Opus: A Multifaceted Research Term

• Opus is a polysemous term, defining innovations from royalty-free low-latency speech/audio codecs to advanced neural resynthesis for low-bitrate applications.
• It underpins workflow generation and evaluation frameworks that use DAGs, intention capture, and quantitative metrics to optimize complex task execution.
• Opus also spans formal reasoning, machine translation, and systems coordination in fields like ML networking, astronomy, and orbital policy modeling.

Searching arXiv for the provided Opus-related papers to ground the article in current records. “Opus” is a polysemous designation in contemporary technical literature rather than a single system. In arXiv-indexed work it denotes, among other things, a royalty-free, low-latency speech/audio codec standardized in RFC 6716; multi-channel compression methods built on that codec for far-field ASR; workflow-generation and workflow-evaluation frameworks centered on Intention capture and DAG-structured execution; Anthropic’s Claude Opus 4.6 as used in formal theorem proving and pentesting studies; an optimizer-aware data-selection method for LLM pre-training; the OPUS ecosystem for neural machine translation; an interoperable astronomical job-control service; a control plane for photonic rails in ML datacenters; and an integrated assessment model for orbital debris and satellite economics (Valin et al., 2016, Drude et al., 2021, Fagnoni et al., 2024, Fagnoni et al., 15 Jul 2025, Seroul et al., 6 Nov 2025, Baudart et al., 20 Mar 2026, Wang et al., 5 Feb 2026, Tiedemann et al., 2022, Servillat et al., 2021, Ding et al., 13 Feb 2026, Rao et al., 2023).

1. Scope and disambiguation

A recurrent misconception is that “Opus” refers only to the Internet audio codec. In the research record represented here, that is only one major usage. The name is attached to systems in speech coding, LLM workflow synthesis, theorem-proving agents, MT infrastructure, datacenter networking, astronomy middleware, and orbital policy modeling.

Domain	Formulation of “Opus”	Representative paper
Audio coding	Speech/audio codec with SILK and CELT	(Valin et al., 2016)
Far-field ASR	Multi-channel Opus compression under fixed bitrate	(Drude et al., 2021)
Workflow generation	Large Work Model with Work Knowledge Graph	(Fagnoni et al., 2024)
Prompted intention capture	Workflow Signal and Workflow Intention framework	(Fagnoni et al., 15 Jul 2025)
Workflow evaluation	Reward–Penalty framework for workflow ranking	(Seroul et al., 6 Nov 2025)
Formal reasoning	Claude Opus 4.6 with Rocq-MCP	(Baudart et al., 20 Mar 2026)
LLM pre-training	Optimizer-induced Projected Utility Selection	(Wang et al., 5 Feb 2026)
Machine translation	OPUS ecosystem and OPUS-MT	(Tiedemann et al., 2022)
Astronomy infrastructure	Observatoire de Paris UWS System	(Servillat et al., 2021)
ML networking	Control plane for photonic rails	(Ding et al., 13 Feb 2026)
Space policy	Orbital Debris Propagators Unified with Economic Systems	(Rao et al., 2023)

This breadth has methodological significance. In some papers “Opus” names a codec format and its algorithmic variants; in others it names a framework, a service, or an acronym whose expansion defines the research problem itself. The commonality is nominal rather than architectural.

2. Opus as a low-delay speech and audio codec

In codec research, Opus is a royalty-free, low-latency speech/audio codec standardized in RFC 6716. Its two principal coding engines are SILK, a linear-predictive voice codec inherited from Skype, and CELT, a modified MDCT transform coder optimized for music and general audio. The encoder can operate in pure SILK mode, pure CELT mode, or hybrid mode, deciding frame by frame which mode gives the best trade-off between distortion and bitrate (Valin et al., 2016).

The CELT side of Opus is designed around very low algorithmic delay. It uses an $N$-point MDCT with a fixed $2.5\,\mathrm{ms}$ overlap and a “flat top” symmetric, power-complementary window borrowed from Vorbis. The same paper describes critical-band energy coding, implicit masking through static allocation curves, allocation tilt, band boosts, PVQ on normalized band vectors, and deterministic frame-level CBR logic without a bit reservoir in CBR mode. End-to-end, the coding chain includes pre-emphasis, optional pitch pre-filtering, transient handling through multiple short MDCTs, energy quantization, bit allocation, PVQ, stereo coupling, entropy coding, and internal framing (Valin et al., 2016).

The same body of work also makes clear that Opus is not a monolith. CELT emphasizes transform coding and psychoacoustic design for music, whereas SILK emphasizes waveform-oriented speech coding. A later low-bitrate study states explicitly that Opus voice mode is “at its core a waveform matching coder,” and that quality degrades quickly below $10\,\mathrm{kb/s}$, where parametric coders tend to perform better than waveform coders (Skoglund et al., 2019).

That low-bitrate limitation motivated backward-compatible neural enhancement. In “Improving Opus Low Bit Rate Quality with Neural Speech Synthesis,” decoded Opus parameters at $6\,\mathrm{kb/s}$ are used as conditioning features for WaveNet and LPCNet. The conditioning vector is

$$\theta_f = [\,\mathbf{c}_f^{(LPC)};\;\mathbf{c}_f^{(W)};\;\tau_f;\;G_f\,],$$

with 18 cepstral coefficients from LPC, 18 cepstral coefficients from the reconstructed waveform, pitch lag, and pitch gain summary. The reported listening test shows that synthesized speech using LPCNet clearly outperforms the standard Opus decoder for the same $6\,\mathrm{kb/s}$ bit stream, while WaveNet provides a higher-quality but impractical upper bound because of its complexity and latency (Skoglund et al., 2019).

A plausible implication is that “Opus” in codec literature should be understood as a format whose bitstream semantics can support substantially different decoder behaviors, including generative neural resynthesis, without breaking compatibility.

3. Multi-channel Opus for far-field ASR

The far-field ASR paper treats Opus not as an endpoint codec for perceptual listening quality, but as a transport bottleneck in cloud ASR. Cloud ASR permits larger models and more powerful multi-channel front-ends than on-device processing, but transmission of multiple microphone-array channels adds inherent latency and bandwidth cost. Client-side Opus compression can reduce bandwidth, yet it may damage beamforming performance by distorting phase and spatial information (Drude et al., 2021).

The paper’s central observation is that off-the-shelf Opus is tuned for subjective quality rather than beamforming-critical phase accuracy. It therefore modifies joint channel coding in two ways: CELT intensity stereo and folding are disabled as much as allowed, and SILK’s rate-distortion optimizer is changed to value accurate waveform reconstruction over perceptual masking. The modified optimization is written as

$$\min_{R_m+R_s=R_{\rm total}} D_{\rm wave}(m,s;\,\hat m,\hat s)+\lambda\,(R_m+R_s),$$

where $D_{\rm wave}$ is mean squared error in the time domain rather than the perceptual distortion function normally used (Drude et al., 2021).

To decorrelate more than two channels, the paper introduces a fixed, array-geometry-agnostic transform across the $D$ real channels:

$X_k(t)=\sum_{d=0}^{D-1} x_d(t)\,e^{-j2\pi dk/D},$

with inverse

$2.5\,\mathrm{ms}$0

Because real-input DFT coefficients are conjugate-symmetric, the transform produces one purely real DC channel and complex AC channels. In the reported power-concentration experiments, the DC channel carries approximately $2.5\,\mathrm{ms}$1 of total power for a typical compact seven-microphone array, making uniform bit allocation across transformed channels materially more efficient than independent coding of each physical channel (Drude et al., 2021).

The experimental setup used seven closely spaced omnidirectional sensors sampled at $2.5\,\mathrm{ms}$2, a mask-based MVDR beamformer, and a single-channel RNN-T acoustic model with an 8-layer LSTM encoder, 2-layer LSTM decoder, and 4K wordpieces, trained on approximately 60 K hours of on-device beamformed audio with SpecAugment. On 126 h of de-identified realistic far-field recordings, independent coding at $2.5\,\mathrm{ms}$3 yielded normalized WER $2.5\,\mathrm{ms}$4 relative to the uncompressed single-channel baseline. DFT decorrelation plus uniform allocation improved nWER to approximately $2.5\,\mathrm{ms}$5, and “DFT+pairwise” real-imaginary pairing improved it to approximately $2.5\,\mathrm{ms}$6, corresponding to a $2.5\,\mathrm{ms}$7 relative WER reduction. Equivalently, holding nWER at $2.5\,\mathrm{ms}$8, the DFT+pairwise method reduced bitrate from $2.5\,\mathrm{ms}$9 to $10\,\mathrm{kb/s}$0, a $10\,\mathrm{kb/s}$1 reduction (Drude et al., 2021).

This use of Opus is thus not primarily about listener-facing perceptual transparency. It is about preserving fine phase cues under a fixed client-to-cloud bitrate budget so that downstream MVDR beamforming and ASR remain effective.

4. Opus in workflow generation and workflow evaluation

Several workflow papers use “Opus” for systems that formalize intent, task structure, and optimization rather than audio. In “Opus: A Large Work Model for Complex Workflow Generation,” Opus is a framework for generating and optimizing executable workflows for complex BPO use cases. A workflow is represented as a DAG $10\,\mathrm{kb/s}$2 whose nodes are Tasks, each Task being a sequence of executable Instructions such as model calls, external tool calls, or human expert review steps. The architecture combines an Intention Encoder, a Work Knowledge Graph $10\,\mathrm{kb/s}$3, a Large Work Model, and a graph-based optimization phase. The Work Knowledge Graph weights historical task transitions by

$10\,\mathrm{kb/s}$4

and the final workflow path is selected by minimizing a composite task cost with a modified Dijkstra algorithm (Fagnoni et al., 2024).

That paper’s medical-coding evaluation is unusually explicit. On CPT Evaluation / Management outpatient coding, using multi-modal clinical records and a reference workflow crafted by senior medical coders, Opus 1α Large obtained Coverage $10\,\mathrm{kb/s}$5, Kendall’s Tau $10\,\mathrm{kb/s}$6, DTW $10\,\mathrm{kb/s}$7, Cosine $10\,\mathrm{kb/s}$8, and BLEU $10\,\mathrm{kb/s}$9, while Opus 1α Small obtained Coverage $6\,\mathrm{kb/s}$0, Kendall’s Tau $6\,\mathrm{kb/s}$1, DTW $6\,\mathrm{kb/s}$2, Cosine $6\,\mathrm{kb/s}$3, and BLEU $6\,\mathrm{kb/s}$4. Averaged across all five metrics via the pentagon-area approach, Opus Large surpassed the best baseline by $6\,\mathrm{kb/s}$5, and Opus Small by $6\,\mathrm{kb/s}$6 (Fagnoni et al., 2024).

A second workflow paper introduces the “Opus Prompt Intention Framework,” which places an intermediate Intention Capture layer between the raw user query and workflow generation. It decomposes processing into Signal Extraction, which produces Workflow Signals for Input, Process, and Output, and Intention Generation, which partitions those signals into one or more Workflow Intention objects $6\,\mathrm{kb/s}$7. The workflow generator then samples from

$6\,\mathrm{kb/s}$8

rather than directly from $6\,\mathrm{kb/s}$9 (Fagnoni et al., 15 Jul 2025).

On a synthetic benchmark of 1,000 multi-intent query–workflow pairs with Mixed Intention Levels $$\theta_f = [\,\mathbf{c}_f^{(LPC)};\;\mathbf{c}_f^{(W)};\;\tau_f;\;G_f\,],$$0, intention-guided generation improved BLEU by $$\theta_f = [\,\mathbf{c}_f^{(LPC)};\;\mathbf{c}_f^{(W)};\;\tau_f;\;G_f\,],$$1–$$\theta_f = [\,\mathbf{c}_f^{(LPC)};\;\mathbf{c}_f^{(W)};\;\tau_f;\;G_f\,],$$2, METEOR by $$\theta_f = [\,\mathbf{c}_f^{(LPC)};\;\mathbf{c}_f^{(W)};\;\tau_f;\;G_f\,],$$3–$$\theta_f = [\,\mathbf{c}_f^{(LPC)};\;\mathbf{c}_f^{(W)};\;\tau_f;\;G_f\,],$$4, BERTScore by $$\theta_f = [\,\mathbf{c}_f^{(LPC)};\;\mathbf{c}_f^{(W)};\;\tau_f;\;G_f\,],$$5–$$\theta_f = [\,\mathbf{c}_f^{(LPC)};\;\mathbf{c}_f^{(W)};\;\tau_f;\;G_f\,],$$6, cosine similarity by up to $$\theta_f = [\,\mathbf{c}_f^{(LPC)};\;\mathbf{c}_f^{(W)};\;\tau_f;\;G_f\,],$$7, and LLM-as-Judge Coverage, Consistency, Integration, and Total scores by $$\theta_f = [\,\mathbf{c}_f^{(LPC)};\;\mathbf{c}_f^{(W)};\;\tau_f;\;G_f\,],$$8–$$\theta_f = [\,\mathbf{c}_f^{(LPC)};\;\mathbf{c}_f^{(W)};\;\tau_f;\;G_f\,],$$9. The same paper reports that as $6\,\mathrm{kb/s}$0 increased, naïve generation collapsed toward zero while intention-guided generation remained stable, with cosine similarity staying above $6\,\mathrm{kb/s}$1 even at $6\,\mathrm{kb/s}$2 (Fagnoni et al., 15 Jul 2025).

A third paper, “Opus: A Quantitative Framework for Workflow Evaluation,” formalizes workflow quality through a probabilistic reward and a normative penalty. For a workflow $6\,\mathrm{kb/s}$3, the expected net-benefit reward is

$6\,\mathrm{kb/s}$4

where $6\,\mathrm{kb/s}$5 aggregates cumulative resources, execution duration on the critical path, and peak releasable resources. Structural and informational defects are captured by Cohesion, Coupling, Observability, and Information Hygiene, which are combined into the Cohesive Independence Penalty, the Signal Integrity Penalty, and the overall penalty $6\,\mathrm{kb/s}$6 (Seroul et al., 6 Nov 2025).

The case study on email-to-ticket classification makes the framework concrete. For three candidate workflows, the reported values were: $6\,\mathrm{kb/s}$7 with $6\,\mathrm{kb/s}$8, $6\,\mathrm{kb/s}$9, and $$\min_{R_m+R_s=R_{\rm total}} D_{\rm wave}(m,s;\,\hat m,\hat s)+\lambda\,(R_m+R_s),$$0; $$\min_{R_m+R_s=R_{\rm total}} D_{\rm wave}(m,s;\,\hat m,\hat s)+\lambda\,(R_m+R_s),$$1 with $$\min_{R_m+R_s=R_{\rm total}} D_{\rm wave}(m,s;\,\hat m,\hat s)+\lambda\,(R_m+R_s),$$2, $$\min_{R_m+R_s=R_{\rm total}} D_{\rm wave}(m,s;\,\hat m,\hat s)+\lambda\,(R_m+R_s),$$3, and $$\min_{R_m+R_s=R_{\rm total}} D_{\rm wave}(m,s;\,\hat m,\hat s)+\lambda\,(R_m+R_s),$$4; and $$\min_{R_m+R_s=R_{\rm total}} D_{\rm wave}(m,s;\,\hat m,\hat s)+\lambda\,(R_m+R_s),$$5 with $$\min_{R_m+R_s=R_{\rm total}} D_{\rm wave}(m,s;\,\hat m,\hat s)+\lambda\,(R_m+R_s),$$6, $$\min_{R_m+R_s=R_{\rm total}} D_{\rm wave}(m,s;\,\hat m,\hat s)+\lambda\,(R_m+R_s),$$7, and $$\min_{R_m+R_s=R_{\rm total}} D_{\rm wave}(m,s;\,\hat m,\hat s)+\lambda\,(R_m+R_s),$$8. The ranking is $$\min_{R_m+R_s=R_{\rm total}} D_{\rm wave}(m,s;\,\hat m,\hat s)+\lambda\,(R_m+R_s),$$9, because $D_{\rm wave}$0 and $D_{\rm wave}$1 tie on reward but $D_{\rm wave}$2 has the lower penalty (Seroul et al., 6 Nov 2025).

Taken together, these workflow papers use “Opus” for a family of ideas centered on explicit intent representation, DAG-structured execution, domain knowledge injection, and quantitative workflow selection.

5. Claude Opus in formal reasoning and pentesting

In a different strand, “Opus” refers to Anthropic’s Claude Opus model family rather than a workflow or codec system. In “Putnam 2025 Problems in Rocq using Opus 4.6 and Rocq-MCP,” Claude Opus 4.6 is evaluated as a tool-augmented theorem-proving agent. Architecture details are not publicly disclosed in the paper; the report states that Anthropic’s documentation indicates a deep Transformer with broad context capabilities, optimized for interactive tool use via Claude Code. The experiment used Rocq-MCP, which provides eight MCP tools split into compilation tools and interactive tools, and follows a “compile-first, interactive-fallback” strategy (Baudart et al., 20 Mar 2026).

The reported outcome was 10 of 12 Putnam 2025 problems solved, with 5,542 lines of Rocq proof code. The environment was an isolated VM with no internet access. Over 3 days, Claude Code launched 141 subagents across four expert roles, consuming approximately $D_{\rm wave}$3 tokens at an API cost of \$5,279, with 17.7 h of active compute and 51.6 h wall-clock elapsed. The paper also records an A3 loophole, in which the agent initially proved the wrong statement because the encoding allowed a “do nothing” strategy, and two unsolved problems, A5 and B6, where even 82–91 subagents and 800 M tokens failed to produce a valid proof (Baudart et al., 20 Mar 2026).

The pentesting paper studies Claude Opus in a PTES-aligned cybersecurity setting. The environment was the GOAD laboratory, simulating a complex Windows Active Directory infrastructure with five virtual machines across three domains, two forests, and dozens of user accounts. Claude Opus and GPT-4 were accessed via the Perplexity platform, while Copilot was used via its web interface. Claude Opus was reported as useful across reconnaissance, vulnerability analysis, exploitation, post-exploitation, and reporting, including full-port Nmap scans, enum4linux, rpcclient, NetExec, CrackMapExec, Kerberoasting, AS-REP Roasting, AD CS abuse with Certipy, Kerberos Relay Up, RBCD workflows, and report drafting with mitigation recommendations (Martínez et al., 12 Jan 2025).

That paper is careful about its quantitative status. Because the original study does not publish raw timings or precise success-failure counts, the reported task-level metrics are derived from the authors’ assessments. Under those approximate metrics, Claude Opus achieved $D_{\rm wave}$4, compared with approximately $D_{\rm wave}$5 for GPT-4 and $D_{\rm wave}$6 for Copilot; time savings were approximately $D_{\rm wave}$7 for Claude Opus, $D_{\rm wave}$8 for GPT-4, and $D_{\rm wave}$9 for Copilot; and error rates were approximately $D$0, $D$1, and $D$2, respectively (Martínez et al., 12 Jan 2025).

A common misconception in this context is to treat Claude Opus as a fully autonomous replacement for expert operators. Both papers point in the opposite direction. In theorem proving, the dominant pattern is iterative proof writing plus compile–verify cycles, with interactive tools used mainly for debugging; in pentesting, the study concludes that the tools cannot fully automate the process and that human validation remains necessary.

6. OPUS in data selection, translation infrastructure, and astronomical job control

The acronym OPUS is also used for systems whose scope is infrastructural rather than agentic. In “OPUS: Towards Efficient and Principled Data Selection in LLM Pre-training in Every Iteration,” OPUS expands to Optimizer-induced Projected Utility Selection. The method defines sample utility in optimizer-induced update space rather than raw gradient space. At step $D$3, the effective update for a subset $D$4 is

$D$5

and the basic projected utility is the inner product between a candidate’s one-step update and a proxy direction derived from an in-distribution proxy pool. Scalability is obtained with Ghost technique and CountSketch, while diversity is maintained through Boltzmann sampling (Wang et al., 5 Feb 2026).

The empirical claims are strong and highly specific. OPUS adds only $D$6 extra compute overhead. In GPT-2 XL pre-training on FineWeb with 30 B tokens, average zero-shot accuracy across 10 benchmarks rose to $D$7 versus $D$8 for QuRating, $D$9 for random selection, $X_k(t)=\sum_{d=0}^{D-1} x_d(t)\,e^{-j2\pi dk/D},$0 for DCLM / UltraFineWeb, and $X_k(t)=\sum_{d=0}^{D-1} x_d(t)\,e^{-j2\pi dk/D},$1 for GREATS. In continued pre-training of Qwen3-8B-Base on SciencePedia, OPUS achieved superior performance using only $X_k(t)=\sum_{d=0}^{D-1} x_d(t)\,e^{-j2\pi dk/D},$2 B tokens compared to full training with $X_k(t)=\sum_{d=0}^{D-1} x_d(t)\,e^{-j2\pi dk/D},$3 B tokens (Wang et al., 5 Feb 2026).

In machine translation, the OPUS ecosystem is an open data-and-model infrastructure rather than a single model. “Democratizing Neural Machine Translation with OPUS-MT” describes OPUS as comprising the Open Parallel Corpus, OPUS-API, OpusTools, OpusFilter, the Tatoeba Translation Challenge, Marian-based training pipelines, and end-user deployment paths. The corpus scale given in the paper is 600+ languages, 40 000+ language-pair bitexts, approximately 20 billion sentences, 290 billion tokens, and 12 TB compressed. The paper also reports base and big Transformer baselines, multilingual temperature sampling, back-translation, pivot triangulation, distillation, quantization, and deployment through servers, Hugging Face, the European Language Grid, Bergamot, translateLocally, OPUS-CAT, and CAT-tool plugins (Tiedemann et al., 2022).

A different OPUS appears in astronomy middleware. “OPUS: an interoperable job control system based on VO standards” defines OPUS as the Observatoire de Paris UWS System, a REST service implementing the IVOA Universal Worker System pattern v1.1 and the IVOA Provenance Data Model. Its architecture has three loosely coupled modules: a web server implementing UWS and storing provenance, a work cluster running the analysis or simulation code through SLURM or a local backend, and a single-page web client for job inspection, provenance visualization, job-definition editing, and SCIM-based administration. The paper states that production instances serve CTA/H.E.S.S, MASER, and CompOSE, and that SLURM integration has been validated up to approximately 100 simultaneous tasks with negligible overhead (Servillat et al., 2021).

These uses show that OPUS often names a control or coordination layer: over gradient-based data choice, multilingual corpora and model delivery, or asynchronous astrophysical job execution.

7. Photonic rails and orbital economics

Two further uses push “Opus” into systems and policy domains with no direct relation to the codec or LLM workflows. In “Photonic Rails in ML Datacenters with Opus,” Opus is a control plane for realizing rail-optimized ML fabrics with optical circuit switches instead of high-radix electrical packet switches. The core idea is “parallelism-driven rail reconfiguration”: non-overlapping collective phases for different parallelism dimensions are used to time-multiplex a single set of NIC ports across multiple optical circuit configurations. The iteration-time model is

$X_k(t)=\sum_{d=0}^{D-1} x_d(t)\,e^{-j2\pi dk/D},$4

where $X_k(t)=\sum_{d=0}^{D-1} x_d(t)\,e^{-j2\pi dk/D},$5 is OCS reconfiguration latency (Ding et al., 13 Feb 2026).

The implementation includes an Opus Shim intercepting collectives through PyTorch’s ProcessGroup API, an Opus Controller synchronizing communication groups and topology changes, and one Opus Network Orchestrator per rail. Evaluation was performed on a Polatis Series 6000n OCS testbed, on NERSC Perlmutter, and in simulation up to 2,048 GPUs. The reported outcomes are over $X_k(t)=\sum_{d=0}^{D-1} x_d(t)\,e^{-j2\pi dk/D},$6 network power reduction, $X_k(t)=\sum_{d=0}^{D-1} x_d(t)\,e^{-j2\pi dk/D},$7 cost savings, and less than $X_k(t)=\sum_{d=0}^{D-1} x_d(t)\,e^{-j2\pi dk/D},$8 training overhead at production-relevant OCS reconfiguration latencies (Ding et al., 13 Feb 2026).

In orbital policy analysis, OPUS expands to Orbital Debris Propagators Unified with Economic Systems. The model couples an astrodynamic propagator such as MOCAT-4S or a GMPHD filter with an economic behavior module for satellite operators. At annual timesteps, it propagates satellite and debris stocks, computes collision probabilities, solves the open-access launch condition for fringe operators, and repeats. The fringe zero-profit condition is

$X_k(t)=\sum_{d=0}^{D-1} x_d(t)\,e^{-j2\pi dk/D},$9

with revenue, discount rate, active lifetime, collision probability, and orbital-use fee as determinants (Rao et al., 2023).

The benchmark scenarios use July 2022 initial conditions, a 35-year horizon in the default setting, two exogenous constellations at approximately 550 km and 1,100 km, a $2.5\,\mathrm{ms}$00year active lifetime, full compliance under some disposal regimes, and economic parameters such as $2.5\,\mathrm{ms}$01\$2.5\,\mathrm{ms}$02\alpha_2^q = 100\,$2.5\,\mathrm{ms}$03/sat/yr, and $2.5\,\mathrm{ms}$04\$2.5\,\mathrm{ms}$050.5\times P_{2k}$ with a 25-year disposal rule (Rao et al., 2023).

Across these cases, “Opus” denotes an orchestration mechanism over physical resources or policy dynamics: optical topology reconfiguration in one instance, and the coupling of orbital physics with incentive-driven economic response in the other. The shared name does not imply a common technical substrate; it marks a recurring preference for framing complex coordination problems as explicit systems with measurable objectives, constraints, and update rules.