GLIDR: Multi-Domain Disambiguation

• GLIDR is an ambiguous term representing distinct methodologies across LLM agents, differentiable rule learning, sparse LiDAR processing, and prediction-powered inference.
• Each variant leverages explicit intermediate structures—such as hierarchical subtasks, graph-like rule templates, topologically regularized backbones, or debiased estimators—to improve performance and interpretability.
• Comparative evaluations emphasize that careful domain-specific disambiguation is key, with measurable improvements in reinforcement learning, rule extraction accuracy, point-cloud reconstruction, and statistical estimation.

GLIDR is not a single universally fixed acronym in current technical literature. It appears as a misspelling of GLIDER, a framework for grounding LLMs as efficient long-horizon decision-making agents via offline hierarchical reinforcement learning (Hu et al., 26 May 2025); as the title acronym of a differentiable inductive logic programming method for graph-like rule learning over knowledge graphs (Johnson et al., 8 Aug 2025); and in closely related spellings such as GLiDR for topology-aware sparse LiDAR densification (Kumar et al., 2023) and GLIDE for prediction-powered inference in GenAI and agentic-system evaluation (Martinon et al., 29 May 2026). In gravitational-wave astronomy, “GLIDR-like” is also used descriptively for galaxy-ranking infrastructures such as the HOGWARTs web application (Salmon et al., 2019).

1. Nomenclature and disambiguation

The term is best understood as a nomenclatural collision across several subfields. Capitalization and expansion matter, because the same or adjacent string is used for unrelated methods, objectives, and mathematical formalisms.

Name	Expansion or description	Domain
GLIDER	Grounding LLMs as EffIcient Decision-Making Agents via Offline HiErarchical Reinforcement Learning	LLM agents and offline HRL
GLIDR	Graph-Like Inductive Logic Programming with Differentiable Reasoning	Differentiable ILP and KG completion
GLiDR	Topologically Regularized Graph Generative Network for Sparse LiDAR Point Clouds	LiDAR densification and navigation
GLIDE	Industrializing Prediction-Powered Inference	Statistical evaluation of GenAI and agentic systems
GLIDR-like systems	Galaxy-targeted follow-up infrastructures such as HOGWARTs	Gravitational-wave astronomy

Two recurrent sources of confusion are explicit in the literature. First, GLIDER is “often misspelled ‘GLIDR’” in the LLM-agent context (Hu et al., 26 May 2025). Second, GLIDE is “sometimes misspelled ‘GLIDR’” in the prediction-powered inference literature (Martinon et al., 29 May 2026). By contrast, the differentiable ILP method is genuinely titled GLIDR (Johnson et al., 8 Aug 2025), while the LiDAR system is titled GLiDR with a lowercase internal $i$ (Kumar et al., 2023).

2. GLIDER in long-horizon LLM decision making

In the LLM-agent literature, GLIDR usually refers to GLIDER, which addresses the observation that LLMs “still struggle with long-horizon decision-making tasks due to deficient exploration and long-term credit assignment, especially in sparse-reward scenarios” (Hu et al., 26 May 2025). GLIDER introduces a two-level hierarchy: a high-level LLM policy that generates natural-language subtasks, and a low-level LLM controller that executes those subtasks as primitive environment actions. Both levels are implemented with a single frozen LLM backbone plus lightweight heads, with hierarchy selected only by short prompt differences such as a planner prompt versus an executor prompt.

The framework is formulated over an MDP

$$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$

where states are textual observations and actions are textual commands. The high-level policy $\pi_\theta^h$ acts every $c$ steps, producing subtasks $g_t$, while the low-level policy $\pi_\theta^l$ emits primitive actions $a_t$ conditioned on $g_t$. This introduces temporal abstraction through macro-rewards

$R_t = \sum_{k=t}^{t+c-1} r_k,$

and restructures offline trajectories into a high-level dataset $\mathcal{D}^h$ over macro-states and a low-level dataset $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$0 over primitive transitions with intrinsic rewards $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$1, where $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$2 is $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$3 if the low-level controller completed the current subtask and $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$4 otherwise. The paper states that these intrinsic rewards are derived automatically from observations, with no hand-coded reward (Hu et al., 26 May 2025).

Training proceeds in three stages: supervised fine-tuning or behavior cloning from hierarchically annotated demonstrations, offline hierarchical reinforcement learning, and optional offline-to-online fine-tuning. The actor uses LoRA adapters on the frozen backbone, the critic uses small MLP heads, the critic objective is an implicit-Q-learning-style expectile loss, and the actor objective is token-level AWAC-style advantage-weighted policy improvement. The stated purpose of the SFT stage is to align the model with valid environment actions and reasonably structured subtasks, while the offline RL stage addresses distribution shift through conservative value learning and behavior-regularized actor updates (Hu et al., 26 May 2025).

Empirically, GLIDER is evaluated on ScienceWorld and ALFWorld with Mistral-7B, Gemma-7B, and Llama-3-8B backbones. On Llama-3-8B, the best baseline on ScienceWorld is ETO at $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$5 for seen and unseen tasks, whereas GLIDER reaches $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$6, corresponding to reported improvements of $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$7 and $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$8. On ALFWorld, the best baseline is $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$9, while GLIDER reaches $\pi_\theta^h$0, with reported improvements of $\pi_\theta^h$1 and $\pi_\theta^h$2 (Hu et al., 26 May 2025).

Setting	Best baseline	GLIDER
ScienceWorld, Llama-3-8B, Seen/Unseen	57.90 / 52.33	77.43 / 68.34
ALFWorld, Llama-3-8B, Seen/Unseen	64.29 / 64.18	71.56 / 75.38

The ablation studies identify hierarchy as crucial. Hierarchical models outperform non-hierarchical ones on unseen ScienceWorld tasks across training regimes, the full SFT+ORL pipeline is best, and a hierarchical Llama-3B with SFT+ORL attains $\pi_\theta^h$3 on ScienceWorld-unseen, outperforming non-hierarchical Llama-7B baselines. The offline data mixture also matters: the best performance is reported at an expert:medium ratio of $\pi_\theta^h$4, with score $\pi_\theta^h$5 on ScienceWorld-unseen; using only expert trajectories gives $\pi_\theta^h$6, and using only medium data gives $\pi_\theta^h$7 (Hu et al., 26 May 2025). Online adaptation experiments on held-out ScienceWorld domains further report that GLIDER starts with higher initial performance and improves faster than flat AC and AWAC baselines.

3. GLIDR as differentiable graph-like rule learning

In knowledge-graph reasoning, GLIDR denotes “Graph-Like Inductive Logic Programming with Differentiable Reasoning” (Johnson et al., 8 Aug 2025). The central claim is that existing differentiable ILP methods such as Neural-LP, DRUM, and NLIL are restricted by chain-like rule structure, whereas many useful logical patterns are graph-shaped and include branches, shared variables, and cycles. GLIDR therefore generalizes chain-based rule learning to a more expressive graph-like rule syntax whose search space is controlled by a maximum number of free variables.

The formal rule template replaces a path-like body with a schema that can place a predicate $\pi_\theta^h$8 between every variable pair: $\pi_\theta^h$9 Specific rules are instantiated by selecting real predicates, inverse predicates, or a special always-true predicate $c$0 for each slot. This yields a maximally connected schematic template from which sparse graph-like logical bodies can be expressed. The paper’s examples include branching and recursive kinship rules as well as loopy biomedical rules (Johnson et al., 8 Aug 2025).

Inference is performed by a differentiable message-passing algorithm reminiscent of arc-consistency in CSPs. For each schema slot $c$1, GLIDR learns logits $c$2, converts them to a soft predicate distribution $c$3, and constructs a soft adjacency matrix

$c$4

Variable states $c$5 encode soft domains over entities. Messages are propagated between variables, and state updates are computed with element-wise minima, approximating set intersection under constraints. The resulting model is exact for trees and chains in the hard setting, but approximate for cyclic rules, where local consistency need not imply global satisfiability (Johnson et al., 8 Aug 2025).

GLIDR is trained as a binary classifier per target predicate using pairwise logistic ranking loss. Multiple rules can be learned per relation and reweighted by validation loss. The model also supports extraction of explicit symbolic rules from learned weights. The paper emphasizes that extracted rules retain significant predictive performance, and gives examples of disjunctive and recursive rule templates recovered from the soft model. A single extracted rule for the Alyawarra kinship term Agngiya is reported to achieve $c$6 in the transductive setting (Johnson et al., 8 Aug 2025).

Empirically, GLIDR is evaluated on Family, Alyawarra Kinships, UMLS, and FB15k-237. On the smaller datasets it substantially outperforms previous differentiable rule learners: on Kinships, the soft model reaches Hits@1 $c$7 and MRR $c$8, compared with DRUM at $c$9 Hits@1 and $g_t$0 MRR; on UMLS, the soft model reaches Hits@1 $g_t$1 and MRR $g_t$2 (Johnson et al., 8 Aug 2025). On FB15k-237, however, the reported run is weaker, with MRR $g_t$3 and Hits@1 $g_t$4, below DRUM and well below embedding methods. The paper attributes this either to a limited benefit from added rule expressivity on that benchmark or to a need for different hyperparameters and more manageable inference cost at scale.

The method is also reported to be highly robust to training-data noise and to support end-to-end coupling with deep neural networks for arbitrary modalities, illustrated by a patch-graph MNIST experiment. Its principal limitations are expensive open-query inference, no recursive unfolding at evaluation time, approximate inference on loopy rules, and separate rule-learning per target predicate rather than meta-generated rule parameters (Johnson et al., 8 Aug 2025).

4. GLiDR for sparse LiDAR point clouds

In autonomous-systems research, GLiDR denotes a “Topologically Regularized Graph Generative Network for Sparse LiDAR Point Clouds” (Kumar et al., 2023). The method addresses sparse dynamic LiDAR scans, which reduce the density of static points available for SLAM and navigation. The paper argues that, even under high sparsity, the global topology of the static environment can often still be inferred, and uses this observation to build a backbone skeleton interpreted as a single connected component that serves as a proxy for the scene’s global static topology.

The model takes a sparse dynamic scan $g_t$5 and generates an augmented scan $g_t$6 intended to approximate a corresponding static scan $g_t$7. The training representation begins with range images, then reshapes them into point sets. Each LiDAR point becomes a graph node with initial feature $g_t$8, and edges are formed by $g_t$9-nearest neighbors with $\pi_\theta^l$0. The graph is rebuilt at each graph layer using updated embeddings. The encoder consists of four LiDAR graph layers with message passing and max-pooling aggregation, and the decoder consists of four 1D convolution layers that map node features back to 3D coordinates (Kumar et al., 2023).

The distinctive regularizer is 0-dimensional persistent homology. The objective is to minimize the persistence of all connected components except the earliest-born one, thereby encouraging the generated point cloud and selected latent spaces to collapse onto a single dominant connected component. The total loss is

$\pi_\theta^l$1

where $\pi_\theta^l$2 and $\pi_\theta^l$3 are latent node embeddings and $\pi_\theta^l$4 is absolute error on range images. The paper explicitly states that no adversarial or contrastive losses are used; GLiDR is a deterministic graph autoencoder with topological regularization (Kumar et al., 2023).

The method is trained and evaluated on KITTI Odometry, CARLA-64, and ARD-16, in both dense and sparse regimes. It is designed to generate precise static points from scans that are $\pi_\theta^l$5 sparser than dense 64-beam LiDAR. Across the reported static-point-augmentation metrics—Chamfer Distance, Earth Mover’s Distance, Jensen–Shannon Divergence, RMSE, and MMD—GLiDR dominates or matches the baselines on almost all metrics and datasets, particularly under sparse settings. Representative values include CARLA-64 sparse CD $\pi_\theta^l$6 versus MOVES $\pi_\theta^l$7 and DSLR $\pi_\theta^l$8, and CARLA-64 sparse EMD $\pi_\theta^l$9 versus MOVES $a_t$0. On KITTI sparse, GLiDR reports CD $a_t$1, compared with MOVES $a_t$2 and DSLR $a_t$3 (Kumar et al., 2023).

The model also yields a binary static–dynamic segmentation as a byproduct: points in the input dynamic scan but absent from the generated static reconstruction are interpreted as dynamic, while generated points correspond either to static points present in the input or to static points previously occluded by dynamic objects. The paper further reports inference speed of approximately $a_t$4 ms per scan on an NVIDIA A100 GPU, with preprocessing at approximately $a_t$5 ms per scan on an Intel Xeon Silver 4208 CPU. Stated limitations include restricted domain generalization, dependence on dynamic–static training pairs, and the use of 0D homology only, since higher-dimensional persistent homology is too expensive for large LiDAR point clouds (Kumar et al., 2023).

5. GLIDE, sometimes queried as GLIDR, for prediction-powered inference

In evaluation research for generative AI and agentic systems, GLIDE is an open-source Python library that is explicitly described as “sometimes misspelled ‘GLIDR’” (Martinon et al., 29 May 2026). Its purpose is to make prediction-powered inference practical for mean estimation when ground-truth human labels are expensive and proxy labels such as LLM-as-judge scores are cheap but biased. The target estimand is the population mean

$a_t$6

with a small labeled set $a_t$7 and a much larger unlabeled set $a_t$8 equipped with proxy predictions $a_t$9.

GLIDE consolidates several estimators under a scipy-style API specialized to means: PPI++, Stratified PPI, Predict-Then-Debias and its stratified variants, and Active Statistical Inference. It also includes samplers for uniform, stratified, active, and cost-optimal sampling. A central point in the paper is that coverage remains valid regardless of proxy quality; a poor proxy yields wider intervals rather than invalid ones (Martinon et al., 29 May 2026).

The default large-sample estimator is PPI++, exposed as PPIMeanEstimator: $g_t$0 Here $g_t$1 recovers the classical labeled-only mean, $g_t$2 recovers the basic PPI estimator, and the optimal $g_t$3 minimizes asymptotic variance. For small labeled sample sizes, GLIDE recommends the bootstrap-based PTD estimators rather than CLT-based intervals. The decision rule given in the paper is operational: if there are at least $g_t$4 labeled samples overall, or per stratum in the stratified case, CLT-based estimators are recommended; if there are fewer than $g_t$5, PTD variants are recommended (Martinon et al., 29 May 2026).

The library ships with Monte Carlo validation notebooks. In the PTD example, the true mean is $g_t$6, the proxy mean is $g_t$7, the correlation $g_t$8 between truth and proxy varies from $g_t$9 to $R_t = \sum_{k=t}^{t+c-1} r_k,$0, the labeled sample size is $R_t = \sum_{k=t}^{t+c-1} r_k,$1, and the proxy sample size is $R_t = \sum_{k=t}^{t+c-1} r_k,$2. The reported findings are that PTD empirical coverage matches the $R_t = \sum_{k=t}^{t+c-1} r_k,$3 target across all $R_t = \sum_{k=t}^{t+c-1} r_k,$4, the CI width is approximately $R_t = \sum_{k=t}^{t+c-1} r_k,$5–$R_t = \sum_{k=t}^{t+c-1} r_k,$6 at $R_t = \sum_{k=t}^{t+c-1} r_k,$7 and drops to about $R_t = \sum_{k=t}^{t+c-1} r_k,$8 at $R_t = \sum_{k=t}^{t+c-1} r_k,$9, and the effective sample size grows from approximately $\mathcal{D}^h$0 to approximately $\mathcal{D}^h$1 (Martinon et al., 29 May 2026).

The agentic-system case study uses R-Judge, a dataset of $\mathcal{D}^h$2 conversations across five application domains with binary security-risk labels. The true risk rate is reported as $\mathcal{D}^h$3, the zero-shot proxy mean from claude-sonnet-4-6 is approximately $\mathcal{D}^h$4, and the proxy–truth correlation is $\mathcal{D}^h$5. With $\mathcal{D}^h$6 labeled trajectories, the $\mathcal{D}^h$7 CI width is approximately $\mathcal{D}^h$8 for the classical labeled-only estimator, $\mathcal{D}^h$9 for PPI++, $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$00 for ASI with active sampling, and $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$01 for stratified PPI++ with Neyman allocation. The corresponding effective sample sizes are reported as $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$02, $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$03, and $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$04 labeled trajectories, respectively (Martinon et al., 29 May 2026). The stated limitations are equally specific: the library targets only means, assumes a single proxy, provides no anytime-valid online guarantees, assumes i.i.d. labeled and unlabeled data, and leaves the annotation workflow external.

6. GLIDR-like infrastructures in gravitational-wave follow-up

In astronomy, the label “GLIDR-like systems” is used descriptively for end-to-end infrastructures that convert LIGO/Virgo alerts and galaxy catalogues into ranked target lists for rapid electromagnetic follow-up (Salmon et al., 2019). The concrete example in the cited paper is the HOGWARTs web application and its backend galaxy-ranking algorithm. Its operational motivation is the combination of large gravitational-wave localization regions, typically $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$05–$$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$06 deg$$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$07, and rapidly fading counterparts such as kilonovae; the paper notes that galaxy targeting can reduce pointings by factors of $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$08–$$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$09 relative to blind tiling.

The system uses LVC HEALPix skymaps from BAYESTAR and LALInference together with the GLADE V2 catalogue. The filtered GLADE V2 subset used in the work contains $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$10 galaxies with both distance and B-band magnitude. The preselection rule in the updated algorithm is to keep galaxies inside the $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$11 credible sky region and within

$$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$12

Scoring is then based on

$$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$13

with location score $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$14, luminosity weight $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$15, and normalized posterior

$$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$16

This ranking uses pixel-wise sky probability, pixel-wise Gaussian distance posteriors, galaxy distance, and B-band luminosity as a proxy for stellar mass (Salmon et al., 2019).

Implementation details are also explicit. The web application is built with the Python-based Flask framework, the backend uses ligo.skymap and astroplan, and results are stored in a PostgreSQL database on Amazon S3 with Heroku deployment. The reported backend ranking time is approximately $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$17–$$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$18 seconds per event, while website and database updates add approximately $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$19 seconds of overhead. The interface provides three workflows: retrieval of the ranked galaxy list, filtering by observatory visibility, and filtering by both visibility and detectability. For S190814bv, the system selected $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$20 galaxies from the $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$21 region and the resulting lists were used within ENGRAVE, GRAWITA, GW@WHT, and GROWTH follow-up campaigns (Salmon et al., 2019).

The paper also states the catalogue-completeness limits that constrain any such GLIDR-like system. Filtered GLADE V2 is $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$22 complete to approximately $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$23 Mpc and approximately $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$24 complete at approximately $$\langle \mathcal{S}, \mathcal{A}, \mathcal{T}, \mathcal{R}, \gamma \rangle,$$25 Mpc. The method therefore assigns zero probability to uncatalogued galaxies, which biases the discrete posterior over catalogued hosts, and it also relies on a 2D credible region plus global distance window rather than a fully volumetric 3D credible region (Salmon et al., 2019).

7. Comparative perspective

Taken together, these usages indicate that GLIDR is not a single research program but a family of unrelated names distributed across LLM agents, differentiable reasoning, point-cloud generation, statistical evaluation, and astronomical alert-response infrastructure. A common misconception is therefore purely lexical: the same character string can denote a misspelling, a valid acronym, or a near-homograph depending on capitalization and field.

A second, more substantive pattern is interpretive. This suggests that the name cluster repeatedly appears around methods that replace flat inference with an explicit intermediate structure. In GLIDER, the structure is a hierarchy of natural-language subtasks and low-level skills (Hu et al., 26 May 2025). In differentiable ILP GLIDR, it is a graph-like logical rule over schematic variables with message-passing constraint propagation (Johnson et al., 8 Aug 2025). In GLiDR for LiDAR, it is a single connected backbone enforced by 0D persistent homology (Kumar et al., 2023). In GLIDE, it is a debiased estimator that explicitly separates proxy prediction, sampling design, and uncertainty quantification (Martinon et al., 29 May 2026). In HOGWARTs as a GLIDR-like astronomy system, it is a ranked posterior over candidate host galaxies derived from sky localization, distance likelihood, and luminosity weighting (Salmon et al., 2019).

For citation and reading practice, the safest disambiguation rule is domain-specific. When the surrounding discussion concerns ScienceWorld, ALFWorld, hierarchical RL, or offline-to-online adaptation, GLIDR almost certainly refers to GLIDER (Hu et al., 26 May 2025). When the discussion concerns knowledge-graph completion, rule extraction, branches, or cycles, it denotes the differentiable ILP method GLIDR (Johnson et al., 8 Aug 2025). When the topic is sparse LiDAR, persistent homology, or static–dynamic segmentation, the relevant system is GLiDR (Kumar et al., 2023). When the topic is PPI, LLM-as-judge debiasing, or effective sample size, the intended library is GLIDE, even if queried as GLIDR (Martinon et al., 29 May 2026). When the phrase is “GLIDR-like” in gravitational-wave follow-up, it refers not to a fixed acronym but to the class of alert-to-galaxy-ranking infrastructures exemplified by HOGWARTs (Salmon et al., 2019).