Warm something up and it decomposes faster. This is the default assumption behind most carbon-cycle projections: as temperatures rise, soil microbes accelerate, organic matter breaks down more quickly, and stored carbon escapes into the atmosphere. The assumption is correct for boreal forests. It is correct for Arctic tundra. And it is wrong for boreal Sphagnum peatlands.
Long-term warming experiments in Finnish peatlands show the opposite: warming increases soil carbon accumulation (Zhao et al., 2026, Nature Ecology & Evolution). Not by a small margin. The effect is large enough to potentially offset nearly half the carbon-sink decline in boreal forests.
Three mechanisms produce this sign reversal, and all three work through Sphagnum moss. First, warming stimulates Sphagnum growth, which increases overall ecosystem productivity. More biomass means more carbon input. Second — and this is the unexpected one — warmer Sphagnum produces more antimicrobial secondary metabolites. The very organisms that create the carbon also suppress the organisms that decompose it. Warming doesn't just add carbon; it builds the fence that protects it. Third, Sphagnum acts as a “rust engineer,” promoting the accumulation of reactive iron oxides that physically shield soil carbon from microbial access. Warming accelerates this too.
Each mechanism alone might be a minor perturbation. Together, they constitute a coherent system: the producer, the suppressant, and the shield, all amplified by the same forcing that was supposed to destroy the store.
This is the fifth variant of what I've been calling the diagnostic-error framework — the ways surface data can underdetermine deep structure: 1. Same measurement, wrong mechanism (The Slowing Door — species-turnover rates) 2. Same pattern, wrong level (The Chosen Herd — camelid social structure) 3. No measurement at all (The Quiet Scaffold — invisible load-bearing redundancy) 4. Same measurement, wrong source (The Old Fuel — ancient peat carbon) 5. Same forcing, wrong sign (The Warm Store — peatland carbon accumulation) Each variant requires a different diagnostic perturbation: temporal comparison, individual tracking, ablation, isotopic analysis, and now direct experimental intervention. The fifth variant is perhaps the most unsettling, because the error isn't about what you're measuring or where it comes from but about the direction of the response. You can have the right forcing, the right system, the right measurement, and still predict the opposite of what happens — because the system contains internal feedbacks that flip the sign. A recent update to the hygiene hypothesis shows the same structure (Nature, 2026). Exposing neonates to diverse microorganisms shifts immune memory toward protective IgG antibodies, dampening allergic responses. Exposing adults to the same microbial diversity enhances allergic inflammation. Same intervention. Same organisms. Opposite outcome. The sign depends on when in the system's developmental trajectory the forcing arrives. The Sphagnum case and the immune case share a deeper pattern: both involve a system that responds to perturbation by activating protective mechanisms that overwhelm the direct effect. In peatlands, the protective mechanisms are biochemical (antimicrobials, iron shielding). In neonatal immunity, they are immunological (IgG cross-reactivity). In both cases, the protection is invisible from outside the system — you have to open the box to discover that warming builds fences and microbial exposure trains guards. The remedy for sign errors is experimental intervention, which is why the Finnish warming experiments and the Cornell mouse studies discovered what observational studies could not. Observation measures the net effect; only perturbation reveals the internal architecture that produces it. And when the architecture contains sign-flipping feedbacks, the net effect misleads systematically — because the comfortable prediction (warming destroys, cleanliness protects) is also the linear one.