S-Impact is: the total number of citations earned by a researcher's S-Index core, the datasets that make up their S-Index. So it already measures how much a researcher's best data gets used. The logical step the panel needs: we then check that number against something completely outside our own scoring, namely whether these researchers are senior authors on highly-cited papers. They are, 1.60 times the odds for those with an S-Impact of 5 or more, p = 0.003. That is the point of using senior authorship, it is an external marker of scientific leadership with no connection to our metric, so the correlation is not circular, the people whose data actually gets reused turn out to be the ones leading influential science.

Honest framing if pressed: senior authorship is the one independent validation of S-Impact we have in the submission, and it is a summary result without a full published model, so lead with the mechanism, S-Impact is citations of the core, and offer the 1.60x as the corroborating external check rather than the whole case. Keep this brief, it complements the dataset-level 5.73x.

The proof · question 8

How much room is there to improve?

Answer

A lot.

Scores are widely dispersed — a 56-point spread, 29 to 85 across repositories, mean about 45, so most data sits far below what's achievable. And the spread is a curation choice, not a budget one: SRA is at 85, NASA at 29, both large and well-funded — so any repository can move up.

Average SHARE by repository · DB-authoritative, v20

SRA

GEO

EDI

OpenAIRE

ClinicalTrials.gov

OpenNeuro

Dryad

Zenodo

NASA

Why it's curation, not budget

Same resources, different scores

SRA (85) and NASA (29) are both large, well-funded repositories — the 56-point gap is metadata policy, not money.

A choice, not a constraint

requiring richer metadata at deposit is a policy any repository can adopt — which is exactly what the nudges push.

Mid-scale = the target

the breadth aggregators (44–56) have the most room to move.

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 19, How scores distribute

Lead with the takeaway: scores are widely dispersed, 29 to 85 across repositories with a global depositor mean around 45, which means most data sits well below what's achievable, there is a lot of room to improve. That dispersion is the opportunity the whole platform targets. Then the reason the spread exists: it's curation intensity, not budget or field size. Average SHARE runs from SRA at 85 and GEO at 74, controlled-access and richly annotated, down through the mid-scale breadth aggregators (OpenAIRE 54, ClinicalTrials 50, Zenodo 44) to NASA at 29, and SRA and NASA are both large and well-funded, so the 56-point gap is metadata policy, not money. The mid-scale aggregators have the most headroom and are exactly what the nudges target. The geographic fairness evidence, that the score is unbiased across countries, lives on the global-reach slide.

Section 4

The live platform

Not a proposal — a running platform, already built and live at sharescore.org.

76M+datasets scored

140K+researchers ranked

500+institutions, 179 countries

APIopen, no-auth, over all 76M+

12live extensions

Self-serveany repo can pledge & be scored

The next slides walk the live surfaces — the global map, the leaderboards, the API, the extensions.

sharescore.org · ConductScience Foundation · conductscience.com

Global reach · /map

179 countries scored, and the score is unbiased. Country researcher-count barely predicts a depositor's score (r = 0.003); Global South 44.7 vs 44.6 worldwide; widest continental gap 1.7 points; dataset-volume barely predicts score (r = +0.02). Quality wins, not resources or size.

sharescore.org · ConductScience Foundation · conductscience.com

Researcher leaderboard · /researchers

Every researcher ranked by S-Index · the h-index for data, made personal · a policy becomes a scoreboard people care about.

Researcher leaderboard ranked by S-Index

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 22, Researcher leaderboard

/researchers. The h-index-for-data made personal, every researcher ranked by S-Index. This turns a policy into a scoreboard people care about.

Be ready for the disambiguation question, and answer it honestly, it is the one real soft spot in the S-Index. How do we know which datasets belong to the same person? ORCID exact-match where it exists, but ORCID coverage in deposit metadata is low (about 0.4% in our aggregated set), so for almost everyone we block by last name plus first initial and cluster with fuzzy name matching at a similarity threshold of 88, boosted when affiliations agree. That means the researcher-level S-Index carries genuine disambiguation error, false merges of same-named people, false splits across name variants, and it is currently strongest within a single repository, since cross-repository linking is still small. The reason this does not undermine the science: every validation result, the 5.73x, the AUC ladder, all of it, is computed at the dataset level on the SHARE score, which needs no author disambiguation at all. So SHARE is the rigorous core, and the S-Index is a researcher-level indicator whose disambiguation we are still hardening. That is the honest framing; do not oversell the leaderboard's precision.

Individual researcher profile

Every researcher: a public, explainable profile. S-Index, S-Impact, deposit-time scores, a SHARE-dimension radar, toggleable extensions. Not an abstraction; a page you can visit today.

sharescore.org · ConductScience Foundation · conductscience.com

Institutional leaderboard · /institutions

Rolls up to institutions, too · 500+ ranked by average S-Index, SHARE score, or data-sharing volume, across 179 countries · the same permissionless score, aggregated to where funding and hiring decisions are made.

sharescore.org · ConductScience Foundation · conductscience.com

Live accountability dashboard · /metrics

A real-time scoreboard for the 2023 data-sharing mandate · a program officer sees whether the policy is producing reusable data today, not in a year-end PDF.

What the Metrics tab shows

Mean SHARE score across all 76M+
Datasets scored & repositories pledged
Score trend, up or down over time
Score distribution + SHARE-dimension breakdown

Fully automated rollout

Weekly new-repository scoring
Daily automated tracking
Quarterly public reports

sharescore.org · ConductScience Foundation · conductscience.com

Open API · api.sharescore.org

API live today, what's possible? A no-auth REST API over all 76M+ scored objects. Three things anyone can already do:

Any publisher

drop SHARE (dataset), S-Index (researcher) & S-Impact onto any page, or expose their own repository's SHARE scores

Any repository

be scored end-to-end, giving every researcher in it a live score

Any researcher

look up any other researcher's S-Index, right now

Because it's open, others build on it — how a standard spreads:

A journal · pre-submission SHARE check · authors fix weak metadata before peer review

A university · S-Index in promotion & tenure packets

A funder · applicants ranked by data-sharing track record

A grant reviewer · proposals triaged by past data quality

An AI research agent · datasets ranked by SHARE, reusable data surfaced first

A lab · dashboard benchmarked against peers

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 26, Public API

This slide is the vision of what an open API unlocks, and the top row is not hypothetical, it's live today. Any publisher can drop SHARE, S-Index, and S-Impact onto any dataset or researcher page, or expose their own repository's SHARE scores. Any repository can be scored end-to-end so every researcher in it gets a live score. And any researcher can look up any other researcher's S-Index right now. Then, because it's open and no-auth over all 76M+ objects, others build on it: journals running a pre-submission SHARE check so authors fix weak metadata before review, universities adding the S-Index to promotion packets, funders and grant reviewers triaging by data-sharing track record, AI research agents ranking datasets by SHARE to surface the most reusable ones, labs benchmarking against peers. The standard spreads by being built on.

Extensions · /extensions, 12 live tools

Twelve lenses ship live · each toggleable in a researcher's own workflow

SHARE Star, overall rating

Originality Lens, anti-duplicate

Compliance Checker, policy check

Data Substance Validator, ghost-data

Pre-submission Feedback, pre-deposit fix

Badges, embeddable proof

Badge Generator, shareable badge

Trend Tracker, over time

Deposit Rate Monitor, activity

SHARE-T, temporal trends

SHARE-C, peer benchmarking

Field Normalizer, field percentile

Example, the Originality Lens (anti-gaming)

Worked example, a 60-dataset profile: S-Index holds at 52 → 52 · all 60 contribute unique concepts, no series inflation · near-duplicate dumps would collapse the unique-concept score, genuine breadth doesn't.

Integrity, in the workflow: several lenses harden it directly · Originality (anti-duplicate) · Data Substance Validator (ghost-data detection) · Compliance Checker (policy & transparency).

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 27, Extensions

The extension store on the left, twelve live tools, each a lens a researcher can toggle in their own workflow. Read the list: SHARE Star, Originality Lens, Compliance Checker, Data Substance Validator, Pre-submission Feedback, Badges, Badge Generator, Trend Tracker, Deposit Rate Monitor, the SHARE-T and SHARE-C analytics, and the new Field Normalizer, which places a dataset's SHARE score on the bell curve of its own research field, a percentile among discipline peers, since sharing norms differ by field. The worked example is the Originality Lens: on a 60-dataset profile the S-Index holds at 52 to 52, all 60 datasets contribute unique concepts, so no series inflation. If instead most deposits were near-duplicate dumps, the unique-concept S-Index would fall below the raw one, and that gap is exactly what the lens surfaces, real contribution, not volume. And this is where a lot of our integrity story lives operationally: Originality is anti-duplicate, the Data Substance Validator does ghost-data detection, the Compliance Checker handles policy and transparency, gaming defenses that ship as usable tools, not just Q&A answers.

Section 5

Implementation

How the standard actually runs, four ways:

1A competitive mechanism that drives adoption

2Governed so no one can capture it

3Changed only by an open process

4Funded to outlast any grant

sharescore.org · ConductScience Foundation · conductscience.com

Compliance becomes a competition

A mandate sets a floor · publishing every repository's SHARE ceiling turns compliance into a race to the top · adoption compounds on its own, no new enforcement.

A competitive ecosystem · the ceiling mechanism

Repository A

15/25 signals

Repository B

20/25 signals

100

Repository C

25/25 → deposits flow here

SHARE_max = signals ÷ 25 × 100 · each new signal a repository supports raises its depositors' ceiling +4 points · repositories compete to host the best data. (numbers illustrative)

sharescore.org · ConductScience Foundation · conductscience.com

Adoption mechanism · both paths work

Anyone can get a repository scored.

Repository pledge

The repository itself

An official maintainer maps the repository's metadata fields to the 25 SHARE signals and we verify them.

→ Confirmed score

Community pledge

Any researcher or third party

Anyone can propose the mapping in one pull request; it scores preliminary immediately and is verified later.

→ Preliminary score, then confirmed

The add-repo wizard · live at sharescore.org/add-repo, no JSON required

Repository info

API configuration

Signal mappings · 0 of 25

Review & verify

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 30, Adoption mechanism

This is how the standard actually spreads, and there are two paths that both work. A repository pledge is the official one: a maintainer maps the repository's metadata fields to the 25 SHARE signals, we verify it, and it earns a confirmed score. A community pledge is permissionless: any researcher or third party can propose the mapping in a single pull request, it scores preliminary right away, and it's verified afterward. So no repository has to opt in first for its data to start being scored, and nobody is gatekept out.

The mechanics are a live product, not a promise: the add-repo wizard at sharescore.org/add-repo walks four steps, repository info, API configuration, signal mappings against the 25 signals, and review and verify, with no JSON knowledge required. Attestation is what separates a confirmed score from a preliminary one, so the leaderboard stays honest while still letting anyone start the process.

ConductScience Foundation · open governance and the COI firewall

1 · External-majority control

Council owns all methodology at 6 external : 1 internal · founder the only ConductScience seat · no single party, including us, can move a score.

2 · No revenue depends on scores

$0 of ConductScience revenue tied to any SHARE score or repository relationship · no incentive to game the metric · commercial support capped at 50%.

3 · Fully open & forkable

Rubric CC-BY, code Apache-2.0, all pledges public, anyone can fork it · fiscal sponsor: HTWB 501(c)(3) (EIN 88-3335809), independent of ConductScience.

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 31, Open governance + COI firewall

Modeled on the Linux Foundation. The org is the ConductScience Foundation, fiscally sponsored by HTWB, an independent 501(c)(3) (EIN 88-3335809), so we get charitable status without ConductScience controlling the money. Three tiers: a Board for fiscal and legal oversight; a SHARE Governance Council that owns signal vocabulary, scoring methodology, pledge policies, and major versions; and an Operations tier on open infrastructure. Explain the Council by the seats, not the names: six external seats each represent an independent constituency, academic medicine, health IT, industry, clinical research, the open research community, and trainees, against one internal founder seat, so the six-to-one external majority is structural, not a favor.

Then answer the conflict-of-interest question head-on, because a commercial-adjacent foundation asking NIH to recognize a metric invites it. Three firewalls. One: the Council controls all methodology at six external seats to one internal, the founder is the only ConductScience seat (Goldstein Harvard/MGH, Liebovitz Northwestern, Marshall Google Health, Goss Mass Eye and Ear, an open research-community seat, an AMSA trainee seat), so no single party, including us, can move a score. Two: zero ConductScience revenue depends on any SHARE score or repository relationship, and commercial support is capped at 50% of the funding mix, so there is no incentive to game it. Three: rubric CC-BY, code Apache-2.0, every pledge public and forkable, with HTWB (EIN 88-3335809) as the independent 501(c)(3) fiscal sponsor. If you don't trust us, fork it, that option is the ultimate check.

How changes happen, one open, evidence-based process

Proposal

Anyone opens a GitHub proposal: rationale + expected impact on scores.

60-day comment

Public for a minimum 60-day comment period.

Evidence

Must show improved accuracy, better reuse prediction or fewer gaming vectors.

Council review

External-majority Council weighs proposal, comments, evidence.

Supermajority

Minor = simple majority; signal or denominator changes = ⅔ + Board ratification.

Public changelog

Approved changes ship with rationale, effective date, and migration notes.

Worked examples (NIH updates the DMS Policy, a new gaming vector, a community-proposed signal) and the ratchet & migration rules for existing scores: in the appendix.

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 32, How changes happen

Changes follow one fixed open process: a GitHub proposal with the expected score impact; a 60-day public comment period; evidence of improved accuracy; external-majority Council review; a supermajority vote (⅔ plus Board ratification for signal or denominator changes); and a public changelog. The examples make it concrete: if NIH updates the DMS Policy, anyone opens an amendment and it runs this exact process; if a new gaming vector appears, a signal is added or adjusted by a ⅔ vote; if the community proposes a signal, it goes through 60 days of comment and must show better reuse prediction. If someone asks the natural follow-up, what happens to existing scores when the rubric changes, that ratchet-and-migration story is its own appendix/FAQ slide, so pull it up only if pressed rather than covering it here.

Next Steps · the growth loop & roadmap to one billion, mostly live Monday

76Mscored today, frozen for the competition

repositories added (x) → objects reached (y) · same pipeline · illustrative

1B+objects · by end of 2026 · re3data 3,000+ vs our 9

Next datasets in the pipeline Figshare · Mendeley Data · Harvard Dataverse · PANGAEA · ICPSR · UK Data Service · OSF · biomedical tier: GEO · SRA · dbGaP · Europe PMC

Q3 2026

9 → 20 repositories
Submit to ODSS as a DMS-Policy reference metric
Ship FAIR-mapping report; nudge top-100 repos

Q4 2026

20 → 35 repos
First journal to require a SHARE score at data submission
Begin post-incentive re-validation of 5.73×

Q1 2027 · flagship

50+ repos · 1 billion objects
Publish the SHAREscore paper
Align with DMS-Policy reporting (the ask)

Success measures · composition-adjusted:

repos & researchers notified

correction / claim conversion

composition-adjusted SHARE lift

NIH-relevant datasets improved

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 33, Roadmap to one billion

This is the Next Steps answer on one timeline. Trajectory first: 76 million objects are scored today, frozen for the finalist competition, with improvements already staged and ready to ship. We aim to reach 1 billion objects scored by the end of 2026, and the same public-metadata pipeline is what makes that reachable; the small graph shows the climb as repositories are added, since re3data lists more than 3,000 repositories against our 9. The next datasets in the pipeline, listed on the slide, are the obvious high-volume sources (Figshare, Mendeley Data, Harvard Dataverse, PANGAEA, ICPSR, the UK Data Service, and the Open Science Framework), which alone roughly doubles coverage, with the biomedical tier (GEO, SRA, dbGaP, Europe PMC) close behind. Then the quarters. Q3 2026: widen to 20 repositories, and take the concrete engagement step of submitting SHAREscore to the Office of Data Science Strategy as a DMS-Policy reference metric, plus ship the FAIR-mapping report and nudge the top-100 repos. Q4 2026: 20 to 35 repos, the first journal to require a SHARE score at data submission, meaning a journal adds a SHARE check to its data-availability step, and begin post-incentive re-validation of the 5.73x, which is the validation-hardening the biostatistician asked for, pulled earlier than before. Q1 2027 is the flagship: 50-plus repos reaching a billion objects, we publish the SHAREscore paper, and we align nudges and reporting with what the DMS Policy needs, which is where the ask lands.

The bottom row is how we will know it is working, composition-adjusted measures rather than a gameable raw mean: repos and researchers notified, correction and claim conversion, composition-adjusted SHARE lift, and the count of NIH-relevant datasets improved. And the whole thing needs no external partner to sign first: the scoring and nudges already run on Monday.

The engine behind the trajectory is a five-step growth loop, first two steps already live: one, score everyone from public metadata, no opt-in, 76M+ today; two, send citation-benchmarked nudges ("a similar repository has better metadata and more citations, here are the exact signals to fix"); three, claim and re-score via the one-PR pledge; four, lead by example, ConductScience deposits exemplar datasets, disclosed and excluded from every comparative benchmark; five, measure composition-adjusted lift and feed it back. We reach researchers by notifying repositories of their standing, building SHAREscore into AMSA (~30,000) and FUN (~540) curricula, our own audience at conductscience.com (~22,000 sessions/month), and letters on file from MGB Biobank, Dr. Jonathan Rosand, and Google for Health.

Dissemination partners · getting SHAREscore into curricula and clinics

Where the next generation learns to value shared data.

~30K

AMSA

American Medical Student Association · SHAREscore built into the teaching curriculum for medical students.

~540

FUN

Faculty for Undergraduate Neuroscience · ~540 faculty carrying it into undergraduate neuroscience programs.

MGB

Biobank

Mass General Brigham Biobank · biomedical integration, letter of support on file.

Plus our own reach at conductscience.com (~22,000 sessions/month) · additional letters of support on file from Dr. Jonathan Rosand and Google for Health.

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 34, Dissemination partners

Adoption isn't only repositories, it's people, so here are the partners who put SHAREscore in front of the next generation. AMSA, the American Medical Student Association, reaches roughly 30,000 medical students, and SHAREscore is built into the teaching curriculum. FUN, Faculty for Undergraduate Neuroscience, is about 540 faculty who carry it into undergraduate neuroscience programs. And Mass General Brigham Biobank gives us biomedical integration, with a letter of support on file.

On top of that, our own audience at conductscience.com runs about 22,000 sessions a month, and we hold additional letters of support from Dr. Jonathan Rosand and Google for Health. The point for the panel: the standard is already being taught and adopted, not waiting on one big partner to sign.

The ask

We don't need funding. We need recognition.

Already built, live, self-funded · scoring 76M+ datasets across 9 repositories, including NIH's own GEO and SRA · what accelerates it: NIH putting its weight behind a standard, not money.

RECOGNIZE

the NIH Office of Data Science Strategy (ODSS) names the S-Index as a reference metric aligned with the 2023 Data Management & Sharing (DMS) Policy

ANNOUNCE

co-announce it to the research community, press & guidance reach

POINT

direct NIH-repository grantees (GEO, SRA, already scored) to their live SHARE scores

Runs with or without NIH · Apache-2.0, forkable, reproducible from public code · recognition accelerates a flywheel that already turns.

Zero cost. Zero risk. · sharescore.org · Shuhan He, MD

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 35, The ask

The ask is deliberately not for money. It's already built, live, and self-funded, scoring 76M+ datasets across 9 repositories including NIH's own GEO and SRA. What accelerates a standard is NIH's endorsement: recognize the S-Index as a reference metric aligned with the 2023 Data Management and Sharing Policy; co-announce it for reach; and point NIH-repository grantees to their live SHARE scores. Zero cost, zero risk.

If a program officer asks "which office, and have you talked to them?" be honest and disarming. The natural home is the NIH Office of Data Science Strategy (ODSS) in the Office of the Director, which stewards the NIH-wide DMS Policy; near term, the National Eye Institute (NEI) administers this Challenge on NIH's behalf. We have not yet sought a formal endorsement, and by design we don't need one: this is a recognition ask for an open, forkable standard, not a contract or a procurement. It runs with or without NIH (Apache-2.0, reproducible from public code), so there is nothing to vet and no dependency to create. Recognition just accelerates a flywheel that already turns. Naming ODSS is naming the appropriate home, not claiming a relationship we don't yet have.

Thank you

The team that builds and runs it, eleven people across medicine, computer science, bioinformatics, math, and machine learning.

Shuhan He, MD

Informaticist · Lead

Yijian Henry He, PhD

Computer Science · CTO

Louise Corscadden, PhD

Community Engagement

Allison Goff, PhD

Bioinformatics

Boyu Peng, MS

Computer Science

Santosh Adhikari, MS

Machine Learning

Lawrence Jiang, BS

Computer Science

Shivani Pimparkar, MS

Bioinformatics

Pedram Safari, PhD

Mathematics

Yaning "Abby" Zheng

Scientific Forecasting

Xinhui Qian, MS

Machine Learning

Operational team, not owners. The ConductScience Foundation 501(c)(3) = a separate governance council controlling the methodology · framework, code, and pledges all open.

sharescore.org · ConductScience Foundation · conductscience.com

Section 8

Appendix

sharescore.org · ConductScience Foundation · conductscience.com · Appendix

Appendix · design philosophy & NIH-criteria mapping

Design philosophy

Deposit-time, not downloads

S·H·A·E come from the deposit · reuse held out to validate · you can't buy your score.

One rubric, everywhere · a ratchet, never a penalty

25 binary signals on any repository · a score never falls · improving is pure upside.

It shows its work

Every score traces to the exact signals behind it.

Covers what NIH asked for

NIH's criteria

FAIR adherence

SHA

Timeliness of sharing

Annotation & completeness

Frequency of reuse

Downstream influence

sharescore.org · ConductScience Foundation · conductscience.com · Appendix

Appendix · why the 5.73× holds, and what we've pre-registered

The derivative-reuse effect survives a full robustness battery · the checks we do not yet have are pre-registered, with a published-negatives commitment · honest by design.

It's not an artifact

Dose–response (per 1,000)0 → 15.06

…top vs bottom bin32×

Temporal 2019 / 2020 / 20215.49 / 5.37 / 5.83

Ecology17.30×

Non-ecology4.91×

Monotonic, temporally stable, and field-general, not a threshold, a fluke year, or one data-rich field.

It's not just prolific authors

Derivative rate, S-Index 11–20 bin37.4% vs 10.4%

…significanceχ²=97.4

Holding researcher productivity fixed (same S-Index bin), consistent sharers still create far more derivatives · not a productivity proxy.

Triangulated across 3 outcomes

Derivative (IsSourceOf)5.73×

Any relatedIdentifier10.04×

Citation > 03.58×

2017–2022 cohort (n=183,872); citation OR here is "any citation > 0," distinct from the 2016 8-yr-follow-up cohort's 3.0×. All predictors VIF < 2.5, no collinearity concern.

Pre-registered next, what we don't yet have

Author/lab-covariate model. Published OR adjusts for year, domain, license only · will add prior productivity, dataset size/richness, funding as covariates (fixed-effects / matched).

Second-repository replication. Reuse OR Zenodo-only today · Phase C-1 replicates in MGB Biobank (165K records), Dryad (DataCite Event Data), dbGaP (NLM access statistics).

Pre-registered bar: success = OR ≥ 3.0; negative results published either way. Q4 2026 – Q1 2027.

sharescore.org · ConductScience Foundation · conductscience.com · Appendix

Speaker notes · Slide 39, Appendix: robustness & pre-registration

Reserve slide for the biostatistician who presses on causality. Three columns. One, it is not an artifact: a monotonic 32-fold dose-response, odds ratios stable at 5.37 to 5.83 across deposit years, and it holds in both ecology (17.30) and everything else (4.91). Two, it is not just prolific authors: within the S-Index 11–20 bin, consistent sharers create derivatives 37.4% of the time versus 10.4% (χ²=97.4; this is the bin where the contrast is significant), and the effect triangulates across three mechanistically independent outcomes (derivative 5.73, any related identifier 10.04, citation 3.58), all with VIF under 2.5. Three, be honest about what is not yet done and is pre-registered: the current model adjusts only for year, domain, and license, so we have pre-registered an author and lab covariate model (prior productivity, dataset size, funding) and a genuine second-repository replication of the reuse odds ratio in MGB Biobank, Dryad, and dbGaP, with a success criterion of OR at least 3.0 and a commitment to publish negative results either way, in Q4 2026 to Q1 2027. Pre-registration with published negatives is the credible posture, it is the opposite of cherry-picking.

Appendix · when the rubric changes, the ratchet & migration rules

The natural follow-up to the change process: what happens to a score you already earned when the rubric evolves? Four rules, and one deliberate exception.

The rules, stability you can trust

Every score is version-stamped. Always tagged with the rubric it was computed under (e.g. v20) · a 2026 score stays comparable to a 2036 score · never apples-to-oranges.

An honest score never falls, the ratchet. New signals only raise ceilings · improving your deposit is pure upside, never a penalty.

Old & new stay comparable. Every version ships published migration notes · records dual-scored across a transition window · trend lines don't break.

Recomputation is deterministic. Scoring a pure function of public metadata · anyone can re-run any version and reproduce the number.

The math. Shown score = max(old-rubric score, new-rubric score) · a new signal can raise the ceiling 84 → 86 · a stricter definition computing 80 leaves the earned 84 intact · no dataset's SHARE ever falls · so the S-Index (largest k with k datasets at SHARE ≥ k) is monotonic non-decreasing across every version · only exception below: a closed exploit.

The one exception

Closing an exploit re-scores, and flags the records that used it.

If a gaming vector is found, we don't grandfather the scores that exploited it · those records re-scored under the fix and flagged, not quietly lowered · in that single case, integrity beats stability · fully transparent in the changelog.

sharescore.org · ConductScience Foundation · conductscience.com · Appendix

Appendix · built to scale, 76M today, 1B+ addressable

Why it scales, the architecture

Public metadata only. Harvest via OAI-PMH, DataCite, OpenAIRE, and open repository APIs (GBIF, NASA CMR, EBI/NCBI), no repository integration, no private access, no opt-in required.

Scoring is O(1) per record. 25 binary signals via a stateless function, embarrassingly parallel and incremental. 76M+ objects scored on commodity infrastructure.

Onboarding a repo is a bounded schema map. ~1–2 person-days each, mostly reused across repositories on the same standard (DataCite, Dublin Core).

No partner has to sign first. Coverage grows unilaterally from public data; pledges only ever raise ceilings.

The addressable ceiling

76M frozen for this competition (~96% OpenAIRE) · all genuine data repositories, not paper indexes · the ceiling depends on the counting unit:

Dataset level · tens of millions

OpenAIRE dataset subset

~93M

DataCite dataset DOIs (subset)

tens of M

Figshare · Dryad · Dataverse · PANGAEA

~10M+

Record / observation level · crosses 1B

GBIF species occurrences

~3B

NASA EarthData granules

1B+

EBI ENA sequence records

billions

1B+

scorable objects from public metadata. Tens of millions at the dataset level; GBIF and NASA EarthData each clear a billion on their own at the record level.

sharescore.org · ConductScience Foundation · conductscience.com · Appendix

Speaker notes · Slide 41, Appendix: scalability

The reserve answer on scale. It scales because of four architectural facts: we use public metadata only, harvested via OAI-PMH, DataCite, Crossref, OpenAIRE, so there's no integration or opt-in gate; scoring is O(1) per record, a stateless function of 25 binary signals, so it's embarrassingly parallel and 76M objects run on commodity infra; onboarding a new repository is a bounded schema map of one to two person-days, mostly reused across repos on the same standard; and no partner has to sign first. The addressable ceiling, and I want to be honest about the counting unit because it matters: these are all genuine data repositories, not paper indexes like Crossref or Europe PMC. At the dataset level, the world's data repositories total tens of millions of dataset DOIs, OpenAIRE's ~93M dataset subset is the largest single pool, plus Figshare, Dryad, Dataverse, and PANGAEA. You cross a billion only when you count at the record or observation level that these same repositories already expose in public metadata: GBIF has about 3 billion species-occurrence records, NASA EarthData over a billion granules, and the European Nucleotide Archive billions of sequence records. GBIF and NASA EarthData each clear a billion on their own. So the billion-plus is real, and it lives at the record level, not by summing paper counts.

Appendix · the live platform, far more than a leaderboard

Datasets browser76M+ records, every one scored & searchable by DOI, repo, reuse

S-Impactthe citation-outcome twin of the S-Index, reuse, not just deposit

Conversational interfacequery mappings & propose pledges in plain language

Pledge registryevery repository's public signal commitment, auditable

Add a repositorypermissionless onboarding, one pull request, no gatekeeper

Live coveragewhat's scored across 9 repositories, updated continuously

sharescore.org · ConductScience Foundation · conductscience.com · Appendix

Appendix · anti-gaming, why the score survives optimization

The gaming move

Why it fails

Mark signals "not applicable"

Fixed /25 denominator. Skipping lowers the score.

Blame the repository

Unsupported scores 0. No free pass.

Repository-shop for a soft grader

Same metadata, same score everywhere.

Hide in a low-scoring venue

A 60 is a 60 anywhere. Ceilings published.

Get an insider deal

Self-service pledge, community-validated.

Claim signals you don't satisfy

Attestation + PR pledge. Checked against metadata.

sharescore.org · ConductScience Foundation · conductscience.com

Speaker notes · Slide 43, Anti-gaming

Reserve slide for the anti-gaming question. Read the table as a scientist trying to pump their own score. Claim "not applicable": fill a few signals and mark the rest as not applicable, but the denominator is fixed at 25, so skipping just lowers the score. "My repo can't do that": unsupported scores 0. Repository-shop for a lenient grader: same metadata scores the same everywhere. Hide in a venue that caps everyone low: scores are absolute, a 60 is a 60, and every ceiling is public. Insider deal: self-service, no gatekeeper. Just claim it: attestation checked against public metadata. The broader "what if everyone games it" Goodhart/LLM-metadata question has its own reserve FAQ slide (slide 48), pull it up only if a judge presses that specific attack.

FAQ · anticipated questions · scope & data types

Does it work for anything besides datasets — code, software, models, protocols?

Yes. SHARE scores any research object with public metadata; the 25 signals are format-agnostic (documentation, licensing, linking, identifiers), so software, models, and protocols score the same way.

How do you score controlled-access data (dbGaP) you can't see?

SHARE scores the metadata record, never the underlying data, so gated datasets are scored like open ones — on how well the deposit is documented, licensed, and linked, which stays visible even when the data isn't.

Does it work across languages and non-English metadata?

The signals are structural (is a license present, a method documented, an ORCID linked), not linguistic, so they compute regardless of language. Identifier and controlled-vocabulary signals are language-independent by construction.

The through-line

One rubric, computed from public metadata on any object type, in any language, open or gated. Nothing here needs the data itself — only its metadata record.

sharescore.org · ConductScience Foundation · conductscience.com · FAQ

FAQ · anticipated questions · fairness & equity

Does SHARE penalize small or null-result datasets?

No. It grades documentation quality, not size or outcome. A small, well-documented null-result dataset can score 100; a huge, poorly-documented one scores low. It rewards stewardship, not volume or "positive" results.

Does the S-Index penalize early-career researchers?

The S-Index is max(k) — breadth of quality capped by count — so it grows with career stage like the h-index. We show it alongside a researcher's best and median SHARE, so a few excellent early deposits still stand out.

Some fields legally can't share raw data (clinical, qualitative). Is SHARE fair to them?

It scores the deposit that is made, including metadata-only and controlled-access records, so a field that shares rich metadata about restricted data still scores well. It measures sharing quality given what can be shared.

Does it favor large, well-resourced institutions?

No. SRA (85) and NASA (29) are both large and well-funded, so the spread is curation policy, not budget. Scoring is from public metadata with no fee or opt-in, so a depositor at any institution is scored identically.

sharescore.org · ConductScience Foundation · conductscience.com · FAQ

FAQ · anticipated questions · method & rigor

Is the score reproducible — would two people get the same number?

Yes. Every signal is a deterministic, binary, machine-checked test on public metadata, computed by open code — no human rating step. Two people, or the public API, always get the same score. Inter-rater reliability is 1.0 by construction.

Is this correlation or causation?

We claim leakage-controlled prediction, not causation: predictor at deposit, outcome later, reuse bucket held out. The pre-registered post-incentive re-validation (roadmap Q4 2026) is the causal test — does nudging a repo up actually raise later reuse.

How good a predictor is it, really?

The four-component deposit-time model reaches AUC 0.930 (0.5 = chance, 1.0 = perfect), out-of-sample and cross-validated. The full ROC and per-band rates are in the robustness appendix.

Why 25 signals — why not more or fewer?

25 is what survived a four-stage funnel from 64 candidate fields (in ≥2 of 4 standards → maps to a FAIR sub-principle → machine-checkable everywhere → discriminative). It wasn't designed to be 25; governance can add or retire signals as evidence warrants.

sharescore.org · ConductScience Foundation · conductscience.com · FAQ

FAQ · anticipated questions · operations, sustainability & positioning

Who pays after the grant — and why is a company running it?

Under $10K a year, self-funded by ConductScience. There's a structural reason it isn't grant-run: NIH's Research Tools Policy encourages dissemination but funds no upkeep, and 2 CFR §200 makes universities poor hosts for standing shared infrastructure. So we run a commercial-in-structure, infrastructure-in-mission hybrid that replaces the grant cliff with ongoing funding for maintenance and version control.

How often are scores updated? What about retractions?

Re-harvested as metadata changes; a ratchet keeps an honest score from falling. The one exception: a retraction or a closed exploit re-scores and flags the affected records rather than grandfathering them.

Can you legally score and publish others' metadata?

Yes — SHARE uses public metadata repositories already expose under open terms. The rubric is CC-BY and the code Apache-2.0, so the whole method is forkable and auditable.

How is SHARE different from F-UJI / FAIRshake / RDA maturity indicators?

Those assess FAIRness against a checklist. SHARE adds three things: it's validated against real reuse (5.73×), it separates deposit-time quality from outcomes to resist gaming, and it rolls up to researcher metrics (S-Index, S-Impact). It complements them, not replaces them.

sharescore.org · ConductScience Foundation · conductscience.com · FAQ

Speaker notes · Slide 47, FAQ: operations, sustainability & positioning

Reserve. Cost/sustainability: under $10K/yr, self-funded, recognition not money. Updates/retractions: re-harvested, ratcheted, with retractions and closed exploits re-scored and flagged. Legality: public metadata, CC-BY rubric, Apache-2.0 code. Positioning vs F-UJI/FAIRshake/RDA: those are FAIR checklists; SHARE is reuse-validated, gaming-resistant by design, and researcher-level, and it complements them.

The full "why a company" argument, if a judge presses on sustainability or conflict of interest: although the NIH Research Tools Policy explicitly encourages dissemination, it provides no standing mechanism to fund the ongoing manufacturing, software maintenance, and version governance required to keep methods executable over time. In parallel, federal cost principles under 2 CFR §200 discourage universities from operating standing, cross-campus, or externally-facing production infrastructure, because strict requirements for cost allowability and allocability make it difficult to sustain shared infrastructure that is not cleanly attributable to a single award. Because these constraints leave no viable academic or grant-based pathway for long-term replication infrastructure, we operate an organizational hybrid, commercial in structure but a data infrastructure in mission, sustained through a continuous operating model that replaces the grant cliff with ongoing funding for manufacturing continuity, software upkeep, and version control. This is the same argument made in our published work, with citations to the NIH Research Tools Policy and 2 CFR §200.

FAQ · anticipated questions

What if everyone games it? The Goodhart & LLM-metadata question

Once a score becomes a target, does it stop meaning anything? Our 5.73× was measured on deposits made before any incentive existed — not a gamed number. Against LLM-mass-produced but empty metadata, three layers. Live today in our extensions store: Data Substance Validator runs 8 checks incl. ghost-dataset detection · Originality Lens flags near-duplicate deposits. Being built now: a semantic check, real science vs boilerplate. The real backstop: SHARE validated against actual reuse · metadata that never earns reuse fails over time. A later rule closing a loophole → records re-scored, not grandfathered.

sharescore.org · ConductScience Foundation · conductscience.com · FAQ

Speaker notes · Slide 48, FAQ: Goodhart & LLM-metadata

Reserve slide, pull up only if a judge presses the "cheap LLM-generated metadata defeats binary checks" attack. Be honest and layered, and do not claim a semantic filter that is fully deployed. Our 5.73× was measured on organic, pre-incentive deposits, so it is not a gamed number yet. What ships today and is live in the extensions store is structural: the Data Substance Validator runs 8 checks including ghost-dataset detection, the Originality Lens collapses near-duplicate deposits, and we audit-sample high-scoring deposits. Semantic screening, NLP perplexity where genuine scientific text runs 45 to 120 versus under 10 for boilerplate, is now being built into the deposit-time pipeline. The real backstop that beats fluent-but-hollow LLM prose: SHARE is validated against downstream reuse, so metadata that never actually earns reuse fails the held-out R-bucket over time. And the ratchet is not a gamer's shield: a score never falls for honest depositors, but when a later rubric version closes an exploit, records that exploited it are re-scored and flagged, not grandfathered.

Appendix · why the dataset is the unit of credit

Why now · AI can write infinite prose, but humans generate the real-world data · the paper is becoming a thin wrapper around the dataset underneath.

Our view of science, the paper is becoming a wrapper

Fig 1 · the dataset

DOI 10.5281/zenodo · SHARE 92

LLMs already generate the prose · the paper: a thin, auto-generated wrapper · the durable asset: the dataset underneath · SHAREscore makes it measurable & rewarded.

sharescore.org · ConductScience Foundation · conductscience.com · Appendix

Appendix · roadmap · SHARE 2.0

A proposed next version: move realized reuse out of the score, so all 25 signals are deposit-time and machine-checked — what you control, not citation luck.

The one change: the R bucket

1.0 R = Reuse (outcomes)

views · downloads · citations · derivatives · community — post-deposit, popularity-driven, not controllable

2.0 R = Reusability (deposit-time)

data dictionary · file-level manifest · fixity / checksums · machine-readable citation · structured provenance

maps to FAIR R1 / R1.1 / R1.2 / R1.3 · candidates pending the same 64→25 derivation funnel

The proof still holds

Our headline result — 5.73× per +10, AUC 0.930, 32× dose-response — is the 4-component deposit-time model. It contains no R at all; realized reuse is already held out as the outcome. Moving R out of the score changes the presentation, not the proof.

Re-runs on existing data only: recompute the composite · re-bin the dose-response · validate the genuinely-new signals · re-run S-Impact on the deposit-time core.

Removes the Principle-1 tension and the S-Impact circularity in one move.

sharescore.org · ConductScience Foundation · conductscience.com · Appendix

Speaker notes · Slide 50, Appendix: SHARE 2.0 roadmap

Reserve/roadmap slide, pull up if a judge probes the Reuse bucket or asks what is next. SHARE 2.0 makes one change: the R bucket moves from realized-reuse outcomes to deposit-time Reusability signals, so all 25 scored signals become machine-readable and controllable by the depositor. Realized reuse does not disappear, it moves out of the score to become the validation target and the basis of S-Impact. Crucially this does not cost us the validation: the headline 5.73x and AUC 0.930 come from the four-component deposit-time model, which already excludes R, so the proof is untouched. It also removes the one bucket that sat in tension with our first principle, score what the researcher controls, and it eliminates any circularity between the score and S-Impact. The five new R signals are candidates, pending the same four-filter derivation funnel that produced the original 25.