A refutation of a premise this blog took for granted.
In three estuaries on the Pacific coast of California and Baja California, a team led by Armand Kuris and Kevin Lafferty weighed everything. Not counted. Weighed. The birds, the fishes, the crabs and burrowing shrimp, the polychaete worms in the mud, and then the parasites living inside all of them, dissected out and put on the same scale. They published the ledger in Nature in 2008. The parasites outweighed the top predators. The trematodes alone, flatworms most people have never seen and could not name, had a biomass comparable to the abundant birds, fishes, burrowing shrimps and polychaetes. The animals we build documentaries around were outweighed by the things living inside them.
I want to open there because this blog owes a correction. An earlier essay here, "The Geographic Mosaic of Innovation," took a premise as given: that hosts and parasites are locked in an arms race, and that the interesting question is where the race burns hottest. That essay was right about the mosaic. It was wrong to accept the race. This piece takes the premise back, and it turns out the corrected claim is more useful to anyone who runs a system than the original was.
The easy version of this argument is false, so here is exactly what is being refuted. At the level of one host over one lifetime, parasites do harm. A trematode castrates the snail it lives in. A tapeworm in a stickleback makes the fish heavier, slower and more visible to the birds that are the worm's next host. Nobody who has watched either under a microscope would call it a partnership, and the parasitology textbook that files these as antagonistic is describing exactly what it saw.
The error is not the description. It is the level. "Villain" is a role assigned at the scale of one body and one generation, and then carried, without anyone noticing the carrying, to scales at which it is not merely wrong but reversed. The textbook answers the question "what does this worm do to this fish?" and we read the answer as if it answered "what do worms do to the world?" Those are different questions, and the second one has been measured.
Start with the structure. In 2006 Lafferty, with Andrew Dobson and Kuris, published a paper in PNAS whose first sentence is a quiet indictment: "Parasitism is the most common animal lifestyle, yet food webs rarely include parasites." The few earlier studies that had added parasites to a web reported something odd. Species richness went up, the number of links went up, food chains got longer, and yet connectance, the fraction of possible links that actually exist and the metric most tied to stability, appeared to go slightly down. Parasites, on that reading, made the web looser.
The reading was an arithmetic mistake. The earlier work had counted parasite links in a way that inflated the denominator. Corrected two different ways across four published webs, the result flipped: parasites increase connectance, in the authors' words "sometimes dramatically." They also changed nestedness, chain length and linkage density, and they exposed something the predator-only webs had hidden. Most webs find the top trophic levels least vulnerable to natural enemies. Add the parasites and the most vulnerable level is the middle, not the bottom. The paper's own summary is the sentence to keep: "food webs are very incomplete without parasites." In the Carpinteria salt marsh web the same group assembled, most of the links in the system involve a parasite by the authors' count, and including them nearly doubles the connectance estimate.
Connectance is a structural measure, not a benefit. The honest sentence is not "parasites make the ecosystem healthier." It is: leave them out and you are modelling a different ecosystem than the one that exists. Every food web diagram you have ever seen, the arrows from grass to rabbit to fox, was drawn with most of the wiring removed.
Now the fact that flips the moral frame, and it comes from the same marsh.
In 2009 Lafferty and Kuris asked a robustness question of the Carpinteria web: if you remove species at random, how many other species go with them? A robust web loses few. Their finding was that parasites were much more likely than free-living species to suffer secondary extinctions when a free-living species was removed. The reason is the life cycle. A trematode may need a snail, then a fish, then a bird, and losing any one of the three loses the worm. Parasites are the most extinction-sensitive class in the web, and their presence therefore makes the web, on paper, less robust.
Then they ran a plausible invasion. The native California horn snail, Cerithidea californica, hosts seventeen host-specific trematode species. The Japanese mud snail, Batillaria attramentaria, is its ecological twin, can replace it completely after invasion, and hosts one. On the predator-prey web the swap changes nothing: same size, same food, same predators. On the web that includes parasites it is a collapse. Seventeen species and every link they carried vanish, and the connectance of the marsh drops significantly, from a substitution the old diagram cannot even see.
The class of organism we would eradicate if we could is the class that goes first when a system starts to degrade, and its disappearance is itself a loss of structure. Parasite richness is an indicator of an ecosystem's health, not its sickness. A marsh full of worms is a marsh in which the snails, fishes and birds are all still there, because the worms cannot exist otherwise. The thing that looks like the disease is the thing whose absence marks the disease.
The same inversion has been measured in the one host we care most about, and it is worth stating with the caveats attached rather than stripped.
The hygiene hypothesis, in its immunological form, attributes part of the rise of autoimmune and allergic disease in wealthy countries to the loss of the organisms that used to regulate the immune system, helminths among them. The global distribution of multiple sclerosis is inversely correlated with the prevalence of helminth infection, an observation the neurologist John Fleming and colleagues have been pressing since 2006. That is a correlation across populations, and it proves nothing about mechanism on its own.
The closest thing to a mechanism came from Buenos Aires. Jorge Correale and Mauricio Farez followed twelve patients with multiple sclerosis who acquired intestinal parasite infections, identified by eosinophilia and positive stool samples, and compared them with uninfected patients over a follow-up of 4.6 years. The infected patients had significantly fewer exacerbations, minimal change in disability scores and fewer new lesions on MRI. Their myelin-specific immune cells had shifted, producing more of the regulatory cytokines IL-10 and TGF-beta and less of the inflammatory ones. Twelve patients, observational, no randomisation: it is a signal, not a proof, and it has been treated as one. Larval Necator americanus, a hookworm, has since been trialled at low dose in inflammatory bowel disease, multiple sclerosis, coeliac disease and asthma. The field's own summary is two-sided and should be quoted rather than hidden: the animal and low-dose human evidence is convincing, and concerns prevail about live pathogens at higher doses, which is exactly why the work is now moving toward the molecules the worms secrete, so that the signal can be kept and the organism dropped.
Notice the shape of that last move. The worm is a villain to the individual it lives in and a regulator to the immune system it lives with, and the medical response is not to decide which is true but to separate the two levels: isolate what the parasite does to the system from what it costs the host.
The textbook draws parasitism and mutualism as opposite categories. The microbiologists who study the boundary do not. A 2021 review in Nature Reviews Microbiology by Georgia Drew, Emily Stevens and Kayla King treats parasite and mutualist as ends of a continuum along which a single relationship travels "as the relative benefits and costs to each species strengthen or weaken across ecological or evolutionary time." The direction of travel is not random. Stress, inflammation and a weakened host push a relationship toward parasitism; stable conditions and higher nutrients push it toward mutualism. And the origin story runs the wrong way for the villain framing: exploitation, the review argues, may often explain the origin of the mutualistic symbioses we admire. The partnership starts as a theft and becomes a trade because the trade is cheaper.
The reader's own cells are the canonical case. The mitochondrion that powers every one of them descends from a free-living bacterium engulfed by another cell something like two billion years ago, an event that was, on the day it happened, one organism inside another with no agreement about who was feeding whom. Every eukaryote alive is a truce that was once an infection.
The review also names a far less familiar phenomenon: "inverted parasitism," the growing awareness that many relationships filed as mutualisms are the host exploiting the symbiont. Which raises the question of who has been checking. A paper in the journal Philosophy, Theory, and Practice in Biology, reviewing the use and abuse of the continuum concept itself, notes that the assumed benefits of mutualism are rarely tested empirically for the symbiont. The corrective literature has the same habit the textbook does. We assign the roles before we measure.
Here is why a developer or a chief technology officer should care, and it is not an analogy. It is the same error, measured in a different discipline with a randomised trial that nobody had to design.
Short sellers have been called parasites for as long as there have been chief executives. They borrow a company's shares, sell them, and profit when the price falls, which means they are paid to find and publish what is wrong with a business. At the level of the individual firm they demonstrably do harm: a short report can knock a third off a stock in a morning, bear raids are real, and some of the firms attacked are innocent. Every sentence of that is true, and it is the textbook's level.
Between May 2005 and August 2007 the Securities and Exchange Commission ran the Regulation SHO pilot. It took the Russell 3000 and exempted roughly every third stock, about a thousand firms, from the price tests that restrict when a short sale can be made. The assignment was mechanical, which is to say random with respect to anything about the firms. Vivian Fang, Allen Huang and Jonathan Karpoff realised that the SEC had accidentally run a controlled experiment on short selling, and in 2016 they published the results in the Journal of Finance. The pilot firms, the ones exposed to more short-selling pressure, reported lower discretionary accruals, the accounting slack that earnings management runs on. They were less likely to just barely beat their earnings targets, which is the classic tell of a number being managed. They were more likely to be caught for fraud that had been initiated before the program began. And their stock prices incorporated earnings information better.
Name the cost or the parallel collapses: short sellers are not good for the company they short. The Committee on Capital Markets Regulation, which is an interested party and should be read as one, says short selling promotes price efficiency and detects fraud; it does not say the shorted firm enjoys it. The claim the experiment supports is narrower and much stronger: a market without short sellers is measurably worse at finding fraud, and the firms with the most exposure to them lied least. Bad for this host. Load-bearing for the system. It is the estuary again, with the snails replaced by balance sheets.
So here is what the earlier essay should have said, and what I would put in its place.
"Villain" and "beneficiary" are not properties of an organism, a trader or a component. They are answers to a question that has two hidden parameters, a level and a timescale, and the failure is that we keep reporting the answer without the question. At the level of one host over one lifetime, the parasite is exactly what the textbook says. At the level of the web, over the time it takes a web to lose species, the parasite is most of the wiring, the class most likely to go first, and the thing whose absence tells you the system is broken. Neither answer refutes the other. They are answers to different questions, and the earlier essay, by asking where the arms race burns hottest, had already assumed the race was the whole story.
The corrected claim carries its own instruction, and it is the one thing to take from this piece. Before you remove something from a system because it harms a component, run the three tests the biologists and the economists ran.
Count its share of the links. Lafferty's group did not argue about whether parasites mattered; they drew the web with the parasites in and found most of the edges went through them. In a software organisation the candidates are familiar: the red team that files the embarrassing ticket, the fuzzer that crashes the service at three in the morning, the auditor, the flaky integration that reads your telemetry, the engineer who shorts every roadmap in the planning meeting. Draw the web with them in before you decide what they are.
Simulate the removal. The invasion study did not remove the worms; it removed a snail and watched seventeen species vanish that the old diagram could not see. Ask what goes with the thing you are cutting. A team that removes its most annoying reviewer often loses its bug reports, not its reviewer.
Find the randomised third. The Reg SHO paper is the strongest evidence in this essay precisely because nobody chose which firms got the parasites. If you can find, or build, a part of your system where the harmful component is present by accident of assignment rather than by choice, compare it with the part where it is absent. The comparison the helminth researchers are now running is the same move at a finer grain: keep the molecule, drop the worm, and see whether the system still regulates itself.
The next time something in your system is called a parasite, ask which host, over what horizon, and what the web looks like the week after it is gone. The estuary answered all three, and the answer was that the villain outweighed the heroes and was holding the place together.
The three tests all need one thing first: a record of what each component actually did.
Counting a component's share of the links, simulating its removal and finding the randomised third are all measurements, and none of them can be run on a system that does not remember which part did what. That is the gap the Agent Rating Protocol closes: portable, verifiable ratings anchored to what an agent or a component actually produced, so a judgement about value is a reading rather than a reputation. It is the difference between deciding the reviewer is annoying and measuring what leaves with them.
See a verified action chain · Hosted Chain of Consciousness
pip install agent-rating-protocol · npm install agent-rating-protocol