SoftGovScore: Governance Scoring Under Uncertainty

• SoftGovScore is a framework combining soft, probabilistic, and expert-elicited data to score governance in various settings.
• It encompasses diverse approaches such as score-based mechanisms, expert constraint graphs, digital OSS maturity indices, and sentiment-driven adaptations.
• The framework addresses challenges like data ambiguity and manipulation by integrating screening costs, normalization methods, and expert validations.

SoftGovScore denotes several technically distinct constructs that combine governance-oriented decision or evaluation tasks with information that is soft, probabilistic, manipulable, expert-elicited, or indirectly observed. In the cited literature, the name is used for at least five families of objects: a score-based approval or selection mechanism under soft and semi-hard information; an expert-opinion scoring procedure based on constraint graphs; an index for governmental Open Source Software enablement; a soft-label relevance score for government-domain text; and governance-focused scoring adaptations for OSS projects and firms (Perez-Richet et al., 2024, Mell, 2021, Linåker et al., 6 Oct 2025, Wang et al., 29 Jul 2025).

1. Score-based mechanism design under soft and semi-hard information

In "Score-based mechanisms" (Perez-Richet et al., 2024), SoftGovScore is specified as a score-based approval or selection mechanism for settings in which the decision-maker can observe and contract on submitted scores, but the score itself combines soft information and semi-hard information. The agent’s type is $t=(\theta,s)\in T$, where $\theta\in\Theta$ is the soft component and $s\in S$ is the natural score. A submitted score is $a\in A\subseteq S$, and falsification is governed by a cost function $c:A\times T\to\mathbb{R}_+$ satisfying $c(s,t)=0$ for all $t$. The agent’s payoff is

$u(x,a,t)=v(x,t)-c(a,t),$

while the principal’s payoff may depend on decisions, submitted scores, and the falsification strategy $\rho:T\to\Delta(A)$ through

$u_P(x,a,t,\rho)=v_P(x,t,a)-c_P(\rho,t).$

The canonical mechanism class is a direct recommendation mechanism

$\theta\in\Theta$0

where $\theta\in\Theta$1 is the set of score-based decision rules. After a type report, the mechanism draws both a score submission request and a decision rule; the agent then chooses what score to submit. Truth-telling and obedience constraints are imposed, and ex-post participation implies obedience when the agent can refuse and obtain a null outcome. The framework also yields a decomposition into a score-based decision rule $\theta\in\Theta$2 and a score request or recommendation rule $\theta\in\Theta$3.

A central result is the score-based principle. Under separability,

$\theta\in\Theta$4

and

$\theta\in\Theta$5

and with deterministic score recommendations, there exists a score-based mechanism $\theta\in\Theta$6 and a falsification strategy $\theta\in\Theta$7 that preserve the agent’s payoff and weakly increase the principal’s payoff. Under these conditions the agent need not make a type report and instead self-selects a submitted score in a fixed score-based rule. In the binary approval specialization, this is implemented as an approval probability $\theta\in\Theta$8, possibly accompanied by an evidence request policy $\theta\in\Theta$9.

The optimal-approval analysis specializes further to $s\in S$0 with objective

$s\in S$1

subject to truth-telling, obedience, and $s\in S$2. For linear falsification costs $s\in S$3, the pointwise solution is $s\in S$4 for $s\in S$5 and $s\in S$6 for $s\in S$7, with

$s\in S$8

and

$s\in S$9

For quadratic costs $a\in A\subseteq S$0, the optimal recommendation is

$a\in A\subseteq S$1

which is increasing under the Monotone Hazard Rate property, and the approval schedule is smooth until saturation at $a\in A\subseteq S$2.

This formulation treats manipulation, screening, and approval as jointly designed objects. The practical interpretation supplied in the paper is explicit: the input score may aggregate soft attributes such as stated preferences and self-reported needs with semi-hard evidence such as residence documentation or medical tests, and the policy levers are the approval schedule $a\in A\subseteq S$3 and the verification intensity $a\in A\subseteq S$4 (Perez-Richet et al., 2024).

2. Expert-elicited constraint-graph scoring

A different SoftGovScore construction appears in "The Generation of Security Scoring Systems Leveraging Human Expert Opinion" (Mell, 2021). There, SoftGovScore is a governance-oriented scoring methodology that uses pairwise expert comparisons to generate a prioritization or scoring system over a set of elements

$a\in A\subseteq S$5

Experts compare pairs of elements on an ordered categorical scale: Much less significant $a\in A\subseteq S$6, Less $a\in A\subseteq S$7, Equal $a\in A\subseteq S$8, Greater $a\in A\subseteq S$9, and Much greater $c:A\times T\to\mathbb{R}_+$0. These judgments are mapped to edge degrees $c:A\times T\to\mathbb{R}_+$1, with $c:A\times T\to\mathbb{R}_+$2 denoting equality and $c:A\times T\to\mathbb{R}_+$3 denoting increasing minimum-separation constraints.

The elicitation engine performs binary insertion sort using the expert as the comparison oracle. Parent nodes correspond to equivalence classes $c:A\times T\to\mathbb{R}_+$4, and each new element is inserted by binary search over parent nodes. Equality places the element inside an existing class; otherwise a new parent node is inserted between adjacent classes. The resulting representation is a directed acyclic constraint graph

$c:A\times T\to\mathbb{R}_+$5

with binary adjacency matrix $c:A\times T\to\mathbb{R}_+$6 and weighted adjacency matrix $c:A\times T\to\mathbb{R}_+$7. By construction, the graph induces a total ordering over equivalence classes and requires $c:A\times T\to\mathbb{R}_+$8 expert comparisons per expert, or $c:A\times T\to\mathbb{R}_+$9 over $c(s,t)=0$0 experts.

When multiple expert graphs are available, the method unifies them through pairwise vote aggregation. For each unordered pair $c(s,t)=0$1, counts $c(s,t)=0$2, $c(s,t)=0$3, and $c(s,t)=0$4 are computed, then adjusted by the paper’s conflict-to-equality rule:

$c(s,t)=0$5

where $c(s,t)=0$6. Pairs are processed in descending order of confidence $c(s,t)=0$7, and constraints are added only if they preserve acyclicity. Degree aggregation can then use a weighted mean, weighted median, or mode, with optional expert weights $c(s,t)=0$8.

Numeric scores are generated by imposing constraints

$c(s,t)=0$9

for every edge $t$0 of degree $t$1, with equality edges enforcing $t$2, and bound constraints $t$3. The paper recommends computing per-node feasible lower and upper bounds and then setting

$t$4

Normalization to $t$5 or $t$6 is then straightforward.

The empirical illustration in the paper includes a CVSS experiment on the top 65 CVSS v3 vectors, with 3 experts and 3 sessions each, mean 1.4 hours per session, mean same-expert inconsistency of 10.58%, mean different-expert inconsistency of 14.96%, and a unified DAG with 65 nodes and 68 edges after processing 9 graphs (Mell, 2021). The method is explicitly positioned as domain-agnostic and is specialized in the supplied details to governance elements such as policy maturity, accountability mechanisms, transparency controls, stakeholder engagement, incident reporting and escalation, and auditability and oversight.

3. Digital-government OSS maturity index

In "Advancing Digital Government: Integrating Open Source Software Enablement Indicators in Maturity Indexes" (Linåker et al., 6 Oct 2025), Linåker and Muto describe a quantitative SoftGovScore for governmental Open Source Software enablement. Here the score is an index over three dimensions: Policy Design, Implementation & Adoption, and Support & Governance. It is grounded in cross-case analysis of 16 digitally mature countries and a synthesized set of 14 indicator areas covering policy incentives and design and implementation and support.

The index conceptualizes governmental maturity in OSS through inbound and outbound policy frameworks, implementation artifacts, and organizational support capacity. Policy Design measures, among other things, inbound policies for acquiring OSS, outbound policies for sharing and publicizing OSS, and externally-oriented OSS policies. Implementation & Adoption covers actual uptake and practice, including guidelines, repository and catalog practices, reuse, contribution behavior, interoperability standards, procurement references, and observable adoption across public-sector organizations. Support & Governance evaluates OSPOs at multiple levels, neutral steward bodies for joint OSS projects, subnational OSS leadership, support programs, knowledge-sharing networks, ecosystem participation, training, catalogs, and national social coding platforms.

The paper formalizes the index through normalized indicator scores $t$7 obtained from raw scores $t$8, with sub-indexes

$t$9

and composite score

$u(x,a,t)=v(x,t)-c(a,t),$0

The details give both an implementation-emphasizing choice $u(x,a,t)=v(x,t)-c(a,t),$1 and an equal-weight alternative. An optional security and sustainability adjustment is

$u(x,a,t)=v(x,t)-c(a,t),$2

The 14 indicator areas are named explicitly: Inbound Policies for Acquiring OSS; Outbound Policies for Sharing and Publicizing OSS; Externally-Oriented OSS Policies; Existence of public sector OSPOs; Existence of neutral steward bodies for joint OSS projects; Subnational entities leading and scaling OSS development; Support and capacity-building for PSO OSS policies; Public sector networks for OSS knowledge sharing; Engagement in national and international OSS ecosystems; Inbound policy guidelines for OSS adoption; Outbound policy guidelines for public software release as OSS; Skills Development and training; Catalog of public sector OSS; and National social coding and version management platform for government OSS. The paper proposes annual measurement and specifies measurement signals for each area, such as maturity scales, prescription scales, coverage breadth, reuse rates, catalog usage, onboarding time, and mandate clarity (Linåker et al., 6 Oct 2025).

The country evidence is used illustratively rather than as a single ranking table. France is presented as combining strong outbound transparency law, DINUM’s Free Software Unit, code.gouv.fr, BlueHats, and contribution expectations for sustainability; the Netherlands as combining "Open, unless," institutional OSPO development, and Developer Overheid; and Spain as combining a National Interoperability Framework, reuse mandates, a Technology Transfer Center catalog, and regional OSPO activity (Linåker et al., 6 Oct 2025).

4. Soft labels for government-domain relevance

In "GovRelBench: A Benchmark for Government Domain Relevance" (Wang et al., 29 Jul 2025), SoftGovScore refers to a soft-label construction for training GovRelBERT, a ModernBERT-based evaluator of government-domain relevance. The motivation is that a binary government versus non-government classifier is inadequate because the boundary is fuzzy. The proposed solution converts hard labels into continuous scores on $u(x,a,t)=v(x,t)-c(a,t),$3.

The construction proceeds in two stages. First, a category label $u(x,a,t)=v(x,t)-c(a,t),$4 is mapped to a hard relevance score

$u(x,a,t)=v(x,t)-c(a,t),$5

Second, an adjusted score $u(x,a,t)=v(x,t)-c(a,t),$6 is used as the mean of a Beta distribution with fixed concentration $u(x,a,t)=v(x,t)-c(a,t),$7:

$u(x,a,t)=v(x,t)-c(a,t),$8

with

$u(x,a,t)=v(x,t)-c(a,t),$9

The resulting variance profile,

$\rho:T\to\Delta(A)$0

is largest in the middle of the interval and smallest near 0 and 1. The model is then trained by mean squared error,

$\rho:T\to\Delta(A)$1

GovRelBERT_A uses a single regression head on a ModernBERT encoder with a final hidden-state representation of dimension 768; GovRelBERT_B adds an 18-way classification head and uses a combined cross-entropy plus MSE objective. The training data comprise 78,200 samples split 8.5:1:0.5 into train, validation, and test, with random seed 3407 and early stopping on validation MSE. The 18 categories include, for example, Government Affairs, Law_A, News, Politics, Government Work Reports, Education, Entertainment, and Others, with mapped hard scores such as Government Affairs $\rho:T\to\Delta(A)$2, Law_A $\rho:T\to\Delta(A)$3, News $\rho:T\to\Delta(A)$4, Politics $\rho:T\to\Delta(A)$5, Government Work Reports $\rho:T\to\Delta(A)$6, Education $\rho:T\to\Delta(A)$7, Entertainment $\rho:T\to\Delta(A)$8, and Others $\rho:T\to\Delta(A)$9 (Wang et al., 29 Jul 2025).

The reported evaluation compares soft-score regression against traditional baselines, an encoder classifier baseline, and prompted decoder-only LLMs. GovRelBERT_A attains 71.47% accuracy and 0.7394 F1 at tolerance $u_P(x,a,t,\rho)=v_P(x,t,a)-c_P(\rho,t).$0, and 93.58% accuracy and 0.9380 F1 at $u_P(x,a,t,\rho)=v_P(x,t,a)-c_P(\rho,t).$1; GovRelBERT_B attains 66.93% and 0.6929 at $u_P(x,a,t,\rho)=v_P(x,t,a)-c_P(\rho,t).$2, and 92.19% and 0.9250 at $u_P(x,a,t,\rho)=v_P(x,t,a)-c_P(\rho,t).$3 (Wang et al., 29 Jul 2025). In this lineage, SoftGovScore is not an index over institutions or a mechanism under strategic falsification, but a probabilistic target used to train a text regressor for government-domain relevance.

5. Dataset-oriented and sentiment-oriented governance adaptations

The name is also used for adaptations built on datasets that were not originally introduced under that label. In "GitHub OSS Governance File Dataset" (Yan et al., 2023), the supplied methodology defines a SoftGovScore for OSS projects using 710 GitHub-hosted OSS projects with root-level governance files matching governance-like patterns around GOVERNANCE.MD. The dataset contains all commits, all issues and comments, full governance-file revision histories, and the latest governance-file content through the last scripted update in June 2022. The proposed governance score combines dimensions such as governance content richness, governance update cadence and recency, adherence or internal consistency, community engagement and responsiveness, and participation distribution and role activity.

That construction uses min–max normalization

$u_P(x,a,t,\rho)=v_P(x,t,a)-c_P(\rho,t).$4

with an additive composite

$u_P(x,a,t,\rho)=v_P(x,t,a)-c_P(\rho,t).$5

or a geometric composite

$u_P(x,a,t,\rho)=v_P(x,t,a)-c_P(\rho,t).$6

The raw ingredients include extracted counts of roles, responsibilities, decision processes, voting references, conflict-resolution references, release-process references, code-of-conduct references, contributing references, text length, section count, governance-file update intervals, issue response times, contributor Gini, and contributor turnover. The paper itself is a dataset paper rather than a SoftGovScore paper; the governance score is an implementation-oriented use of the dataset supplied in the details (Yan et al., 2023).

A second adaptation appears in "Creating a Systematic ESG (Environmental Social Governance) Scoring System Using Social Network Analysis and Machine Learning for More Sustainable Company Practices" (Patel et al., 2023). There, a governance-only SoftGovScore is built from the paper’s ESG sentiment pipeline. The underlying data collection covers approximately 937,400 data points across roughly 470 S&P 500 companies, using Wikipedia, Google News, Twitter through Snscrape, and LinkedIn through Selenium and BeautifulSoup. Governance keywords include governance, compensation, corruption, ethical, fraud, justice, and transparency. Governance sentiment is aggregated as

$u_P(x,a,t,\rho)=v_P(x,t,a)-c_P(\rho,t).$7

with effectively uniform $u_P(x,a,t,\rho)=v_P(x,t,a)-c_P(\rho,t).$8 in the paper’s implementation. The full ESG predictor evaluates Random Forest Regression, Support Vector Regression, K-Nearest Neighbors Regression, and XGBoost Regression against S&P Global ESG ratings on a 0–100 scale; Random Forest performs best with MAE $u_P(x,a,t,\rho)=v_P(x,t,a)-c_P(\rho,t).$9, correlation $\theta\in\Theta$00, and $\theta\in\Theta$01 on the holdout set (Patel et al., 2023). The supplied governance adaptation defines

$\theta\in\Theta$02

where $\theta\in\Theta$03 is the governance-specific feature vector.

These two uses differ from the mechanism-design, public-sector maturity, and government-text settings. One is repository- and community-centric, grounded in governance files and activity traces; the other is firm-centric, grounded in governance-related public sentiment and supervised learning.

6. Shared structure, divergences, and recurring limitations

Across these uses, the common thread is not a single canonical algorithm but a recurring attempt to operationalize governance with information that is soft in some sense. In the mechanism-design framework, softness refers to freely manipulable or semi-hard evidence and the need to induce costly screening (Perez-Richet et al., 2024). In the expert-opinion framework, it refers to latent significance that lacks direct ground truth and is elicited by pairwise human judgments (Mell, 2021). In the OSS maturity index, it refers to policy and organizational capacities that must be encoded by rubrics and annual indicators (Linåker et al., 6 Oct 2025). In GovRelBench, it refers to soft labels that better represent fuzzy government-domain relevance than discrete classes (Wang et al., 29 Jul 2025). In the OSS-project and ESG adaptations, it refers to textual governance artifacts, public interaction traces, and social sentiment (Yan et al., 2023, Patel et al., 2023).

The divergences are correspondingly substantial. The mathematical objects range from direct recommendation mechanisms $\theta\in\Theta$04, falsification strategies, and approval schedules $\theta\in\Theta$05, to DAGs with equivalence classes and separation constraints, to composite maturity sub-indexes, to Beta-diffused regression targets, to Random Forest outputs on an S&P-aligned 0–100 scale. The term therefore names a family resemblance rather than a single standard.

Several limitations recur. The expert-elicitation method assumes experts can reliably compare pairs and that transitivity holds across the induced order; the paper also notes that unified graphs may lose total order and require heuristic tie-handling (Mell, 2021). The OSS maturity index notes heterogeneous government levels and legal contexts, outcome-versus-intention gaps, and resource constraints in interviews and document collection (Linåker et al., 6 Oct 2025). The GovRelBench formulation depends on category-to-score mappings and Chinese-heavy source corpora, and it does not claim that relevance is equivalent to correctness or fairness (Wang et al., 29 Jul 2025). The OSS-project governance adaptation is constrained by root-level governance-file coverage and the absence of pull requests and review events in the dataset (Yan et al., 2023). The ESG-derived governance score inherits platform bias, sarcasm, misinformation, and coverage bias, and the paper’s reported predictive performance վերաբers to the full ESG target rather than a separately reported governance-only ground truth (Patel et al., 2023).

Taken together, these works suggest that SoftGovScore is best understood as a recurring research label for governance scoring or governance-aware selection under uncertainty, manipulability, partial observability, or fuzziness. What unifies the label is the attempt to replace purely hard, binary, or self-reported governance signals with formal structures that preserve ambiguity, screening costs, expert uncertainty, or indirect evidence rather than suppressing them.