Data access request under Article 40 of the DSA¶
What this document is, and is not
It is a template ready to file, not a filed request. Article 40(8)(a) DSA reserves vetted researcher status to persons affiliated to a research organisation within the meaning of Article 2(1) of Directive (EU) 2019/790. This repository, kept by one person, is not eligible as it stands: the request below is written to be filed by an organisation that meets the conditions, and the parts to complete are in brackets.
What makes this request unusual
It asks for no logs and no personal data. It asks for four aggregate tables, which notebook 18 verifies are sufficient to recompute identically the four measurements that decide everything — and which weigh, on a real log, 95 times fewer rows than the log itself.
The reasoning that leads to this request¶
Three public datasets have been measured, and none permits the announced evaluation:
- MIND carries editorial categories but not the rank — its recorded order is indistinguishable from a shuffle (\(z = +0.12\));
- Baidu-ULTR carries the rank (\(z = -206\)) but no usable label;
- the Open Bandit Dataset carries rank, true propensity and a random bucket, but its item attributes are anonymised.
Article 40 remains. And a request is admissible only if it is necessary and proportionate — Article 40(8)(e) DSA, Article 8(d) of Delegated Regulation (EU) 2025/2050, in force since 29 October 2025. "Give us your logs" is neither: it demands personal data the analysis has no need of, and it hands the platform the easiest ground for refusal — service security and trade secrets, Article 40(5).
Hence the form adopted: a specification, not an access.
I. Identification and legal basis¶
| Applicant | [RESEARCH ORGANISATION], within the meaning of Art. 2(1) of Directive (EU) 2019/790 |
| Principal researcher | [NAME], [POSITION] |
| Data provider | [PLATFORM], designated a VLOP on [DATE] |
| Digital Services Coordinator of establishment | [AUTHORITY] — for most VLOPs, the Irish authority |
| Route | Article 40(4) of Regulation (EU) 2022/2065, procedure of Delegated Regulation (EU) 2025/2050 |
| Filing | DSA data access portal (Arts. 3 and 5 of the Delegated Regulation) |
On the route chosen. Article 40(12) — access to publicly accessible data, open to researchers affiliated to not-for-profit bodies — does not fit: the data requested here are not publicly accessible, since it is precisely the served order and the exposure that public datasets lack. Route 40(4) is therefore the only one.
II. Subject of the research (Art. 8(f) of the Delegated Regulation)¶
The subject is the detection and measurement of a systemic risk under Article 34(1)(c) DSA — negative effects on civic discourse — in a measurable form:
A platform bound by a diversity floor bearing on the composition of its feeds can comply by burying divergent items at ranks nobody reads. The composition stays compliant and the actual exposure does not.
This repository established the point by exhaustive enumeration: a platform certified at 0.70 by a rank-blind measure exposes only 0.36 of real diversity, and closing the loophole doubles the engagement cost (adversarial rank). It also established that the correction requires knowing exposure, hence the served rank, and that ignoring it makes the evaluation vacuous by construction (MIND).
The rank requested must be verifiable, not declared
At least one public dataset supplies an order column that is not the served rank, without lying and with nothing signalling it: on EB-NeRD the recorded order fails the exchangeability test (\(z = +1.05\)) even though the log would have detected a severity a hundred and thirty times smaller. The request therefore bears on a rank whose informativeness can be tested, and that test is the first thing to run on the delivered data. → the index measured
The research requested consists of measuring, on the feeds actually served by [PLATFORM]:
- the gap between composed diversity and exposed diversity, by period;
- the position-bias severity proper to this surface, rather than transported from another — the study measured 1.10 on a search results page and 0.04 to 0.11 on a three-thumbnail banner, an order of magnitude apart;
- the engagement cost of a re-ranking that respects a rank-aware floor, estimated by importance weighting rather than by replay, whose measured median error reaches 201 %.
III. Data requested (Art. 8(c) of the Delegated Regulation)¶
Four aggregate tables, in CSV or Parquet, for the period [PERIOD] and the market [MEMBER STATE].
Table 1 — feed profiles
| Column | Type | Description |
|---|---|---|
profile |
integer | identifier of the rank profile (see table 1 bis) |
feed_clicks |
integer | number of items clicked in the feed |
feeds |
integer | number of matching feeds |
Table 1 bis — profile definitions: profile, rank — the list of ranks each profile
occupies. Indexing by feed length alone would be shorter and wrong as soon as a surface
skips ranks, which a results page does.
Table 2 — clicks by rank
| Column | Type | Description |
|---|---|---|
profile |
integer | profile of the feed |
feed_clicks |
integer | number of items clicked in the feed |
rank |
integer | served rank |
clicks |
integer | clicks observed at that rank |
The second key is not an ornament: without it, feeds where everything was clicked — which constrain nothing — cannot be excluded from the computation.
Table 3 — (item, rank) cells
| Column | Type | Description |
|---|---|---|
item |
pseudonymous identifier, stable over the period | item served |
rank |
integer | served rank |
impressions |
integer | times the item was served at that rank |
clicks |
integer | clicks observed |
propensity |
real, optional | probability of service, if the platform knows it |
displays |
integer | impressions actually displayed to the reader |
format |
label | the item's display format, a category declared by the platform |
The displays column removes three problems at once
It is the only addition to this request since it was drafted, and it comes from a measurement: on Baidu-ULTR, the same quantity obtained by display rather than by click gives \(\hat\eta = 0.88 \pm 0.05\) over 143 documents, against \(1.09 \pm 0.09\) over 55. → Measured exposure
Without it, exposure must be estimated through clicks, which imposes three assumptions none of which is verifiable: that examination follows a power law, that it does not depend on what was clicked above, and that an item's attractiveness does not vary with its rank. The first is false on measured data, the second undecidable without it, and the third inflates the estimate by 23 %.
It is moreover less sensitive than clicks: knowing an item was displayed says less about a reader than knowing they chose it.
The format column serves to check, not to compute
It was first requested for a reason that did not hold: on Baidu-ULTR a rich format above appeared to remove 8.4 points of examination from what follows. With the item held fixed, only 6 % remains, and the bearing on the index falls to \(0.0025\) — fourteen times less than that of the \(1/R\) convention. → The page effect
The column is still requested, for two more modest reasons: it allows this check to be reproduced on another platform, where the effect might be stronger; and it is the least sensitive of the six — an item's format says nothing about its reader. It does not condition the computation of the index, and a platform failing to supply it would not prevent it.
Table 4 — exposure by viewpoint
| Column | Type | Description |
|---|---|---|
rank |
integer | served rank |
viewpoint |
label | category of the reference catalogue declared by the regulator |
impressions |
integer | number of displays |
clicks |
integer | clicks observed |
The viewpoint catalogue is a decision of the regulator, not of the platform nor of the researcher: it is the political discretisation that the critical audit identifies as the chief reservation on using the index. This request concerns its application, not its choice.
IV. Necessity and proportionality (Art. 8(d) of the Delegated Regulation)¶
What these tables allow — and nothing more.
| Measurement | Tables needed |
|---|---|
| exchangeability test — is the log even correctable? | 1, 1 bis, 2 |
| exposure severity \(\eta\), measured | 3, displays column |
| position-bias severity \(\eta\), estimated failing that | 3 |
| form of examination (cascade or not) | 3, displays column |
| effect of format on exposure, item held fixed | 3, format and item columns |
| counterfactual estimation and effective sample size | 3, with propensities |
| composed and exposed diversity, burial gap | 4 |
What is verified, not asserted. Notebook 18
recomputes these measurements twice — directly on a complete log, then on the tables alone — and
publishes the gap. It is \(3 \times 10^{-12}\) for the exchangeability test and exactly zero
for severity. The code that consumes the tables is
ide.aggregates,
and eleven tests lock it.
What it weighs. On Baidu-ULTR, 524,164 documents served yield 5,543 rows requested, a ratio of 95. On MIND, 5,843,444 items served yield 57,906 rows, ratio 101.
What is not requested: no reader identifier, no browsing sequence, no content, no text, no ranking parameter, no model. A feed appears in these tables only by its shape and by the number of clicks it received.
Consequence for Article 40(5). A platform may request amendment of a request if disclosure would create security vulnerabilities or reveal trade secrets. Counts by rank and by declared viewpoint reveal neither the ranking, nor its parameters, nor the items: they say what was exposed, not how the platform decided to expose it.
V. Risks, confidentiality and data protection (Art. 8(e))¶
Personal data. The tables requested contain none: the smallest unit is the cell, and no row designates a reader. The processing therefore does not involve a communication of personal data, which removes the request's principal risk and the platform's principal ground for objection.
Confidentiality threshold. The request proposes a suppression threshold of [THRESHOLD] impressions per cell, and asks that it be published with the data. This threshold is not neutral, and its effect is measured rather than assumed: on Baidu-ULTR, moving from 5 to 20 impressions shifts the estimated severity from 1.10 to 1.40, i.e. +27 %, because rare cells are those of deep ranks, from which the estimate draws most of its information.
Access modalities. A secure processing environment within the meaning of Article 9(5) of the Delegated Regulation is accepted without reservation. Given the aggregate nature of the data, a simple transmission nonetheless appears proportionate.
Security. [TECHNICAL AND ORGANISATIONAL MEASURES OF THE ORGANISATION].
Funding (Art. 8(b)): [FUNDING SOURCES]. Independence (Art. 40(8)(b)): [DECLARATION].
VI. Publication of results (Art. 40(8) and Art. 8(a))¶
The publication commitment is already met in advance: all code, derived data, figures and results — including negative results — are published in a public repository, under MIT for code and CC BY 4.0 for content. The six negative results already published by this repository are the most useful demonstration of it.
The data received will not themselves be republished; the derived aggregates will be, along with the code producing them, in line with the practice already applied to the three datasets measured.
VII. Timetable and remedies¶
| Step | Deadline | Source |
|---|---|---|
| Coordinator's decision: reasoned request or rejection | 80 working days | Art. 7(1) of the Delegated Regulation |
| Amendment request by the platform | 15 days | Art. 40(5) DSA |
| Coordinator's decision on the amendment | 15 days | Art. 40(6) DSA |
| Notification that access has been given | 3 working days | Art. 15(1) of the Delegated Regulation |
| Request for mediation | 5 working days | Art. 13(1) |
| Opening, then duration of mediation | 20, then at most 40 working days | Arts. 13(4), 13(9) |
Article 15(3) of the Delegated Regulation further forbids the provider from imposing limitations on the use of standard analytical tools, unless the reasoned request mentions them explicitly.
VIII. What we will do with a refusal¶
The refusal will be published, quoted as received, and analysed like any other result: a request of this form cannot be set aside for disproportion without the ground bearing on something other than its size. That is the point of the specification — to shift the burden of proof, and to make refusal arguable.
Reservations attaching to this request¶
These tables do not permit everything. They say nothing of a reader's trajectory, hence nothing of individual filter bubbles, nor of the temporal dynamics of one user's feed. That is deliberate — those are the analyses that would require personal data — but it must be said: this request is narrow, and part of the repository's research programme falls outside it.
An aggregate is an assumption. Tables 1 and 2 are keyed by rank profile and by feed click count precisely because two simpler assumptions turned out false in use, on real data, without producing a visible error. A platform delivering these tables under a coarser key would deliver wrong figures of the right order of magnitude — the worst case.
And eligibility is missing. This document is a template, not a procedure under way. Its value is in being ready: a research organisation wishing to ask this question no longer has to design the specification, only to file it.
Implementation: ide.aggregates · Notebook:
18 — What to ask for ·
MIND's real exploration · logs that record the rank ·
memorandum