Cartridges: Modular Function Units

• Cartridges are modular, replaceable units with standardized interfaces that enable calibration, exchange, and reuse in larger systems.
• They are applied in fields ranging from optical imaging and radio astronomy to microfluidics, computing, and forensic analysis, optimizing performance and reliability.
• Design challenges such as packaging constraints, thermal limits, and bus saturations emphasize that modularity shifts, rather than eliminates, system coupling.

In recent technical literature, the term cartridge denotes a modular, usually insertable and replaceable unit that packages a bounded function inside a larger system. The same word is used for interchangeable filter holders in astronomical imagers, cryogenic heterodyne receiver units in radio astronomy, plug-in field-emission cathodes, self-powered microfluidic consumables, hot-swappable edge-AI modules, learned KV-cache prefixes for long-context LLMs, and ammunition units or cartridge cases in firearm acoustics and forensic studies (Kim et al., 2016, Baryshev et al., 2015, Grimes et al., 2024, Alvarez-Braña et al., 2024, Brogan et al., 23 Jul 2025, Eyuboglu et al., 6 Jun 2025, Gurny et al., 16 Jun 2026, Dorfman et al., 2022). Taken together, these uses suggest that a cartridge is less a material category than an interface concept: a subsystem is externalized so that it can be fabricated, calibrated, exchanged, or recombined without redesigning the host apparatus.

1. Conceptual scope and systems role

In several of these literatures, the cartridge is explicitly identified as the architectural unit that makes modularity practical. In SQUEAN, the cartridge is described as the “key mechanical unit” that turns the filter wheel into a configurable filter-handling system, because “each filter is mounted on its own cartridge” and the carousel is “designed to install twenty normal or ten large interchangeable cartridges” (Kim et al., 2016). In ALMA Band 9, a cartridge is a complete receiver unit with a “uniform mechanical, thermal, and electronic interface” to the telescope front end (Baryshev et al., 2015). In CHAMP, a cartridge is the “fundamental unit of modular capability,” simultaneously a hardware accelerator module and a software-advertised capability that can be inserted into a live pipeline (Brogan et al., 23 Jul 2025). In long-context LLM work, a cartridge is a learned KV cache loaded as a prefix so that a frozen model can answer questions about a corpus without prefilling the full document text (Eyuboglu et al., 6 Jun 2025).

The host system usually supplies the invariant infrastructure, while the cartridge supplies the specialized function. That pattern is visible across disposable analytical cartridges coupled to reusable electronics, plug-in cathodes inserted into vacuum hardware, cryogenic receiver units inserted into standardized front ends, and per-document KV memories loaded into otherwise unchanged transformer stacks (Jang et al., 2024, Qiu et al., 2017, Baryshev et al., 2015, Talaei et al., 1 Jul 2026). A plausible implication is that cartridge architectures are most valuable when three conditions coincide: the host platform is expensive or durable, the inserted function changes across tasks or bands, and serviceability or throughput matters.

2. Optical and observational-instrument cartridges

The SQUEAN camera provides a clear mechanical example. Its science camera module comprises a focal reducer, a customizable filter wheel, and a CCD camera at the focal plane, while the rotating disk module contains “a carousel, interchangeable cartridges, and filters” (Kim et al., 2016). The carousel is “designed to install twenty normal or ten large interchangeable cartridges.” The normal cartridge supports filters “of less than 10 mm thickness and a 50 mm × 50 mm size,” and the largest cartridge supports filters “up to 86 mm × 86 mm,” allowing “twenty filters of 50 mm × 50 mm size, ten filters of 86 mm × 86 mm size, or many other combinations” (Kim et al., 2016). This design was driven by a constrained packaging problem: the total wheel thickness was limited to 21 mm by the focal reducer flange focal length, yet the system still had to support many filters and permit in situ exchange through “a window to replace filter cartridges and filters” without disassembling the wheel from telescope and camera (Kim et al., 2016).

Calibration in SQUEAN ties cartridge modularity directly to optical performance. The initial filter mask was used to calibrate wheel angle and slot registration, and the measured filter position was repeatable at “much less than one pixel accuracy,” with the shift after wheel movement described as “unmeasurable ($\ll 1$ pixel)” (Kim et al., 2016). Finite element analysis of the cartridge-loaded wheel reported maximum flexure of 32 $\mu$m and optical-axis tilt of 0.′5, which the paper states does not affect optical performance (Kim et al., 2016). One internal detail remains unresolved: the main text states that the maximum filter thickness is limited to 10 mm, whereas the Figure 1 caption says the normal design concept assembles twenty 50 mm × 50 mm filters of “less than 7 mm thickness”; the paper does not reconcile this discrepancy (Kim et al., 2016).

3. Cryogenic, vacuum, and pulsed-power cartridges

In submillimeter astronomy, the cartridge becomes a full receiver instrument. The ALMA Band 9 receiver cartridge is a self-contained heterodyne module for 602–720 GHz, installed as one of the standardized front-end receiver positions in each antenna (Baryshev et al., 2015). Inside the cartridge, a compact mirror arrangement refocuses the telescope beam, a grid separates two orthogonal linear polarizations, beam splitters inject the local oscillator, and two independent double-sideband SIS mixers down-convert the signal to a 4–12 GHz IF chain (Baryshev et al., 2015). Across 73 constructed receivers, the reported mean aperture efficiency was $81.89 \pm 0.90\%$, mean beam squint was $4.34 \pm 2.29\%$, and all receivers complied with ALMA requirements (Baryshev et al., 2015). Here the cartridge is not a consumable holder but the deployable unit of cryogenic optical, RF, mixer, and IF integration.

The wSMA upgrade uses the same word for a different but related systems transition. The original SMA front end had one large cryostat per antenna with eight cartridge positions, though only four single-polarized DSB receivers were active. The new Stage I system replaces this with a two-cartridge cryostat containing two dual-polarized receiver cartridges spanning 194–286 GHz and 264–376 GHz (Grimes et al., 2024). Each cartridge is inserted from the bottom, makes thermal contact through two automatic thermal links, and is aligned by a G-10 “hat” mating to three ball-and-vee-groove kinematic mounts (Grimes et al., 2024). The change is simultaneously cryogenic, optical, and architectural: selector optics move inside the cryostat and are cooled to about 50 K, while the two cartridges become a platform for future sideband-separating, wider-IF, dual-band, and eventually tri-band operation (Grimes et al., 2024).

Vacuum-electronic and pulsed-power systems use cartridge language in a similar modular sense. The “nano-diamond thin film field emitter cartridge” is a plug-in stainless-steel cathode insert, coated with Mo and then with thin-film $(\mathrm{N})$UNCD, inserted into the back wall of a 9.17 GHz single-cell injector (Qiu et al., 2017). The reported performance is $\gtrsim 1\ \mathrm{mA/cm^2}$ at 28 MV/m, with a 4.4 mm diameter cartridge producing a beam that can be as small as 1 mm (Qiu et al., 2017). In pulsed power, the Saturn “flux-compression cartridge” is an 18-cm-long insertable module that compressed flux from about 33 nH into about 3 nH, doubled current into the load to about 6.3 MA, and reduced rise time to about 100 ns, while an unexpected nearly $1~\mu\text{s}$ delay was interpreted as plasma flow switch behavior (Felber et al., 2010). In both cases, “cartridge” identifies a preassembled functional stage inserted into a larger high-field apparatus.

4. Microfluidic, biosensing, and scent-delivery cartridges

Microfluidic work often treats the cartridge as the application-facing consumable, with actuation provided separately. In the self-powered microsystem architecture of polymeric micropumps and plastic microfluidic cartridges, the cartridge is a laminated COC/PSA device whose serpentine channel defines the hydraulic path over which front propagation is measured (Alvarez-Braña et al., 2024). The reported channel section is 0.127 mm$^2$, the channel length is 900 mm or 1800 mm, and one coated pump maintained flow for 23 hours, advancing the fluid front 1.8 meters (Alvarez-Braña et al., 2024). The cartridge is therefore both container and calibrated hydraulic load; the pump is external, modular, and attached at the outlet.

Disposable analytical cartridges extend that logic to biochemical sensing. In the FET-based cholesterol platform, the paper-based analytical cartridge is the biochemical reaction environment, while the MOSFET is reused as an electrical transducer (Jang et al., 2024). The cartridge contains a porous sensing membrane loaded with dried cholesterol esterase, cholesterol oxidase, peroxidase, surfactants, BSA, and buffer; a 20 $\mu$L plasma sample is injected into the inlet; and the cartridge eliminates the need for the transistor to carry its own surface-functionalized biochemistry (Jang et al., 2024). The reported consumable cost is < \$0.15/test, and the deep-learning analysis achieves CV < 6.46% and $r^2 > 0.976$ when blindly compared to a CLIA-certified laboratory (Jang et al., 2024).

The fluorescence vertical flow assay uses a more explicitly cassette-like cartridge. Its top and bottom cases are 3D printed and twist-assembled around a stacked paper microfluidic architecture, culminating in a nitrocellulose sensing membrane with 17 wax-isolated immunoreaction spots (Goncharov et al., 2023). Operation requires three injection/loading steps through a single loading inlet, uses 50 $\mu$0L of serum, and completes in <15 min (Goncharov et al., 2023). Blind testing on 46 individually activated cartridges yielded $\mu$1 and $\mu$2 for myoglobin, CK-MB, and FABP, with mean triplicate CVs of 12.4%, 12.6%, and 12.5% (Goncharov et al., 2023). Here the cartridge is a single-use assay stack whose geometry, spot pattern, and passive fluidics are co-designed with the optical reader and neural inference pipeline.

Olfactory interfaces illustrate a boundary case. AromaGen explicitly criticizes prior systems that rely on “fixed scent cartridges” and “a fixed set of preloaded cartridges or reservoirs,” but its hardware still derives from a 12-channel SCENTAC X-Scent 3.0 device (Wen et al., 2 Apr 2026). The authors state that they “replaced the original cartridges with custom modules filled with our base odorants,” one per channel, and the system realizes mixtures by sequential fan-driven release rather than true simultaneous blending (Wen et al., 2 Apr 2026). The actuation rule is simple: for ratio vector $\mu$3 with $\mu$4, the release duration is $\mu$5 with $\mu$6 seconds (Wen et al., 2 Apr 2026). This clarifies that cartridge-based design can shift from fixed final products to fixed basis elements with runtime compositional control.

5. Computational cartridges: modular hardware and learned KV memories

In edge computing, CHAMP uses cartridges as physical modules on a high-throughput bus. A CHAMP cartridge contains a low-power compute element such as an FPGA, VPU, or ASIC, plus local memory and a bus interface controller; on insertion it advertises its capability and I/O schema to the VDiSK runtime (Brogan et al., 23 Jul 2025). The current prototype uses USB 3.1 Gen1 at 5 Gbps, supports live insertion/removal with staggered power pins, and experimentally shows about 0.5 seconds pause on removal of a middle pipeline stage and about 2 seconds pause on reinsertion while the model reloads (Brogan et al., 23 Jul 2025). End-to-end latency for a simulated three-stage pipeline is approximately the sum of stage latencies plus only about 5% overhead (Brogan et al., 23 Jul 2025). In this setting, the cartridge is a hot-swappable hardware capability block.

In long-context LLM research, the cartridge becomes virtual rather than physical. “Cartridges: Lightweight and general-purpose long context representations via self-study” defines a Cartridge as a trained KV cache $\mu$7 that replaces the full corpus prefix at inference time (Eyuboglu et al., 6 Jun 2025). Instead of storing the KV pairs for all corpus tokens, the model stores only the learned cartridge plus the normal KV pairs for the query. Self-Study generates synthetic conversations about the corpus and trains the cartridge with a context-distillation objective so that $\mu$8 approximates $\mu$9 (Eyuboglu et al., 6 Jun 2025). Reported results are that Self-Study Cartridges match ICL performance while using 38.6x less memory and enabling 26.4x higher throughput, and extend effective context length from 128k to 484k tokens on MTOB (Eyuboglu et al., 6 Jun 2025).

Subsequent work extends and analyzes this paradigm. “Cartridges at Scale” replaces monolithic corpus compression with per-document cartridges, dynamic distractor mixing, and a budget manager that rotates cartridges between GPU and persistent storage; at comparable token budgets it improves over a monolithic cartridge by 10–31 points, oracle cartridge accuracy falls within 2–6 points of full ICL, and retrieval-augmented cartridge selection matches or exceeds conventional RAG while using 3–4x fewer prompt tokens (Hardalov et al., 3 Jun 2026). “Learned Structure in CARTRIDGES” argues that keys act as shareable routers and values carry most of the compression, with foreign-key ablations producing only about 4–5% loss on Llama and about 7% on Qwen3, and introduces Sampled Chunk Initialization as a faster-converging initializer (Diaz, 23 Aug 2025). “Distill to Detect” then reuses the cartridge form as a KV-cache prefix adapter for auditing stealth biases: a small cartridge distilled from the difference between a suspected model and its base can amplify hidden preferences into generated text, with 16-token cartridges giving the strongest detection in the reported inverted-U capacity sweep (Talaei et al., 1 Jul 2026). Taken together, these papers generalize cartridge from a hardware module to a reusable, corpus-specific or behavior-specific memory block.

6. Ammunition cartridges, cartridge cases, and associated analyses

In firearm acoustics, “cartridge” denotes ammunition specification rather than a replaceable module, but the same emphasis on structured metadata and modular variation remains. The C3GD dataset was designed to include a “wide variety of firearms, calibers, cartridges, microphones, and microphone locations,” with metadata files that record “exact cartridges used,” “ammunition type,” and “ammunition specifications” alongside platform and event metadata (Gurny et al., 16 Jun 2026). The dataset comprises more than 8000 field-collected recordings from 28 firearms across 16 calibers (Gurny et al., 16 Jun 2026). The paper argues that caliber alone is often too coarse, citing ambiguities such as 5.56 NATO vs .223 Remington, 7.62 NATO vs .308 Winchester, slug vs shot in 12 gauge, and subsonic, supersonic, and/or overpressure 9x19 mm loadings (Gurny et al., 16 Jun 2026). It also notes that “supersonic ammunition typically produce muzzle blast and ballistic shockwave whereas subsonic projectiles lack the shockwave” (Gurny et al., 16 Jun 2026).

A very different cartridge appears in pharmaceutical manufacturing, where the object of study is the breakage of glass cartridges on a conveyor belt (Boso et al., 2020). An integrated DEM-FEM analysis modeled 840 cartridges/spheres under dense packing and found that the most dangerous condition is impact under confinement rather than isolated impact (Boso et al., 2020). The best tested operating parameters were $81.89 \pm 0.90\%$0, $81.89 \pm 0.90\%$1, and $81.89 \pm 0.90\%$2 (Boso et al., 2020). In the critical scenarios, maximum principal stress reached about $81.89 \pm 0.90\%$3 at the bottom of cartridge 1 and about $81.89 \pm 0.90\%$4 inside the wall at the contact zone, leading the authors to conclude that a crack may originate inside the cartridge and propagate outward (Boso et al., 2020). Here the cartridge is analyzed as a brittle thin-shell primary container whose integrity depends on process flow.

Forensic studies focus on cartridge cases as evidentiary items. The re-analysis of the Ames-USDOE-FBI study reports that, for cartridge cases, interexaminer disagreement with all six AFTE categories retained was 36.4% for matching sets and 59.7% for nonmatching sets (Dorfman et al., 2022). Even after pooling inconclusives, disagreement remained between 23.6% and 45.1% (Dorfman et al., 2022). The paper argues that the original report misused expected agreement as though “observed agreement exceeds expected agreement” were itself a satisfactory criterion, and re-centers the analysis on Cohen’s kappa, $81.89 \pm 0.90\%$5 (Dorfman et al., 2022). This is one of the clearest controversies in the cartridge literature: the object is stable and standardized, but the interpretive framework applied to conclusions drawn from it is disputed.

7. Cross-cutting design themes and controversies

Across these domains, cartridge performance depends less on the isolated artifact than on the quality of its system interface. SQUEAN’s cartridges succeed because the wheel has an access window and an image-based calibration procedure; ALMA Band 9 cartridges succeed because the front end standardizes mechanical, thermal, and electronic interfaces; CHAMP cartridges succeed because VDiSK standardizes discovery, message typing, and routing; and microfluidic cartridges become predictive only when paired with characterized pumps and known channel geometry (Kim et al., 2016, Baryshev et al., 2015, Brogan et al., 23 Jul 2025, Alvarez-Braña et al., 2024). This suggests that a cartridge is best understood as an interface-stabilized subsystem rather than merely a replaceable part.

The same literature also shows recurring bottlenecks. Packaging constraints can dominate the design, as in SQUEAN’s 21 mm wheel thickness and filter-thickness tradeoff (Kim et al., 2016). Bus constraints can dominate modular compute, as CHAMP reports practical saturation around 4–5 devices on USB3 (Brogan et al., 23 Jul 2025). Thermal and alignment constraints can dominate cryogenic cartridges, as wSMA prototype work identified optical-alignment difficulty caused by thermal distortion of the selector wheel at 50 K (Grimes et al., 2024). In virtual memory cartridges, naïvely combining independently trained per-document cartridges can collapse performance to near chance, so compositionality must itself be trained (Hardalov et al., 3 Jun 2026). The shared lesson is that modularity does not remove coupling; it relocates coupling to interfaces, calibration procedures, and resource budgets.

A final clarification is terminological. Cartridge does not always mean a consumer-style, hot-swappable consumable. AromaGen remains physically dependent on cartridge-derived channel modules even while rejecting fixed final-scent cartridges (Wen et al., 2 Apr 2026). LLM papers use “cartridge” for learned KV caches that are loadable but immaterial (Eyuboglu et al., 6 Jun 2025, Talaei et al., 1 Jul 2026). Firearm and forensic papers use the term for ammunition units or cartridge cases whose relevance lies in metadata, acoustics, breakage, or examiner agreement rather than modular insertion (Gurny et al., 16 Jun 2026, Dorfman et al., 2022). Taken together, the literature suggests a broad technical definition: a cartridge is a bounded functional unit—mechanical, cryogenic, fluidic, chemical, electrical, or computational—whose encapsulation enables reuse, exchange, calibration, or controlled analysis within a larger host system.