Pathogen-Agnostic Surveillance

Established Pathogen-agnostic surveillance is the attempt to detect an outbreak without being told in advance what causes it. Almost every routine system in use does the opposite. A notifiable-disease list, a respiratory PCR panel, a case definition and a sentinel-site protocol are all instructions to look for something that already has a name, and they are extremely good at that. The agnostic question is the residual: what is happening that nobody wrote a test for.

Established Four instrument families carry the agnostic load, in very different condition. Wastewater sampling is operational at national scale in several countries and is the cheapest population-level sensor ever deployed for infectious disease. Metagenomic sequencing — sequence everything, ask afterwards what it was — is routine in research and a narrow clinical niche, and is a standing surveillance mode nowhere. Animal-side sampling is fragmented and has just had its flagship programme cancelled. Hospital-side agnostic work — systematic sequencing of severe illness with no diagnosis — is the strongest idea in the field and the least implemented.

Frontier The claim this brief lands is that none of the binding constraints is a detection-sensitivity problem. The instruments detect. What fails is everything around the detection: the denominator that would turn a signal into a rate, the coverage in the places where spillover happens, the absence of any published false-alarm rate, the money that has to exist in the years when nothing is happening, and the legal right to the sample itself, which is an unclosed annex in a treaty negotiation that has slipped past two deadlines.

Scope boundary. Resistance, stewardship and the therapeutics pipeline belong to the sibling Post-Antibiotic Medicine commission and are not restated here. The countermeasure that a detection is supposed to trigger — and the measurement problem that keeps it from being licensed — is Universal Vaccines, which owns the correlate-of-protection argument this brief depends on and does not repeat. The tail case, in which surveillance has already failed and the question is what was stockpiled, is Civilization Resilience Planning. What is owned here is the sensing layer: what it measures, what it cannot, who pays for it on a quiet Tuesday, and who is allowed to hold the sample.

Established A note on sourcing. This brief was commissioned in September 2026 from the Institute’s research base. Reading-list entries without links are cited from the bibliographic record rather than re-fetched, and claims are dated no later than early 2026 unless carried by a linked source.

Established Wastewater surveillance is the one part of this field that scaled. The United States National Wastewater Surveillance System was stood up in September 2020 and grew to well over a thousand reporting sites; the Netherlands has sampled essentially all of its municipal treatment plants since the same autumn; WastewaterSCAN runs a multi-target panel across roughly one to two hundred United States sites. Established The oldest working instance predates all of them: poliovirus environmental surveillance, which is how vaccine-derived poliovirus was found in London sewage between February and June 2022 and in New York State the same year, in both cases identifying transmission clinical surveillance had not seen.

Frontier The lead-time evidence is real, smaller than advertised, and misdescribed. Reviews of the SARS-CoV-2 wastewater literature report leads over clinical indicators clustering between roughly zero and fourteen days, central estimates commonly four to eight, and a substantial minority of studies reporting no lead or a lag. Established The spread is structural rather than technical: lead time is not a property of sewage, it is the latency of the clinical system being compared against. Where testing was scarce and reporting slow, sewage led by a week or more; where testing was dense and same-day, the lead collapsed. A programme buying a fixed one-week warning has bought a number that shrinks as the rest of the system improves.

Established What wastewater cannot do is produce a rate. Converting a sewer concentration into a statement about people requires catchment population, flow, in-sewer decay, the shedding fraction, the duration and magnitude of shedding, and a normaliser — pepper mild mottle virus, crAssphage, flow-weighting — on which no standard exists. The errors multiply. Frontier The honest description of the output is an ordinal trend for one catchment over time: not an incidence estimate, and not comparable between two cities.

Established Metagenomic sequencing is cheap per base and expensive per answer. A hundred gigabases of short reads costs a few hundred dollars on current platforms, and that is the wrong denominator. Agnostic detection is a ratio problem: early in an outbreak a novel agent is a vanishing fraction of the nucleic acid in a sewer or a swab pool, plausibly one read in a billion to one in a hundred billion against a background of human, bacterial, dietary and phage sequence. Frontier Published modelling of the depth required — the most careful of it from the Nucleic Acid Observatory — produces annual cost estimates spanning several orders of magnitude depending on assumed shedding, growth rate and the incidence at which one wants the alarm to fire. That spread is the result, and nobody has narrowed it with data.

Established Clinical metagenomics works, in one narrow proven place. A 204-patient study of metagenomic sequencing of cerebrospinal fluid in idiopathic meningitis and encephalitis identified thirteen infections conventional testing had not, while missing organisms conventional testing caught. Established The wider opportunity is untouched: in the largest prospective American study of pneumonia requiring hospitalisation, a pathogen was detected in only about 38 per cent of some 2,250 adults. Roughly three in five severe pneumonias are already pathogen-unknown, in wealthy hospitals with full diagnostic access — and that residual is discarded rather than sequenced.

Frontier H5N1 in North American dairy cattle is the live case, and it is about detection latency rather than detection capability. Highly pathogenic avian influenza was confirmed in dairy cattle on 25 March 2024 in the Texas panhandle; molecular-clock analyses published later that year placed the single introduction from wild birds into cattle in late 2023, roughly four months earlier. Established By early 2025 the United States Department of Agriculture’s own running tally recorded on the order of nine hundred to a thousand affected herds across sixteen or more states, California carrying the largest share; some seventy human cases were reported during 2024 and the first death was announced in Louisiana in January 2025. Established The instrument that finally worked was agnostic to the individual animal: bulk-tank milk testing, formalised as the National Milk Testing Strategy announced on 6 December 2024, which found infected herds on-farm clinical observation had missed.

Frontier The same episode produced the field’s cleanest demonstration of an uninterpretable agnostic signal. H5 sequence was detected in municipal wastewater in multiple Texas cities during spring 2024. The detection was real and the attribution was not: milk-processing waste, animal-facility discharge and human infection are not separable in a sewer. Frontier A signal that cannot be attributed to a host cannot trigger a proportionate response, and the episode showed both limbs at once — agnostic sensing found something months of veterinary reporting had not, then could not say whose it was.

Established Animal-side virus discovery has been tried at scale and stopped. USAID PREDICT ran from 2009 to 2019 at roughly two hundred million dollars and catalogued on the order of a thousand novel viruses; its successor, announced in October 2021 at about a hundred and twenty-five million, was wound up within roughly two years after biosafety and utility criticism. Frontier The critique that killed it is the substantive one: cataloguing wildlife viruses does not predict which will spill over, and the same money spent on sentinel surveillance of severe illness at the human-animal interface would detect emergence rather than inventory potential. That argument was made in 2018 and has not been refuted in either direction.

Established Representativeness is the field’s structural defect, and it is geographic. Wastewater surveillance requires sewers; roughly half the world’s population is not connected to a managed network, and non-connection correlates closely with the land-use, livestock-density and wildlife-contact conditions under which spillover is most likely. Established Genomic surveillance shows the same gradient: the overwhelming majority of SARS-CoV-2 genomes shared during the pandemic came from a small number of high-income countries, with many low-income countries sequencing well under one per cent of confirmed cases. Frontier A sensing system whose sensitivity is highest where emergence is least likely is not a global system; it is a national one with an export market in reassurance.

Frontier What is the false-alarm rate of agnostic detection? Nobody has published one. Every operational system rests on human adjudication of anomalies, and the number of anomalies adjudicated and discarded per true detection is not reported anywhere the research behind this brief could locate. A method with unmeasured specificity cannot be costed, cannot be compared against a targeted panel, and cannot be defended when it triggers an expensive response that turns out to be nothing.

Frontier Can sequence depth substitute for sampling design? The dominant framing treats detection as a depth problem — sequence harder until the needle appears. The competing framing treats it as a sampling problem: a hundred well-chosen severe-illness specimens from a sentinel hospital may carry more information per dollar than a terabase from a sewer, because the clinical specimen has a host, a date, a syndrome and a denominator. This is genuinely unresolved and is the single question a well-designed programme could settle in three years.

Frontier What does “novel” mean computationally? A large fraction of reads in any environmental metagenome already matches nothing known. Separating genuinely novel pathogen sequence from the permanent dark matter of uncharacterised phage and environmental genomes is an unsolved classification problem, and benchmarking exercises comparing taxonomic classifiers report large tool-to-tool divergence on the same data.

Frontier Is bulk-commodity sampling generalisable beyond milk? Bulk-tank milk worked because milk is pooled, chilled, tracked and already sampled for other reasons. The analogous substrates — abattoir wastewater, poultry-house litter, market drainage, aircraft lavatory waste — have the pooling structure and lack the regulatory infrastructure that made milk testable. Whether the milk result was a principle or a lucky commodity is open.

Established The denominator is the first bottleneck and it is not a sequencing problem. Catchment population, connection rates, industrial influent, transient population and flow variability all enter any attempt to convert a wastewater concentration into a population statement, and none of them is routinely measured at the precision the inference requires. This is a water-utility metering problem wearing epidemiological clothes.

Established Normalisation is unstandardised. Faecal-strength normalisers, flow-weighting and per-capita adjustment are all in use, they do not agree, and inter-laboratory comparison studies routinely find method-driven differences of the same order as the signal being reported. Until a reference method exists, cross-site comparison is not defensible and cross-country comparison is decorative.

Frontier Depth costs scale with the inverse of relative abundance, and relative abundance early in an outbreak is unknown. The published cost estimates for agnostic wastewater sequencing differ by orders of magnitude precisely because the shedding parameters for a pathogen that does not yet exist cannot be assumed. Any procurement built on the low end of that range is buying a number, not a capability.

Established Reagents and flowcells are a supply bottleneck that fails exactly when needed. Every sequencing surge since 2020 has brought allocation of consumables, and a surveillance system drawing from the same pool as the diagnostic response loses that competition when its output is most valuable.

Frontier Metadata, not sequence, is the binding data bottleneck. The dairy episode is the demonstration: sequences were deposited, and the early depositions lacked collection dates and locations precise enough for the epidemiological inference researchers wanted. A sequence without a date, a place and a host is a curiosity, and the instruments that produce metadata are administrative.

Established It depends on a clinical specimen stream it does not control. Residual diagnostic specimens are the highest-information substrate available for agnostic sequencing, and access to them is governed by consent frameworks, laboratory retention policy and institutional review, none of which were designed for population surveillance.

Frontier It depends on veterinary and agricultural cooperation that is voluntary and economically adverse. A confirmed detection on a farm imposes movement restrictions, culling and market loss on the producer who reported it. The dairy episode showed what that produces: testing that becomes systematic only when a federal order makes it so, and even then at the bulk-commodity level rather than the animal level.

Frontier It depends on a countermeasure pathway to make the detection worth anything. Universal Vaccines establishes that the field can raise broad responses in humans and cannot yet say which responses predict protection; until that changes, a faster detection hands its output to a countermeasure pipeline whose own yardstick is unresolved. The sibling Post-Antibiotic Medicine commission owns the equivalent dependency on the therapeutic side.

Frontier The decisive experiment is a pre-registered parallel run, and nobody has funded it. In one defined catchment of a few million people, operate agnostic metagenomic sequencing of wastewater and of residual severe-illness clinical specimens at a declared depth for three years, alongside the existing clinical and laboratory surveillance system, with detection criteria, alarm thresholds and adjudication rules fixed in advance, and report every detection, every false alarm and the cost of each. That trial would produce the three numbers the field argues about without possessing: the true-detection count that agnostic sensing adds over targeted surveillance, the false-alarm rate per unit of depth, and the cost per additional true detection.

Established The natural experiment already running is the dairy panzootic, and it is being under-instrumented. Bulk-tank milk testing, farm-level veterinary reporting, municipal wastewater and human case-finding have been operating over the same herds and the same counties since 2024. Comparing what each saw and when, retrospectively and with the metadata unsealed, would recover most of the parallel-run result at a fraction of the cost. It requires no new hardware and it has not been published.

Frontier A third test is cheap and would settle the substrate question. Sequence the residual — the sixty-odd per cent of hospitalised severe pneumonia with no identified pathogen — agnostically for two seasons at a handful of sentinel hospitals, and count how much resolves to known agents, to known agents with divergent sequence, and to genuinely unknown ones. That answer is assumed rather than measured, and the assumption drives most of the field’s strategic argument.

Frontier The negative result worth designing for is a null. If three years of agnostic sequencing in a well-instrumented catchment adds no true detection ahead of the existing system, that is the most useful finding available, and the field currently has no mechanism to produce it because no funder buys a negative surveillance result.

Established The sampling front end determines the result more than the sequencer does. Grab samples and time-proportional or flow-proportional composite samplers give materially different answers from the same sewer, and autosampler deployment, refrigeration and chain of custody are the unglamorous determinants of whether a site produces comparable data over a year.

Frontier Depletion is the lever that actually moves agnostic economics. Human and ribosomal sequence dominates most clinically useful samples. Host depletion, targeted capture panels for broad viral families, and adaptive sampling on nanopore devices each raise the effective depth per dollar by an order or more — and each reintroduces a prior about what to look for, which is exactly the property agnostic surveillance was supposed to shed. This trade-off is the engineering centre of the field.

Frontier Airport and conveyance sampling is the one deployed form of border-adjacent agnostic sensing. Traveller-based genomic surveillance with pooled nasal sampling and aircraft lavatory wastewater has operated at a handful of major hubs since 2021 and 2022 respectively, and the programmes report having caught variants close to or ahead of their first domestic clinical identification. The reported detections are the programme’s own; an independent evaluation of their marginal value against domestic surveillance has not been published.

Established Countermeasures determine this field’s value. Universal Vaccines owns whether a broad-coverage product can be licensed at all, and its finding — that the binding constraint is an immune correlate orthogonal to the target, not antigen design — sets the ceiling on what an earlier detection buys. The sibling Post-Antibiotic Medicine commission owns resistance surveillance, stewardship and the antibacterial pipeline; resistance genes in wastewater are that commission’s instrument, not this one’s.

Established Catastrophe planning is adjacent on the far side. Civilization Resilience Planning begins where detection has failed and asks what was stockpiled and what can be restarted; its treatment of severe pandemic as one of a small scenario set is the correct frame for the residual risk this brief cannot reduce.

Frontier Engineered-organism risk runs the other way through the same instruments. Biosecurity screening regimes govern what may be designed and synthesised; agnostic surveillance is the downstream detector of anything that escapes them, and the only detector that does not require knowing what was built.

Established Microbial community measurement shares the whole toolchain. Microbiome Engineering rests on the same distinction between amplicon sequencing, which generally cannot resolve species, and shotgun metagenomics, which recovers strain and function at much higher cost.

Frontier Vector and reservoir control is adjacent at the animal interface. Designer Organisms covers engineered vector-control technologies whose release would itself require exactly the standing environmental sequencing described here, to detect unintended spread.

Established The right to the sample is an open item in an unfinished treaty, and this is the most consequential institutional fact in the field. The WHO Pandemic Agreement was adopted on 20 May 2025 by a recorded vote of 124 in favour, none against and 11 abstentions, with the United States not participating. Its Pathogen Access and Benefit-Sharing system — which would govern who gets samples and sequences and what flows back — was deferred to an annex, and the Agreement does not open for signature until that annex is adopted. Established The annex has slipped twice. The intergovernmental working group’s seventh meeting, 6 to 17 July 2026, closed with contractual arrangements, laboratory networks and the definition of benefits arising from pathogen sharing all still open, and an eighth set for 14 to 18 September 2026; the outcome is now expected at the Assembly in May 2027 or an earlier special session. The ratification count is necessarily zero until the annex closes.

Established The substantive divide is not procedural. Specialist reporting through early 2026 describes the split as developing countries seeking mandatory benefit-sharing with technology transfer against high-income countries seeking voluntary commitments with innovation protection. The Agreement’s own implementation literature records the bargain: an allocation of pandemic product output to the Organization, One Health surveillance duties in the early articles, and hortatory technology-transfer language beside binding language in the same instrument. Frontier For a surveillance brief the implication is direct and uncomfortable: the input to every agnostic system is a sample somebody has to hand over, and the terms of that handover are unsettled.

Established The International Health Regulations were amended in parallel and did enter into force. The 2024 amendments, which add a pandemic emergency category and a National IHR Authority requirement, entered into force in 2025 for states that did not reject them. This is the binding instrument that actually governs notification, and it is weaker on sample sharing than the Agreement would be.

Frontier Nobody funds standing surveillance between outbreaks, and the arithmetic is public. The G20-commissioned estimate of additional international financing required for pandemic preparedness was of the order of ten billion dollars a year. The instrument built to supply it, the World Bank-hosted Pandemic Fund launched in November 2022, awarded roughly three hundred and forty million dollars in its first call to thirty-seven countries and a comparable order in its second — a mechanism operating at a few per cent of its own diagnosis of need. Established National systems show the same pattern: much of the American wastewater network was built on emergency supplemental appropriations, and site counts fell as those lapsed.

Frontier Preparedness indices do not predict outcomes. The strongest cross-national test available found higher pre-pandemic Global Health Security Index scores associated with worse early COVID-19 outcomes across 36 OECD countries. One index, one disease, an early snapshot: the safest reading is that the index is a weak instrument, not that preparedness harms. But a field that cannot tell a good surveillance system from a bad one by any published measure has unanchored procurement.

Frontier Declaration is discretionary and demonstrably slow. Analysis of ten emergency events between 2009 and 2023 finds the criteria vague enough to leave determination to members’ own interpretations; the 2018 to 2020 Ebola outbreak in the Democratic Republic of the Congo produced three refusals to recommend a declaration, in October 2018, April 2019 and June 2019 — the last despite a confirmed cross-border case — before one came in July 2019. Detection speed cannot compensate for a process that takes nine months to convert evidence into authority.

Established Wastewater surveillance is population surveillance without individual consent, and its defence is aggregation. At treatment-plant scale the ethical position is strong: no individual is identifiable, no one is enrolled, and the output is a community-level trend. The defence weakens monotonically as the catchment shrinks. Building-level and institutional sampling — dormitories, prisons, single workplaces, aircraft — can identify small enough groups that the result functions as a test of named individuals who never consented to one.

Frontier The agnostic property is itself the ethical hazard. A system that sequences everything in a sample sequences human genetic material, drug metabolites and anything else present. Programmes state that they do not analyse human sequence; the data that would permit it is nonetheless generated, retained and in some cases deposited. There is no widely adopted technical standard for host-read destruction at the point of generation, and that is a solvable engineering problem being handled as a policy assurance.

Established Benefit-sharing is an equity claim with a documented grievance behind it. That samples flow out of low-income countries and products flow back at market prices is not rhetoric; it is why the benefit-sharing annex is the hardest item on the Agreement’s agenda. A surveillance system with no answer to it will get fewer samples, later.

Frontier Detection imposes costs on the detected. A herd, a farm, a neighbourhood or a country that reports first bears trade restriction, culling, stigma and travel loss. Until reporting is indemnified — compensation for culled animals is the only established instance — the incentive at the point of observation runs against the system, and no amount of sensing technology reverses it.

Frontier The civilizational argument for agnostic surveillance is that it is the only defence that does not require a correct prediction. Every other layer — stockpiles, vaccine platforms, therapeutic libraries, border policy — is specified against an expected threat. An agnostic sensing layer is specified against the residual, and the residual is where the events that matter have historically come from.

Established The value is asymmetric and therefore chronically underprovided. Surveillance that works produces a smaller event, and a smaller event produces no benefit attributable to the programme. That is the canonical public-good failure, and it is why spending oscillates with emergencies rather than tracking risk: the funding instruments built since 2020 operate at single-digit percentages of their own estimates of need.

Frontier A working agnostic layer would change what an outbreak is administratively. If unexplained severe illness were routinely sequenced, the unit of action would shift from a declared emergency to a continuously updated anomaly register — a different institution, with different thresholds, and no state has one.

Speculative The strongest long-horizon claim is that sensing becomes a condition of trade rather than of health policy. Bulk-commodity testing of milk, meat and animal movement is already closer to a market-access requirement than to a public-health programme. If that generalises, standing surveillance gets funded for reasons that have nothing to do with pandemics, which is probably the most plausible route to it being funded at all.

Handwave The claim that comprehensive surveillance would have prevented COVID-19 is not supportable. It requires simultaneous assumptions about detection, attribution, declaration and response, each of which the record shows failing independently. The honest version: a working agnostic layer moves the detection date earlier by an unknown amount, and what that buys depends on a response system this brief does not own.

These horizons track the sensing layer specifically — instruments, coverage, and the institutional right to the sample — and not the countermeasures a detection would trigger.

• 10 yr: Frontier Wastewater monitoring is a standing utility function in most high-income countries with a reference method and comparable cross-site data; agnostic sequencing is routine on unexplained severe illness in a minority of sentinel hospitals; the first published false-alarm rates exist; the benefit-sharing annex has closed one way or another and the sample-access regime is known.
• 25 yr: Speculative Depletion chemistry and classification improve enough that agnostic detection costs less per catchment-year than the targeted panel it replaces, and coverage reaches non-sewered settings through commodity, abattoir and conveyance sampling rather than sanitation build-out.
• 50 yr: Speculative A continuously updated anomaly register with agreed thresholds replaces discretionary emergency declaration as the trigger for response in at least one region, making detection administratively actionable rather than politically actionable.
• 100 / 250+ yr: Handwave Global biological sensing with representativeness proportional to spillover risk rather than to income, funded as infrastructure. Nothing in the current financing or treaty record supports this, and it is stated as an aspiration, not a forecast.

• Depends on results mapped elsewhere: the correlate-of-protection failure recorded by Universal Vaccines, which bounds what an earlier detection converts into; the shotgun-versus-amplicon ground held by Microbiome Engineering; and the treaty-instrument analysis in Global Cooperation Models and Existential Risk Governance, from which the Pandemic Agreement record is taken rather than re-derived.
• Requires (not on this map) six things nobody supplies: an appropriation that survives the quiet years between emergencies rather than a supplemental that lapses with the last one; a published false-alarm rate per unit of sequencing depth, without which agnostic sensing cannot be costed or compared; a sample-and-sequence access regime that keeps working while benefit-sharing is contested; sewer and sentinel coverage where spillover occurs rather than where sewerage already exists; a consumables supply not competing with the diagnostic surge; and a funder willing to pay for a three-year parallel run that might return a null.
• Enables a response system triggered by a measured anomaly rather than a declared emergency; bulk-commodity and animal-interface monitoring as a market-access function; and the detection layer that engineered-organism screening regimes assume exists downstream of them.
• Adjacent Designer Organisms on engineered vector control and the monitoring it would require, Civilization Resilience Planning on the tail after detection fails, and Scientific Advisory Institutions on how evidence becomes a declaration.

Handwave “Wastewater gives you a one-to-two week warning.” This is quoted as a property of the method and it is not one. Measured leads run from roughly two weeks to zero to negative, and the variation is explained almost entirely by how fast the clinical comparator was. A jurisdiction with fast testing should buy wastewater surveillance for coverage and cost, not for warning time.

Handwave “Sequencing is cheap now, so agnostic surveillance is cheap.” Cost per base has collapsed; cost per detection has not, because detection scales with the inverse of the target’s relative abundance in a background that is overwhelmingly not the target. The published estimates for agnostic wastewater sequencing differ by orders of magnitude for exactly this reason, and quoting the low end is a choice of assumption, not a finding.

Frontier “Cataloguing wildlife viruses prevents pandemics.” This was the premise of a decade of programme spending and it is contested in the open literature by senior figures in the field, who argue that discovery does not predict spillover and that the money buys more at the human-animal interface. The flagship discovery programmes were ended rather than refuted, which is a political outcome and not an evidential one; the question remains genuinely open and is testable.

Established “A detection is a response.” The 7-1-7 assessments are unambiguous on this: detection and notification targets are met far more often than the seven-day response target. The marginal value of detection speed is capped by response latency, and a programme that improves only the first half of that chain has bought a faster number.

Handwave “We detected H5N1 in wastewater, so wastewater surveillance caught the dairy outbreak.” The detections were real; the attribution was not available. Milk-processing discharge, animal waste and human infection are not separable in a municipal sewer, and the episode is better read as showing that an unattributable agnostic signal cannot by itself justify a proportionate response.

Frontier “Preparedness scores tell you who is ready.” The best available cross-national test found the opposite sign. One index against one disease in an early snapshot is a weak instrument rather than a refutation of preparedness, and the defensible claim is narrower: the field has no validated measure of surveillance system quality, so both the rankings and the complaints about them are unanchored.