Adaptive Autonomous Underwater Sampling for Marine Carbon Dioxide Removal Verification

1. Problem Statement

Buyers have paid for marine carbon removal that cannot be delivered, because nobody can prove it happened.

As of October 2025, offtake agreements had secured 578,000 tonnes of marine carbon dioxide removal, and 0.3% of that contracted volume had been formally issued as credits. Fifty-six developers are active in a sector that has absorbed $209 million of private investment (AlliedOffsets, Marine-CDR: Market Landscape, Momentum, and Challenges, 22 October 2025).

The chemistry is not what is failing. Adding alkalinity to seawater shifts the carbonate equilibrium and draws down atmospheric carbon dioxide, and the largest disclosed transaction prices it: Frontier buyers committed $31.3 million to Planetary Technologies for 115,211 tonnes delivered 2026 to 2030, or $271.68 per tonne. What a registry requires before issuing is a quantified alkalinity field, and producing one means measuring a dispersing plume in a moving ocean.

A 2025 field study led by the Woods Hole Oceanographic Institution measured that difficulty. A rhodamine-traced release south of Martha's Vineyard was followed for 36 hours, during which the patch displaced roughly 14.8 kilometres while the R/V Connecticut sampled at intervals of ten minutes or longer, which the authors translate to one sample every 1,200 metres travelled. Tracer fell from a peak near 720 parts per billion to 4 to 5 ppb above a 0.1 ppb baseline. And for a simulated 20-tonne sodium hydroxide addition, only 10% of the potential carbon removal had occurred by the time the campaign ended (Subhas et al., Biogeosciences, 2025). The signal dilutes toward the detection floor over hours, the patch outruns the sampling grid, and the uptake being verified is still mostly in the future when the ship goes home. The authors state it without hedging: "Dispersion and dilution of water masses, both horizontally and vertically, will be a critical factor, as will the ability to track the intervention through space and time. This is the central challenge of in-water MRV."

The R/V Connecticut, a 27.4-metre coastal-class vessel, bills $10,937 per 24-hour day at its 2025 rate, crew and fuel included (University of Connecticut Department of Marine Sciences, effective July 2024). Section 5 turns that into a cost per verified tonne. Verification, not removal, is what the sector cannot yet afford.

2. State of the Art

Four capability families bear on in-water verification, and each has a ceiling that Section 3 quantifies.

Autonomous carbonate sensing is solved and flying. The National Oceanography Centre's lab-on-chip alkalinity analyser has flown with a matching pH sensor on the Autosub Long-Range AUV in the Celtic Sea at registry-grade accuracy (Schaap et al., ACS Sensors, 2025; Hammermeister et al., Environmental Science & Technology, 2025). The ceiling is not accuracy: the vehicle flew a pre-planned transect, and its authors caution that special care is required "in waters with high spatial biogeochemical variability." A dispersing plume is the definition of that.

Fixed-position autonomous monitoring is proven and stationary. An analyser of the same class ran 40 days beside a live alkalinity discharge (Zabihihesari et al., Communications Engineering, 2026). Its ceiling is geometric: it observes one point, and a registry prices a volume.

Autonomous feature tracking is proven, on other variables. MBARI's long-range AUV tracked the deep chlorophyll maximum inside a North Pacific eddy for four days without surfacing (Zhang et al., IEEE Journal of Oceanic Engineering, 2020), and the same group published front tracking on the same platform in 2022. The ceiling is the variable: chlorophyll and salinity are fast, continuous and cheap, and no control law has been closed around a carbonate observable arriving once per ten minutes.

Adaptive plume tracing exists mostly in simulation, tested on a ground vehicle rather than an AUV in water (Wu et al., Robotics, 2024).

The claim this brief makes is deliberately narrow: no published deployment has used a carbonate-system observable as the onboard state variable in a closed-loop adaptive survey of an intentional alkalinity-addition plume. Three bodies of work sit closest and are named here because they are what a reviewer will raise. Sub-seabed carbon storage leak monitoring (STEMM-CCS at Goldeneye in 2019, QICS in 2012) flew AUVs and ROVs carrying real-time pH and pCO2 through a controlled sub-seabed CO2 release, resolving pH of 7.965 and pCO2 of 942.1 µatm within 8 m of the bubble stream. Hydrothermal plume tracing has closed autonomous loops on chemical observables for two decades, generally on redox potential, turbidity or methane rather than carbonate. BGC-Argo floats and gliders carrying solid-state pH sensors collect carbonate data on Lagrangian or pre-planned trajectories. The distinction is the same in each case and it is not cosmetic: those are detection and localisation problems answering whether and roughly where, under a monitoring obligation, while verification is a quantification problem answering how much, to a stated uncertainty, for credit issuance. The objective function differs, and so does the trajectory it produces.

3. Foundational Research

Subhas AV, Rheuban JE, Wang ZA, McCorkle DC, Michel APM, et al. (2025). "A tracer study for the development of in-water monitoring, reporting, and verification (MRV) of ship-based ocean alkalinity enhancement." Biogeosciences, 22, 5511–5534. DOI: 10.5194/bg-22-5511-2025. Fifty-six kilograms of rhodamine WT dye in 1,000 L were released over 75 minutes at 0.2 L/s south of Martha's Vineyard (41.5307° N, 70.6456° W) on 1 to 3 September 2023, then tracked for 36 hours from the R/V Connecticut with nine CTD casts, four Lagrangian drifters and three satellite images. Of 214 paired measurements, 168 (78%) were in-patch. Detection thresholds were alkalinity above 10 µmol/kg, pH above 0.01 units and fCO2 above 10 µatm; a dynamic rather than static baseline cut alkalinity variability by 60%. The patch travelled 14.8 km and the simulated 20-tonne addition realised 10% of its potential removal. The authors recommend sampling above 1 Hz over campaigns far longer than a ship charter allows.

Hammermeister EM, Papadimitriou S, Arundell M, Ludgate J, Schaap A, et al. (2025). "New Capability in Autonomous Ocean Carbon Observations Using the Autosub Long-Range AUV Equipped with Novel pH and Total Alkalinity Sensors." Environmental Science & Technology, 59(14), 7129–7144. DOI: 10.1021/acs.est.4c10139. PMID: 40168248. Lab-on-chip pH and alkalinity sensors were integrated onto the 3.6 m, 750 kg Autosub Long-Range and deployed in the Celtic Sea from 19 to 30 March 2022 across depths of 100 to 3,000 m. The vehicle returned 947 pH measurements (one per 7.5 min, precision below 0.001, uncertainty ±0.010) and 423 alkalinity measurements (one per 10 min, accuracy better than 5 µmol/kg), with mean residuals against co-samples of 1 µmol/kg (σ = 4, n = 191) on the shelf transect and 2 µmol/kg (σ = 5, n = 129) on the deep transect. Platform envelope: 10-day endurance, 550 km at 0.6 m/s, 60 W hotel load. Registry-grade carbonate measurement already rides an endurance AUV, which removes sensing from the risk register and leaves tasking standing alone.

Zabihihesari A, Burt W, Sonnichsen C, Motahari S, Whitworth A, et al. (2026). "High frequency in situ total alkalinity measurement for monitoring ocean alkalinity enhancement field trials." Communications Engineering, 5. DOI: 10.1038/s44172-026-00665-w. PMID: 42020707. The first field deployment of an autonomous lab-on-chip alkalinity analyser during a live trial using magnesium hydroxide slurry. Positioned 60 m from the discharge alongside pH, salinity and temperature sensors, it performed 314 alkalinity and 52 onboard certified-reference-material measurements over 40 days from roughly 3,300 optical readings. Alkalinity rose about 40 µmol/kg above a baseline alkalinity-salinity relationship after roughly 210 tonnes of addition and did not return to baseline between dosing intervals, indicating cumulative retention. The geometry forces the commercially decisive result: a 40-day record from one fixed point cannot say how much alkalinity is where across the affected volume.

Zhang Y, Kieft B, Hobson BW, Ryan JP, Barone B, Preston CM, et al. (2020). "Autonomous Tracking and Sampling of the Deep Chlorophyll Maximum Layer in an Open-Ocean Eddy by a Long-Range Autonomous Underwater Vehicle." IEEE Journal of Oceanic Engineering, 45(4), 1308–1321. DOI: 10.1109/JOE.2019.2920217. In March and April 2018 a long-range AUV carrying a third-generation Environmental Sample Processor tracked and sampled the deep chlorophyll maximum inside a cyclonic eddy in the North Pacific Subtropical Gyre for four days without surfacing, holding vertical position by locking onto the isotherm corresponding to the chlorophyll peak rather than onto chlorophyll itself, and flying tight circles while drifting with the eddy to produce a quasi-Lagrangian time series. Two things transfer: four-day autonomous feature following is existing capability, and the isotherm lock shows a fast proxy can stabilise a control loop whose target variable is too slow to steer on.

Zhang Y, Yoder N, Kieft B, Kukulya A, Hobson BW, Ryan S, Gawarkiewicz GG (2022). "Autonomous Tracking of Salinity-Intrusion Fronts by a Long-Range Autonomous Underwater Vehicle." IEEE Journal of Oceanic Engineering, 47(4), 950–958. DOI: 10.1109/JOE.2022.3146584. The same platform tracked a laterally propagating salinity front. An alkalinity patch advects and dilutes laterally, so this is the closer geometric analogue, and it establishes that the tracking primitive generalises across feature topologies on production hardware.

Schaap A, Papadimitriou S, Mawji E, Walk J, Hammermeister E, Mowlem M, Loucaides S (2025). "Autonomous Sensor for In Situ Measurements of Total Alkalinity in the Ocean." ACS Sensors, 10, 795–803. DOI: 10.1021/acssensors.4c02349. The instrument behind both deployments above: a microfluidic analyser performing spectrophotometric titration in situ, characterised against certified reference materials. It matters here for a reason that is not scientific. It is a laboratory-built assembly, and every deployment cited here drew on a handful of hand-built units.

Wu Z, Wang S, Shao X, Liu F, Bao Z (2024). "Adaptive Path Planning for Subsurface Plume Tracing with an Autonomous Underwater Vehicle." Robotics, 13(9), 132. DOI: 10.3390/robotics13090132. A double deep Q-network learns a plume-tracing policy and is compared against the canonical lawnmower pattern, with the learned policy favoured for large-scale exploration. The stated scope is the finding: experiments ran in numerical simulation and on a ground vehicle. Reinforcement-learning plume tracing has not been flown on an AUV in water, which locates the readiness gap precisely.

4. Competitive Landscape

This market has three layers, and conflating them produces a competitor count that will not survive contact with an oceanographer. The sensor layer is crowded: Sunburst Sensors, Sea-Bird (SeaFET), ProOceanus, 4H-Jena, ANB Sensors, and Clearwater Sensors, the National Oceanography Centre lab-on-chip spinout whose product line is the instrument this analysis depends on. The platform layer is crowded and well capitalised: Saildrone, Teledyne Webb (Slocum), Kongsberg (HUGIN), Ocean Infinity, Bedrock Ocean, Sofar Ocean, Open Ocean Robotics. The decision layer is empty. No company sells onboard adaptive autonomy that estimates a carbonate-anomaly field from its own measurements and re-plans the vehicle's trajectory to reduce the uncertainty a registry prices. That is the claim, and it is scoped to that layer wherever it appears.

Open Ocean Robotics (Victoria, British Columbia) is the closest direct competitor. It builds solar-powered uncrewed surface vessels and launched the Gen-2 DataXplorer specifically for marine CDR monitoring, with sales to research institutions. It closed a CA$2.8 million round in October 2024 co-led by Antares Ventures and Spring Impact Capital, roughly $7.22 million raised in total, and partners on the Ocean Frontier Institute's SCALE MRV project at Dalhousie. Two limits bound it: an alkalinity plume disperses vertically as well as horizontally, so a surface craft samples one face of a three-dimensional object; and the product is a data-collection platform, executing missions rather than deciding them.

atdepth MRV (Cambridge, Massachusetts), an MIT spin-off, sells verification services to operators including Deep Sky, built on GPU-accelerated ocean and biogeochemical modelling, funded through ARPA-E SEA-CO2 at $2,524,964 and the BlueSwell accelerator, and is the other SCALE MRV partner. Its limitation mirrors Open Ocean Robotics': atdepth owns no vehicle, so its accuracy is bounded by a sampling design someone else executed. A faster model does not fix a plume sampled once every 1,200 metres.

Saildrone (Alameda, California) is the platform incumbent and is a genuine carbonate-observation operator, not merely an adjacent one. Its vehicles carried NOAA PMEL's ASVCO2 system through a 22,000 km, 196-day autonomous circumnavigation of Antarctica producing direct air-sea CO2 flux measurements (Sutton et al., Geophysical Research Letters, 2021). It has raised more than $345 million including a $60 million round in early 2025 led by Denmark's EIFO and $50 million from Lockheed Martin in October 2025, with commercial emphasis since moving toward defence and maritime domain awareness. The limit is neither capability nor capital: its vehicles are surface platforms executing operator-specified missions, and it does not sell a policy that re-tasks the vehicle from the carbonate signal.

[C]Worthy, a non-profit building an open modelling and data-integration framework (ARPA-E-funded at $3,884,825), is the most consequential entry here and threatens the thesis rather than competing with it. If registries accept substantially model-based quantification, willingness to pay for in-water verification is structurally weaker than contracted tonnage suggests. The counter is in the protocol text, which requires quantification of model skill through data-model comparison: a model's credit-bearing output is bounded by the observations available to constrain it, and constraining observations are exactly what a dispersing plume denies to a fixed sampling plan. This dependency is carried as a named risk in Section 8 rather than assumed away.

Operators running their own verification are buyers and competitors at once. Planetary Technologies, Ebb Carbon, Vesta, Captura and Equatic each carry in-house MRV capability and could internalise this layer; Running Tide's 2024 wind-down is the cautionary case in the same set. MRV Systems manufactures profiling floats and supplies hardware to several parties above.

The space has not commoditised for three reasons, one technical. The buyer did not exist until recently: Isometric published the first dedicated ocean alkalinity enhancement protocol in 2024, so a company selling registry-grade verification before then had no standard to sell against. The capability spans marine autonomy, carbonate chemistry, microfluidic manufacture and biogeochemical modelling, and institutions holding one rarely hold two. And capital inside marine CDR flows to removal rather than measurement, because removal is what the tonne is sold as. Commoditisation becomes plausible 24 to 36 months after registry protocols stabilise and operators reach continuous commercial dosing.

5. Total Addressable Market

The market is persistent autonomous in-water monitoring services for marine CDR verification: producing the time-resolved three-dimensional alkalinity and carbonate-system field a registry requires before issuing credits. It is not the ocean-sensor market and not the AUV market, both larger and neither what the buyer purchases.

What the registry demands. The Isometric Ocean Alkalinity Enhancement protocol requires characterisation across three regimes: the mixing zone over seconds to hours, the coastal domain over hours to years and potentially several hundred square kilometres, and the open ocean at basin scale over months to years. It requires a time-variable three-dimensional alkalinity forcing function and sets a 5% materiality threshold for the totality of omissions and errors. That specification defines the workload.

Bottom-up, from a measured campaign. At $10,937 per day, the Subhas study's 36 hours of monitoring is roughly $16,400 of vessel day-rate, excluding personnel, sample analysis and instrumentation, so true campaign cost is higher. The simulated intervention was 20 tonnes of sodium hydroxide: at 40 g/mol that is 5.0 × 10⁵ moles of alkalinity, which at a conversion of 0.8 moles CO2 per mole of added alkalinity represents an uptake potential of about 17.6 tonnes of CO2. Potential is not removal. The realised fraction is governed by plume dilution, residence time and air-sea gas exchange rate, and the study measured that only 10% of it had occurred at 36 hours. Roughly 1.76 tonnes of uptake were therefore observed for $16,400 of ship time: about $9,300 per verified tonne against an implied credit price of $271.68, the unit price of the single largest disclosed transaction rather than a market mean.

Two qualifiers travel with that ratio wherever it is quoted. It is not a steady-state cost, because research campaigns are deliberately over-instrumented and buy process understanding rather than credits. And the gap between uptake potential and realised removal is itself the argument, because dilution, residence time and gas-exchange rate are none of them observable from a ship snapshot and all of them require sustained spatially resolved tracking of an advecting patch.

Why the cost does not collapse at scale. The obvious objection is that 1.5 days of ship time spread across a 10,000-tonne deployment costs under $2 per tonne, dissolving the bottleneck. It does not, for three reasons. Dilution drives the anomaly toward the detection floor as the treated domain grows, so a larger deployment is a fainter signal rather than a stronger one. Sampling requirements scale with advected area and residence time rather than with tonnage: a patch from a large dose still translates roughly 10 km per day and still must be followed. And registry protocols specify spatial and temporal coverage across defined regimes rather than a fixed sample count, so the observational obligation grows with the domain. Verification effort tracks the geometry of the water mass, not the mass of reagent added to it.

Serviceable market today, and it is small. A 20% MRV share of the 578,000 tonnes contracted at $271.68 per tonne gives roughly $31 million of embedded verification spend across deliveries to 2030. This is below the $100 million threshold and flagged as such: at today's issued volumes this is a development-stage market, and an investment case ignoring that is mispricing the timing. Federal research money is larger and nearer, at $36 million from ARPA-E and $24.3 million from NOAA, so the near-term revenue mix is grant and contract research rather than per-tonne fees. Spend then scales linearly: every additional 1 million tonnes per year adds roughly $54 million per year of verification demand.

Top-down cross-check. MarketsandMarkets sizes the autonomous underwater vehicle market at $3.13 billion in 2025 growing to $4.64 billion by 2030, an 8.2% CAGR, with unit volume rising from 997 to 1,424 vehicles (Autonomous Underwater Vehicle (AUV) Market, November 2025). Verification service revenue is a fraction of a fraction of that, so the check is a sanity bound: this business is constrained by demand for verification, not by platform supply.

Payment pathway. No CPT or HCPCS analogue applies, because nothing here is a reimbursed clinical service, and any brief implying otherwise has misread the buyer. The equivalent rail is the carbon registry protocol: Isometric issued the first verified ocean alkalinity enhancement credits under its 2024 protocol, to Planetary Technologies and CREW Carbon. Verification is a project expense recovered out of credit revenue, which is why $271.68 per tonne and the 5% materiality threshold together set the cost envelope. Two secondary rails carry money today: federal research funding, and operators' own venture capital, of which $209 million has entered the sector.

6. Research Gap and Commercial Opportunity

Every deployed in-water verification workflow is open loop: ships run pre-planned transects, moorings sit at fixed points, AUVs fly pre-programmed lines, and the plume moves while the sampling plan stays as written before the vessel left the dock. Sections 1 and 3 quantify the consequences. The patch outruns the grid, the measurement misses most of the uptake, and the one platform with the endurance to stay was aimed blind.

The opening is the closed loop, and its shape is dictated by an awkward measurement property. Alkalinity arrives once per ten minutes; pH, salinity and temperature arrive continuously, so a vehicle steering on alkalinity steers on a signal that updates after it has already travelled several hundred metres. The resolving architecture is the one MBARI validated on a different variable: lock the fast control loop onto a proxy and use the slow measurement to correct the proxy relationship rather than to fly the vehicle. MBARI locked chlorophyll tracking onto an isotherm; an alkalinity plume locks onto the pH-salinity anomaly, with each alkalinity sample updating the estimated relationship between that anomaly and true alkalinity. Around that inner loop sits a planner treating the plume as a partially observed field and choosing each next leg to reduce the uncertainty the registry's materiality threshold prices. This is informative path planning against a regulatory objective rather than generic information gain, and that framing is the defensible part: what is optimised is credit issuance, not scientific coverage.

The business consequence reframes what is sold. Today an operator buys ship days and receives a partial snapshot. The closed-loop equivalent sells verified tonnes: persistent, self-tasking, subsurface monitoring priced against the credit revenue it unlocks, on a platform costing a fraction of a vessel day.

Why have the well-funded parties not closed it? The evidence is in ARPA-E's own portfolio. Of the eleven SEA-CO2 projects itemised in Section 7, every one is a sensor or a model, and not one funded the decision layer that decides where the sensor should be. That reflects how the field is organised. Sensor groups are analytical chemists who do not build autonomy stacks. Modelling groups are physical oceanographers whose deliverable consumes data rather than commissioning it. Marine autonomy groups such as MBARI hold the tracking capability but are research institutions with no commercialisation mandate and no reason to specialise a general platform for one buyer. And CDR operators are capital-constrained companies whose engineering attention goes to the dosing system producing the tonne, not the instrument proving it.

The least visible constraint is manufacturing. Every alkalinity measurement cited here came from a hand-built microfluidic analyser produced in small numbers by two laboratories, and a verification service needs tens of units at consistent tolerance with a documented calibration trail, because an instrument without a manufacturing record is one an auditing registry can decline.

7. Comparable Funded Projects

Agencies have committed roughly $60 million to adjacent work in three years, validating both the problem and a willingness to fund technology rather than services alone.

ARPA-E SEA-CO2 (Sensing Exports of Anthropogenic Carbon through Ocean Observation). Department of Energy, $36 million across eleven projects in eight states, announced 26 October 2023. The five largest: University of Colorado at $5,904,233 for broadband-laser optical carbon sensors; Woods Hole Oceanographic Institution at $4,802,245 for thorium-decay carbon-flux sensors and $3,738,960 for an integrated system-on-chip sensor; GE Research at $4,274,658 for multi-kilometre fibre-optic chemical sensing; and [C]Worthy at $3,884,825 for a community modelling framework. The remaining six, from $2,004,554 (University of Utah) to $2,524,964 (atdepth MRV), are likewise sensors or models. The portfolio's shape is as informative as its size: it brackets the gap on both sides without funding the middle.

NOAA Ocean Acidification Program, FY23 NOPP marine CDR awards. $24.3 million across 17 projects involving 47 institutions, announced 7 September 2023, of which $14.36 million came from the Inflation Reduction Act under NOAA's Integrated Ocean Observing System priorities. Named projects include Dennis McGillicuddy's multiscale observing-system simulation experiments, co-funded by NOAA Global Ocean Monitoring and Observing, the Ocean Acidification Program and NSF, and Galen McKinley's work on data requirements for quantifying natural variability in marine CDR models. Andreas Andersson at Scripps received $1.45 million and Andrew Dickson at Scripps just under $1 million. That an observing-system design project and a data-requirements project both won establishes NOAA already treats sampling design, not only sensing, as fundable.

EPA MPRSA research permit, LOC-NESS Wilkinson Basin Study. Woods Hole Oceanographic Institution, permit issued April 2025 under the Marine Protection, Research, and Sanctuaries Act, docket EPA-HQ-OW-2024-0189, field trial conducted August 2025 in the Gulf of Maine. Not a funding award but the more consequential precedent: the first EPA-permitted open-ocean alkalinity enhancement trial in the United States, deploying ships, gliders, long-range AUVs, drifters and satellite imagery over roughly five days. The applicant had previously withdrawn the Martha's Vineyard phase following stakeholder engagement with fishing organisations.

Frontier and Planetary Technologies offtake. $31.3 million for 115,211 tonnes delivered 2026 to 2030, announced August 2025. Private rather than federal, included because it sets the price the verification cost envelope must fit inside.

Federal money funds the science, private money funds the removal, and the verification step connecting them is where contracted tonnes are stalling.

8. Opportunity Assessment

TRL 4, with a subsystem-level evidence chain. On the Section 3 evidence: autonomous carbonate sensing on an endurance AUV sits at TRL 5 to 6; autonomous feature tracking on the same platform class sits at TRL 6; fixed-point registry-grade alkalinity monitoring sits at TRL 6; reinforcement-learning plume tracing sits at TRL 3, simulation and a ground vehicle only. The integrated system exists at no readiness level, which makes it TRL 4 as a whole. TRL 5 requires the planner flown on an instrumented AUV against a traced release in open water, measured against a pre-planned transect over the same event.

Risk 1: the slow measurement cannot close a control loop. Alkalinity arrives once per ten minutes while the vehicle covers roughly 360 metres in that interval at 0.6 m/s, so a planner steering on it steers on stale information. Mitigation is architectural: a fast inner loop on the pH-salinity anomaly with alkalinity updating the proxy relationship. The falsifiable threshold is whether proxy-guided tracking holds the vehicle inside the patch for a defined fraction of mission time relative to a lawnmower transect over the same release; failing that, the fallback trades onboard inference for a multi-vehicle formation buying spatial coverage.

Risk 2: the anomaly falls below detection before the mission ends. Tracer fell to 4 to 5 ppb within 36 hours and alkalinity detection needs roughly 10 µmol/kg against natural variability, so a vehicle can be perfectly guided toward a signal that no longer exists. Mitigation is that the objective is uncertainty reduction against the materiality threshold rather than signal chasing: as the anomaly approaches the floor, the planner correctly redirects to characterising the baseline field, the input that let the WHOI team cut alkalinity variability by 60%.

Risk 3: registries may accept model-based quantification, weakening demand for in-water measurement. The Isometric protocol states that quantification of atmospheric CO2 removal currently relies on biogeochemical ocean models. If protocol evolution leans further that way, willingness to pay for in-water verification falls. Two things bound the risk: the same protocol requires quantification of model skill through data-model comparison, so observations remain the constraint on what a model may claim, and the 5% materiality threshold applies to the total of omissions and errors regardless of method. This remains the demand-side assumption most worth testing before committing capital, and buyer-side due diligence and ratings pressure rather than protocol mandate alone may prove the stronger driver.

Risk 4: reagent-limited instrument endurance bounds any sampling policy. The alkalinity analyser performs a wet-chemical titration, so its sample budget is finite: 423 measurements across an eleven-day AUV deployment, 314 across a 40-day mooring. A policy that samples greedily exhausts the instrument before the campaign ends, which makes sample budget a hard planner constraint rather than an operational detail, and a further argument for spending the expensive measurement on estimator correction rather than on steering.

Risk 5: the dye tracer makes the problem look easier than it is. Both field studies cited here tracked rhodamine, which is bright, cheap and conclusive. Repeated dye release at commercial dosing cadence carries its own permitting burden under the same statute governing the alkalinity release, so a service depending on a tracer inherits a constraint that scales badly. Building against the carbonate signal itself is the only version surviving to commercial scale, which makes this risk the moat: a competitor whose product assumes a tracer has built for the research phase.

Regulatory pathway. This is not an FDA-regulated system and no 510(k), De Novo or PMA pathway applies. The governing United States regime is the Marine Protection, Research, and Sanctuaries Act, administered by the EPA's ocean dumping programme, under which the material released rather than the measurement platform is the permitted object; the WHOI permit in Section 7 is the operative precedent. Internationally the London Protocol governs marine geoengineering placement, and vehicle operations fall under Coast Guard navigation rules for uncrewed systems. Regulatory burden is a moat for the verification layer specifically: the permitting record shows in-water evidence quality is what a review scrutinises and what stakeholder objections attack, so better monitoring compounds into an advantage across public, precedential permit applications.

Locked versus adaptive algorithm. The distinction the FDA formalised for medical AI has a direct analogue and is the most consequential design decision here. A planner that keeps learning at sea produces a verification dataset whose sampling design is not the design an auditor reviewed, and under a 5% materiality threshold an unauditable sampling policy is a defect rather than a feature. The recommended architecture is a locked planning policy in the credit-relevant path, trained ashore, versioned, frozen for the duration of a campaign, and logged so the trajectory can be replayed, with adaptation confined to parameters declared in advance with stated bounds. This mirrors the FDA's Predetermined Change Control Plan framework. Carbon registries have no equivalent today, and adopting PCCP discipline voluntarily is both the fastest route through verification audit and a standards-shaping position while protocols are still being written.

9. Team Requirements

The mismatch between the capability profile this demands and the institutions holding the science is the basis of the opening.

Marine autonomy and sequential decision-making under uncertainty. The two-loop architecture: a proxy-locked inner controller for patch following, and a planner maintaining a spatial estimate of a partially observed carbonate field that selects trajectories reducing the uncertainty a registry prices. This includes the estimator fusing fast pH and salinity against slow alkalinity, the simulation environment for pre-training against advection-diffusion realisations, and the evaluation methodology measuring closed-loop performance against a pre-planned transect over the same release. The benchmark design is itself a deliverable, because no standard exists for scoring a verification sampling policy.

Analytical carbonate chemistry and ocean physical sciences. Carbonate-system state estimation and propagation of measurement uncertainty into a credit calculation; plume dispersion physics and the advection prior the planner is warm-started from; experimental design for traced-release trials; and the baseline characterisation strategy determining whether an anomaly is detectable at all.

Manufacturing engineering. Required from the first month, not the last. A verification service requires tens of lab-on-chip analysers at consistent tolerance carrying a documented calibration trail, because credits are audited and an instrument without a manufacturing record is one a registry can decline. Design-for-manufacturability applied to microfluidic production, tolerance analysis on optical path lengths and reagent handling, and a quality system surviving audit separate a demonstration from a deployable service, and their absence is the most common reason funded work in this class stalls between TRL 4 and TRL 7.

This is deliberately a computational and manufacturing profile rather than a full-stack oceanographic-services profile. An entrant competing head-on with a $36 million federal sensor portfolio on analytical chemistry would lose. An entrant supplying the decision layer that portfolio did not fund is complementary to it, and positioned to partner, license or subcontract rather than displace.