antonsaccount/open-selection-graph

Open Selection Graph (OSG)

Open Selection Graph is a dataset for studying how candidates move through observable selection processes in science. It brings together public evidence from peer review, research funding and patent examination. The common unit is an institutional cycle: venues, tracks, calls, panels, journal periods, or examination processes governed by a defined set of rules.

Public records begin at different points. Conferences may expose submissions after desk screening, funders may publish the proposals considered by a panel, and patent sources may begin with publication or a recorded office action. OSG records starting points for populations to make the available denominator, hidden stages, and admissible comparisons visible to researchers from the outset.

Version 2.0.0 is a frozen release with a source cutoff of 23 August 2026.

Paper and interactive discovery

The data descriptor, Open Selection Graph (OSG): a dataset for mapping observable populations and selection processes in science, documents the dataset design, construction, validation and recommended uses. The archival release and citation record are available at doi:10.5281/zenodo.22081233.

The Dataset Viewer exposes data/gate_cycle_descriptive_atlas.parquet, a 350-row CC0 discovery table describing reconstructed institutional cycles, population coverage, stage counts and analysis-eligibility flags. The complete licence-separated release remains available through the archives listed below.

What the dataset contains

OSG represents candidates, document versions, institutional policies, evaluations, decisions, lineage, and later outcomes in one relational graph. Native terminology remains attached to each record. An acceptance decision, a fundable-band outcome, and a notice of allowance therefore retain their domain-specific meanings.

Resource	Contents	Scale in version 2.0.0
Observability census	Public population boundaries, earliest visible stages, source coverage, and policy surfaces	4,872 cycle records
Process atlas	Reconstructed publication-selection cycles with coverage and rate eligibility	350 cycles
Audited OpenReview cohorts	Public candidate populations, decisions, withdrawals, and official-review counts	55,369 candidates and 31,992 reviews
Research funding panels	UKRI application-panel events and anonymised SNSF proposal votes	21,883 UKRI events and 4,811 SNSF vote cells
Patent examination	PANORAMA office-action trajectories, claim-state transitions, prior-art links, and capacity cells	8,143 application trajectories
Released HUPD comparison population	Privacy-minimised metadata for published English-language US utility applications filed from 2004 through 2018	4,518,254 applications

Additional tables describe manuscript versions, source-declared lineage, fixed-window publication outcomes, institutional rules, evaluation criteria, and semantic or recombinatorial measurements. SPECTER2 and Qwen3 provide separate semantic rulers over 8,501 shared time-valid document versions. Their disagreement is released as a measurement result.

Understanding population coverage

The observability grade states which population a source can support.

Grade	Public evidence	Appropriate interpretation
A	The observable population begins at entry to the modelled process	Entry-selection and stage-transition analysis
B	The observable population begins after a named institutional screen	Analysis conditional on reaching the public stage
C	Review histories are selected, published, or opt-in	Description of visible evaluation and revision histories
D	Awards, publications, or other outcomes form the public registry	Portfolio and outcome description
U	The public population boundary remains unresolved	Source auditing, disclosure research, and missingness analysis

Grades belong to a source, cycle, and object type. They can change over time. Independent count evidence and stage-flow checks determine whether a cycle is eligible for denominator-based analysis. Version 2.0.0 contains 210 Grade A or B process cycles, including 200 with an observed candidate population. Sixty-one cycles support observed selection rates and 197 support review-incidence measures.

Research uses

OSG supports questions such as:

• Which parts of a selection process become public, and when?
• How do review intensity and selection shares vary across observable institutional architectures?
• How do policy changes coincide with shifts in evaluation, workload, or candidate flow?
• How do manuscripts change across versions, venues, and later publication states?
• How stable are novelty results across scientific-document encoders and recombination measures?
• How do panel-stage funding choices and patent-examination trajectories compare with publication review at the level of process architecture?
• Which findings remain stable when hidden screens, linkage uncertainty, and source coverage are varied?

The dataset is suited to metascience, research evaluation, institutional comparison, measurement research, source auditing, and reproducible method development.

Files and access

The release is organised into six licence-separated components.

Component	Main contents	Licence
metadata	Coverage atlas, process flows, policies, rubrics, schemas, documentation, and benchmark definitions	CC0 1.0
derived_features	Evaluations, novelty measures, lineage, outcome windows, construct benchmarks, and transport diagnostics	CC BY 4.0
funding_open_panels	UKRI application-panel events and anonymised SNSF proposals, votes, and outcomes	CC BY 4.0 and OGL 3.0; source notices apply
funding_noncommercial	Funding opportunity, panel, portfolio, recurrence, and grant-output tables	CC BY-NC-SA 4.0
patent_noncommercial	PANORAMA process tables and the full privacy-minimised HUPD boundary census for 2004 through 2018	CC BY-NC-SA 4.0
validation	Machine-readable coverage, leakage, linkage, privacy, licensing, and recovery reports	CC0 1.0

Each component includes a LICENCE.json file. The release root contains VERSION.json, PACKAGE.json, MANIFEST.json, VALIDATION.json, TABLE_COUNTS.csv, an analytical DuckDB database, and machine-readable schemas. The HUPD census remains a standalone Parquet table because of its size. Its row count and hash are verified with the rest of the package.

The frozen version 2.0.0 release is archived at doi:10.5281/zenodo.22081233. Code and metadata are available from GitHub, and the licence-separated archives are mirrored on Hugging Face.

Quick start

Query the analytical database with DuckDB:

import duckdb

con = duckdb.connect("observatory.duckdb", read_only=True)

cycles = con.sql("""
    SELECT platform_label,
           effective_observability_grade,
           count(*) AS cycles
    FROM gate_cycle_descriptive_atlas
    GROUP BY 1, 2
    ORDER BY 1, 2
""").df()

Query the released HUPD comparison population directly:

hupd_count = con.sql("""
    SELECT count(*) AS applications
    FROM read_parquet(
      'components/patent_noncommercial/hupd_application_population.parquet'
    )
""").fetchone()[0]

For selection-rate research, begin with gate_cycle_descriptive_atlas. Filter on descriptive_rate_allowed, confirm a positive observable_count, and report earliest_public_stage with the result. Evaluation research should join evaluation_objects to the exact candidate version and retain the native rubric and scale.

Data sources

The dataset was assembled from public APIs, OAI-PMH services, bulk files, structured publisher records, public workbooks, and openly available research corpora. Every source has a versioned card describing its role, public stage, access method, identifiers, licensing, expected-count evidence, and realised release state.

The catalogue uses three reader-facing states:

• Released means that the source contributes records or computational dependencies to version 2.0.0.
• Pointer means that OSG supplies provenance, retrieval instructions, or validation evidence while the upstream object remains with its provider.
• Catalogued means that the source was assessed and retained in the source census. Its current coverage, access, or licensing state prevents row-level inclusion.

Complete source catalogue: all 53 registered source cards

Publication platforms and process records

Source	Role	State
OpenReview public configuration	Venue groups, invitations, permissions, rubrics, and policy surfaces	Released
OpenReview audited API cohorts	Public submission-state populations, decisions, withdrawals, and reviews	Released
OpenReview legacy venue connector	Earlier venue enumeration pathway retained for provenance	Catalogued
Copernicus public discussion	Discussion papers, referee comments, replies, and versions	Released
Crossref/Copernicus denominator layer	Independent discussion-population reconciliation	Released
Copernicus outcomes	Later article and discussion relations	Released
eLife Reviewed Preprint census	Editorial-model source census	Catalogued
eLife process records	Reviewed Preprints, reviews, responses, assessments, and policy cohorts	Released
F1000Research records	Published articles, versions, reports, and responses	Released
F1000 platform-family process records	Shared provider-native process representation	Released
Wellcome Open Research	Public post-publication review process	Released
Gates Open Research	Public post-publication review process	Released
NIHR Open Research	Public post-publication review process	Released
HRB Open Research	Platform-family source census	Catalogued
SciPost source census	Public submissions and refereeing pages	Catalogued
SciPost process records	Editor-assigned public submissions, reports, replies, and decisions	Released
PeerJ review histories	Published review histories	Pointer
PLOS review histories	Author opt-in published histories	Pointer
EMBO Press transparent review	Published process files	Pointer
Royal Society review histories	Published process files	Pointer
BMC open review	Open-review journal histories	Pointer
Qeios	Public preprints and evaluation surface	Catalogued
PREreview	Community reviews of public preprints	Catalogued

Derived peer-review corpora

Source	Role	State
ReviewArena	Derived review corpus	Pointer
Re2	Derived review corpus	Pointer
NLPeer	Multi-domain peer-review corpus	Pointer
ResearchArcade	Derived OpenReview research corpus	Pointer
PeerRead	Papers, reviews, and decisions from selected venues	Pointer
MOPRD	Transparent-review histories	Catalogued
Review Arcade / ACL ARR consent programme	Consented peer-review research corpus	Pointer

Scholarly metadata and content

Source	Role	State
Crossref	Bibliographic metadata and work relations	Released
OpenAlex	Scholarly identities, works, citations, and concepts	Released
arXiv	Preprint metadata and version history	Released
Cornell/Kaggle arXiv snapshot	Bulk mirror assessed for recovery	Catalogued
Europe PMC	Biomedical identifiers, metadata, JATS, references, and corrections	Released

Research funding

Source	Role	State
Demystifying Funding / UKRI public sources	Panel and outcome evidence	Pointer
Austrian Science Fund Open API	Public funded-project registry	Released
Swiss National Science Foundation Data Portal	Public funded-project registry	Released
UKRI public outcomes	Council and call outcome workbooks	Catalogued
Public lottery-instrument pages	Funding-policy and allocation-mechanism census	Catalogued
SNSF individual-vote dataset	Anonymised proposal, panel, grade, conflict, and outcome cells	Pointer

Patent examination

Source	Role	State
USPTO Office Action Research Dataset	Public prosecution events	Released
USPTO PatEx	Public application and examination records	Released
USPTO Patent Claims Research Dataset	Published and granted claim records	Released
USPTO Open Data Portal	Application-level retrieval pathway	Catalogued
PANORAMA	Public rejection-to-allowance process trajectories	Pointer
Patent-CR	Rejected-to-granted claim-revision pairs	Pointer
PatRe	Public prosecution cases and revision evidence	Pointer
Harvard USPTO Patent Dataset	Public US utility-application population	Pointer

Models and human benchmarks

Source	Role	State
SPECTER2	Scientific-document embedding ruler	Released
Qwen3-Embedding-0.6B	Independent text-embedding ruler	Released
Peer Review Analyze 1.0	Human-labelled free-form review benchmark	Pointer
PreprintToPaper Gray Zone	Human-adjudicated lineage benchmark	Pointer

Normalised facts retain links to source and provenance records. Content governed by source-specific terms is represented through identifiers, retrieval recipes, and verification hashes.

Construction and measurement

Semantic measurements use revision-pinned SPECTER2 and Qwen3 encoders. Each target year is compared with earlier documents. Citation and textual recombination are calculated as separate measures because they describe different opportunity spaces. Source-declared relations form the analysis-grade lineage layer; probabilistic candidates are supplied for discovery and sensitivity analysis.

Policy, evaluation, and outcome fields preserve provider terminology before normalisation. Point-in-time views align document versions, policy states, and downstream records with the date of the modelled decision.

Validation

Validation operates at the population, record, temporal, and package levels.

• Five OpenReview shards reconcile with provider-reported public candidate counts for all 176 audited cycles.
• Stage-flow identities check observable, reviewed, withdrawn, rejected, selected, and unresolved counts.
• Temporal tests exclude later document versions, future citations, and identities revealed after a decision.
• Native outcomes and rubric values are preserved through normalisation.
• Funding workbooks reconcile panel totals, outcomes, repeated appearances, and duplicate rows.
• Every PANORAMA case has one HUPD reconciliation state. The states account for all 8,143 cases.
• Every pointer row has an absolute HTTPS retrieval target, an object locator, and at least one payload-derived SHA-256 verification hash.
• The release scanner checks schemas and packaged paths for protected identity fields, contacts, raw objects, and restricted content.
• Clean-room tests rebuild representative public fixtures in an isolated environment.
• Package verification checks 72 Parquet tables, the DuckDB tables, the external HUPD table, and every file hash.

Machine-readable reports for these checks are included in the validation component.

Scope and limitations

OSG describes public evidence. Hidden editorial, administrative, and legal stages create selection into observation. The coverage grade identifies the earliest visible stage, while partial-identification and sensitivity tables describe compatible hidden populations.

Institutional populations differ in field, calendar period, architecture, size, selection share, and review intensity. Pairwise transport diagnostics measure overlap in the recorded dimensions. Cycle-level domain, policy, and disclosure modifiers have incomplete coverage and remain required inputs for stronger cross-source transport claims.

Publication outcomes use fixed observation windows. Exact publication dates are unidentified for much of the linked panel, and unmatched candidates retain an unknown state. Free-form review constructs and probabilistic lineage are external benchmarks or discovery aids. Validated source-declared evidence governs canonical links.

The funding tables begin at the observed panel stage. Pre-panel applicants and randomised allocation arms fall outside the available public choice sets. The HUPD comparison population covers published English-language US utility applications filed from 2004 through 2018. PANORAMA supplies a separate 8,143-case process sample. Its reconciliation is complete: 3,416 cases match HUPD exactly, 3,532 fall outside HUPD's filing-year boundary, and 1,195 fall within the boundary but are absent from the frozen HUPD population. Confidential applications and applicant response documents are outside the released evidence.

Privacy, licensing, and responsible use

The public package minimises person-level information. Evaluation identifiers, protected authorship identifiers, direct contact fields, source full text, proposal text, claim text, and raw acquisition objects are excluded. Funding and reviewer-supply measures are aggregated or de-identified according to their source conditions.

OSG is intended for research on institutions, populations, and measurement. The release policy prohibits reviewer deanonymisation, identity inference aimed at anonymous people, individual employment or funding evaluation, personnel ranking, harassment, surveillance, automated targeting, and reconstruction of restricted text.

Licensing is component-specific. Consult the LICENCE.json file beside each component and the source attribution records before reuse. A verification hash establishes provenance; access and redistribution remain governed by the applicable source terms.

Citation and contact

Please cite the dataset and its Data Descriptor:

Waagaard, A. Open Selection Graph (OSG): a dataset for mapping observable populations and selection processes in science. Version 2.0.0 (2026).

Creator: Anton Waagaard
Affiliation: Dimensional Impact Lab
Contact: anton@dimpactlab.org

Corrections and takedown requests are documented in TAKEDOWN_AND_CORRECTIONS.md. Versioned manifests record every amended release.