Nitsum: Disambiguating 'CoRD' and 'CORD'

• Nitsum is a term addressing disambiguation challenges, highlighting the polysemous use of 'CoRD,' 'CORD,' and 'cord' across diverse fields.
• In biomedicine, 'cord' refers to both umbilical cord blood banking—with defined collection and preservation protocols—and spinal cord stimulation based on electrophysiological models.
• Other domains, such as topology, NLP, machine learning, and systems research, use the term to denote distinct mathematical constructs, structured datasets, and performance-driven frameworks.

Nitsum is not defined as a standardized technical term in the arXiv materials considered here. The surrounding literature instead records several unrelated usages of the forms “CoRD,” “CORD,” and “cord,” spanning umbilical cord blood banking, transverse string topology and knot invariants, spinal cord stimulation, COVID-19 research corpora, cross-modal reasoning for audio-LLMs, cooperative multi-agent reinforcement learning, RDMA networking, and multicore real-time scheduling (Sivakumaran et al., 2018, Basu et al., 2012, Wang et al., 2020, Hu et al., 23 Jan 2026, Planeta et al., 2023). This suggests that the relevant encyclopedic problem is one of disambiguation: the neighboring terminology is strongly polysemous, and its meaning is determined almost entirely by disciplinary context.

1. Terminological status and disambiguation

In the biomedical literature represented here, “CoRD” refers to umbilical cord blood and its banking (Sivakumaran et al., 2018). In low-dimensional topology, “cord” denotes homotopy classes of cords or transverse open strings associated with a framed codimension-2 submanifold, especially a knot in $\mathbb{R}^3$ (Basu et al., 2012, Petrak, 2019). In spinal neuromodulation, the salient anatomical referent is the spinal cord in transcutaneous spinal cord stimulation (Shneider et al., 2021). In scientific information retrieval and natural language processing, “CORD-19” is the COVID-19 Open Research Dataset, with derivatives such as CORD-NER and CORD-19-Vaccination (Wang et al., 2020, Wang et al., 2020, Singh et al., 2024). In machine learning and systems, “CORD” is reused as an acronym for distinct methods, including cross-modal distillation for large audio LLMs and role-diverse cooperation in MARL, while “CoRD” also names a converged RDMA dataplane (Hu et al., 23 Jan 2026, Matsuyama et al., 4 Jan 2025, Planeta et al., 2023).

A common misconception would be to treat these usages as variants of a single research program. The literature does not support that interpretation. The same token identifies unrelated objects: a stem-cell source, a knot invariant framework, a pandemic corpus, an audio-text alignment method, a networking architecture, and a real-time scheduling framework. A plausible implication is that any unqualified use of the neighboring forms “CoRD,” “CORD,” or “cord” is semantically unstable outside a clearly identified field.

2. Biomedical usages: cord blood and spinal cord modeling

In “Umbilical Cord Blood Banking and its Therapeutic Uses” (Sivakumaran et al., 2018), umbilical cord blood is the blood that remains in the umbilical cord and placenta after a baby is born. It is presented as a rich source of hematopoietic stem cells characterized by CD133, CD34, and CD45, together with mesenchymal stem/stromal cells, endothelial progenitor cells, unrestricted somatic stem cells, very small embryonic-like stem cells, multi lineage progenitor cells, neuronal progenitor cells, and natural killer cells. The paper emphasizes immunological immaturity, high plasticity, neutral differentiation capabilities, and the possibility of producing functional neural cells. Cord blood banking is defined as storing umbilical cord blood collected immediately after delivery, with three bank types—public, private, and direct donation—and with explicit collection and cryopreservation parameters such as clamping and transection within 10–30 seconds, storage at $4\,^\circ\mathrm{C}$ before processing within 24 hours, and freezing with 60% DMSO followed by storage in liquid nitrogen freezers. The article reports that nearly 80 diseases can be completely cured using umbilical cord blood stem cells and that over 50,000 transplants have been successfully carried out worldwide (Sivakumaran et al., 2018).

The same biomedical cluster also includes the spinal cord as an electrophysiological object rather than a banking resource. “Theoretical model of external spinal cord stimulation” (Shneider et al., 2021) presents an analytically tractable model for transcutaneous spinal cord stimulation over the lumbosacral cord. Axons are modeled as thin-walled cylindrical capacitors immersed in electrolyte, and external current pulses generate a transmembrane potential

$$U_m(\theta) = -\frac{2 a I_0}{\pi R_{\text{el}}^2 \sigma}\cos\theta,$$

leading to a current threshold

$$I_{\text{th}} = \frac{\pi R_{\text{el}}^2 \sigma}{2 a} U_{\text{th}}.$$

The model predicts $I_{\text{th}} \propto 1/a$, so larger-diameter axons have lower thresholds, and it interprets “multiple motor pools” as populations of motoneurons with different axon diameters. Using $R_{\text{el}} = 0.9\times 10^{-2}\ \text{m}$, $\sigma \approx 0.061\ \text{S/m}$, and $U_{\text{th}} \approx 10\ \text{mV}$, the paper reports threshold estimates of about 10.5 mA for $\alpha$-motoneurons, about 20 mA for $\beta$-motoneurons, and about 38 mA for $4\,^\circ\mathrm{C}$0-motoneurons, consistent in order of magnitude with segment-specific thresholds reported experimentally (Shneider et al., 2021).

These two biomedical usages are conceptually unrelated except for the anatomical word “cord.” One concerns progenitor-cell collection, cryopreservation, and transplantation; the other concerns membrane charging, recruitment thresholds, and neuromodulation. This suggests that biomedical disambiguation requires attention not merely to discipline, but to subfield.

3. Topological usages: cord algebra and transverse string topology

In topology, “cord” has a precise geometric-algebraic meaning. “Transverse string topology and the cord algebra” (Basu et al., 2012) studies smooth maps $4\,^\circ\mathrm{C}$1 with endpoints on a framed codimension-2 submanifold $4\,^\circ\mathrm{C}$2, interior transverse to $4\,^\circ\mathrm{C}$3, and a rigid radial behavior near endpoints and internal intersections determined by the framing and two antipodal sections $4\,^\circ\mathrm{C}$4. The singular chain complex on the corresponding string space $4\,^\circ\mathrm{C}$5 is equipped with two operations: a resolve $4\,^\circ\mathrm{C}$6 at an internal intersection and a split coproduct $4\,^\circ\mathrm{C}$7. Together with the singular boundary $4\,^\circ\mathrm{C}$8, these define a differential graded coalgebra

$4\,^\circ\mathrm{C}$9

After cobar construction, the degree-zero homology $$U_m(\theta) = -\frac{2 a I_0}{\pi R_{\text{el}}^2 \sigma}\cos\theta,$$0 is identified with a twisted group ring of the knot complement modulo longitude relations, and the resulting string algebra $$U_m(\theta) = -\frac{2 a I_0}{\pi R_{\text{el}}^2 \sigma}\cos\theta,$$1 is shown to be isomorphic to a specialization of Ng’s cord algebra as a $$U_m(\theta) = -\frac{2 a I_0}{\pi R_{\text{el}}^2 \sigma}\cos\theta,$$2-algebra (Basu et al., 2012). In the knot case, the construction yields a non-trivial knot invariant not determined by the Alexander polynomial, Jones polynomial, HOMFLY polynomial, Kauffman polynomial, signature, Khovanov homology, or Ozsváth–Szabó invariants (Basu et al., 2012).

“Definition of the cord algebra of knots using Morse Theory” (Petrak, 2019) recasts the same invariant in Morse-theoretic terms. A linear cord is a straight line segment in $$U_m(\theta) = -\frac{2 a I_0}{\pi R_{\text{el}}^2 \sigma}\cos\theta,$$3 with endpoints on a knot $$U_m(\theta) = -\frac{2 a I_0}{\pi R_{\text{el}}^2 \sigma}\cos\theta,$$4, so the cord space is identified with $$U_m(\theta) = -\frac{2 a I_0}{\pi R_{\text{el}}^2 \sigma}\cos\theta,$$5. On this torus-like space one studies the Morse perturbation of the energy

$$U_m(\theta) = -\frac{2 a I_0}{\pi R_{\text{el}}^2 \sigma}\cos\theta,$$6

whose non-diagonal critical points are binormal chords. The algebra $$U_m(\theta) = -\frac{2 a I_0}{\pi R_{\text{el}}^2 \sigma}\cos\theta,$$7 is the noncommutative unital $$U_m(\theta) = -\frac{2 a I_0}{\pi R_{\text{el}}^2 \sigma}\cos\theta,$$8-algebra generated by index-0 critical points, while $$U_m(\theta) = -\frac{2 a I_0}{\pi R_{\text{el}}^2 \sigma}\cos\theta,$$9 is the free abelian group on index-1 critical points. A map $$I_{\text{th}} = \frac{\pi R_{\text{el}}^2 \sigma}{2 a} U_{\text{th}}.$$0 is built by following unstable manifolds under the negative gradient flow and applying four geometric relations corresponding to contractible cords, framing intersections, base-point crossings, and splitting at knot intersections. The cord algebra is then defined as

$$I_{\text{th}} = \frac{\pi R_{\text{el}}^2 \sigma}{2 a} U_{\text{th}}.$$1

The paper computes the unknot algebra as $$I_{\text{th}} = \frac{\pi R_{\text{el}}^2 \sigma}{2 a} U_{\text{th}}.$$2 and gives an explicit noncommutative presentation for the right-handed trefoil (Petrak, 2019).

Here the term “cord” is neither anatomical nor acronymic. It denotes a path-like or segment-like geometric object whose algebraic relations encode fine structure of the knot complement. A plausible implication is that this is the most historically stable mathematical usage among the materials surveyed.

4. Corpus and NLP usages: CORD-19 and derived resources

In the COVID-19 computational literature, “CORD” most commonly refers to the COVID-19 Open Research Dataset. “CORD-19: The COVID-19 Open Research Dataset” (Wang et al., 2020) describes a continuously updated corpus of scientific publications and preprints on COVID-19 and related coronaviruses, released in March 2020 by the Allen Institute for AI and partners. As of 2020-06-14 it contained over 140,000 papers, more than 72,000 full-text papers, and 47,000+ 2020 papers with 7,000+ preprints. The corpus integrates PMC, PubMed, the WHO COVID-19 Database, bioRxiv, medRxiv, arXiv, and direct publisher feeds, and it represents full text in S2ORC JSON after parsing PDFs with GROBID or PMC JATS XML with a custom parser. The dataset supports text mining, information retrieval, question answering, summarization, knowledge graph construction, and shared tasks such as the Kaggle CORD-19 Research Challenge and TREC-COVID (Wang et al., 2020).

“Comprehensive Named Entity Recognition on CORD-19 with Distant or Weak Supervision” (Wang et al., 2020) builds CORD-NER on the 2020-03-13 snapshot of CORD-19, using 29,500 documents and covering 75 fine-grained entity types. Its annotation pipeline combines four sources: SpaCy general NER, SciSpacy BioNER, UMLS-based distantly supervised NER, and seed-guided weak supervision using CatE for new COVID-specific types. The paper reports evaluation on more than 1000 manually annotated sentences from the COVID-19 corpus. CORD-NER achieves overall precision 81.29, recall 73.65, and F1 77.28, compared with overall F1 62.98 for SciSpacy(BIONLP13CG) and 65.66 for SciSpacy(BC5CDR); on chemical and disease recognition it outperforms SciSpacy, while gene F1 remains lower than SciSpacy(BIONLP13CG) (Wang et al., 2020).

“Constructing the CORD-19 Vaccine Dataset” (Singh et al., 2024) introduces CORD-19-Vaccination as a curated subset focused on COVID-19 vaccine research. Starting from CORD-19 version 109, with “over one million journals” in metadata.csv, the authors filter for publish_time >= '2020', vaccine-related title or abstract patterns across languages, presence of full-text JSON, and non-null abstracts, yielding about 30,000 research papers. They add language fields (lang_id, lang_id_confidence, lang_id_predictions) using fastText, author-demography fields (aff_lab_inst, aff_location, aff_country) with JSON parsing and Google search augmentation, YAKE keywords, LDA-based topics, and labeled_abstract from sequential sentence classification. Coverage of country-of-affiliation is increased from about 63% to 93%, and five topic groups are reported: Vaccine development (20%), Vaccination side-effects (14%), Vaccination efficacy (16%), Methodologies for COVID studies (25%), and Vaccine uptake (25%) (Singh et al., 2024).

This corpus lineage illustrates a distinct semantic pattern: CORD is an acronymic label for data infrastructure rather than a domain object. It also shows how one acronym can anchor a hierarchy of derivative resources, evaluation tasks, and annotation layers.

5. Machine-learning usages: cross-modal reasoning and cooperative MARL

“CORD: Bridging the Audio-Text Reasoning Gap via Weighted On-policy Cross-modal Distillation” (Hu et al., 23 Jan 2026) defines CORD as a unified alignment framework for large audio LLMs. The method aligns audio-conditioned reasoning with text-conditioned reasoning inside a single model, using the text modality as an internal teacher. Token-level alignment uses on-policy reverse KL divergence with importance-aware top-$$I_{\text{th}} = \frac{\pi R_{\text{el}}^2 \sigma}{2 a} U_{\text{th}}.$$3 weighting and position-based decay, while sequence-level alignment uses a judge-based global reward optimized with Group Relative Policy Optimization. The final objective is

$$I_{\text{th}} = \frac{\pi R_{\text{el}}^2 \sigma}{2 a} U_{\text{th}}.$$4

Using only 80k synthetic training samples, the method reduces the average audio-text gap on Qwen2-Audio-7B-Instruct from 15.25 to 8.90, a 41.6% reduction, and on Step-Audio2-mini from 10.86 to 6.00, a 44.8% reduction; it also preserves general audio capabilities better than supervised fine-tuning or forward-KL distillation on MMAU (Hu et al., 23 Jan 2026).

A different use appears in “CORD: Generalizable Cooperation via Role Diversity” (Matsuyama et al., 4 Jan 2025), where CORD is a hierarchical MARL method for cooperative Dec-POMDPs. A high-level controller assigns latent roles to low-level agents by maximizing constrained role entropy, and the authors show that this objective decomposes into causal influence in role and role heterogeneity. The intrinsic rewards are

$$I_{\text{th}} = \frac{\pi R_{\text{el}}^2 \sigma}{2 a} U_{\text{th}}.$$5

and

$$I_{\text{th}} = \frac{\pi R_{\text{el}}^2 \sigma}{2 a} U_{\text{th}}.$$6

combined with the environment reward as $$I_{\text{th}} = \frac{\pi R_{\text{el}}^2 \sigma}{2 a} U_{\text{th}}.$$7. Across resource collection, MPE, and SMAC tasks, the method improves both training performance and generalization to unseen teammates and team sizes; in SMAC generalization it achieves the best win rate in 8 of 12 tasks reported (Matsuyama et al., 4 Jan 2025).

These two works share the acronym “CORD” but not their mathematical objects, datasets, or objectives. One addresses cross-modal alignment in autoregressive reasoning; the other addresses role-conditioned cooperation in hierarchical MARL. The overlap is purely nominal.

6. Systems and real-time computing usages: RDMA dataplanes and resource-aware scheduling

In systems research, “CoRD” denotes a redesign of the RDMA stack. “CoRD: Converged RDMA Dataplane for High-Performance Clouds” (Planeta et al., 2023) argues that kernel bypass is not mandatory for high-performance RDMA because zero-copy and polling matter more than bypass itself. CoRD routes RDMA dataplane operations back through the kernel while preserving the verbs API and direct DMA from registered user memory. The prototype modifies Linux 6.0.0-rc7, with about 250 lines added or changed in the kernel-level mlx5 driver and about 20 lines in the user-level RDMA driver. Microbenchmark dissection attributes only a small constant overhead of about 0.07 $$I_{\text{th}} = \frac{\pi R_{\text{el}}^2 \sigma}{2 a} U_{\text{th}}.$$8s to the removal of kernel bypass in the emulation experiment, whereas removing busy-polling adds several microseconds for small messages and removing zero-copy dominates large-message cost. In NAS Parallel Benchmarks on Azure HB120, CoRD’s runtime is effectively indistinguishable from kernel-bypass RDMA, while IP over InfiniBand is up to 2$$I_{\text{th}} = \frac{\pi R_{\text{el}}^2 \sigma}{2 a} U_{\text{th}}.$$9 slower (Planeta et al., 2023).

A separate real-time systems usage appears in “CORD: Co-design of Resource Allocation and Deadline Decomposition with Generative Profiling” (Gifford et al., 14 Jan 2025). There CORD is a framework for DAG-based multicore real-time tasks with partitionable shared LLC and memory bandwidth. It combines a generative resource-profiling algorithm based on a multimarginal Schrödinger bridge with an offline co-allocation method that jointly assigns resource budgets and decomposed deadlines under global EDF. The shared resources are partitioned into $I_{\text{th}} \propto 1/a$0 cache partitions and $I_{\text{th}} \propto 1/a$1 bandwidth partitions, while profiling uses only $I_{\text{th}} \propto 1/a$2 and $I_{\text{th}} \propto 1/a$3 profiles per budget to synthesize profiles for all 400 budgets. The execution model is phase-based,

$I_{\text{th}} \propto 1/a$4

with worst-case execution time

$I_{\text{th}} \propto 1/a$5

The paper reports that solving the generative profiling problem and generating synthetic profiles for all budgets takes about 15 minutes per benchmark, and that CoRD-DA substantially improves schedulability over the Stretch baseline, with all task sets schedulable at utilization 2.0 in one reported 4-core setting where Stretch schedules none (Gifford et al., 14 Jan 2025).

This systems cluster again demonstrates acronymic reuse with no shared ontology. One CoRD is about reinstating kernel mediation in RDMA; the other CORD is about coupling stochastic profiling with deadline decomposition in multicore real-time analysis.

7. Comparative assessment

Taken together, the literature indicates that Nitsum has no established technical identity in the sources surveyed, while the neighboring forms “CoRD,” “CORD,” and “cord” are highly domain-specific. In biomedicine they denote either umbilical cord blood banking or spinal cord stimulation; in topology they denote cords of knots and associated algebras; in information science they denote a large COVID-19 corpus and its derivatives; in machine learning they denote cross-modal alignment or role-diverse cooperation; in systems they denote either a converged RDMA dataplane or a co-design framework for resource-aware real-time scheduling (Sivakumaran et al., 2018, Shneider et al., 2021, Basu et al., 2012, Petrak, 2019, Wang et al., 2020, Wang et al., 2020, Singh et al., 2024, Hu et al., 23 Jan 2026, Matsuyama et al., 4 Jan 2025, Planeta et al., 2023, Gifford et al., 14 Jan 2025).

A plausible implication is that the only rigorous treatment of Nitsum available from these materials is a disambiguating one. The neighboring terminology is not merely broad; it is structurally heterogeneous, crossing anatomy, algebraic topology, corpus construction, neural alignment, cooperative control, operating systems, and schedulability theory. Any future stabilization of the term would therefore require an explicit definition tied to one of these lineages rather than reliance on lexical resemblance alone.