Software engineering has a long romance with elegant patterns: publish-subscribe, map-reduce, event sourcing. The pattern currently reshaping one of the largest applied-AI markets in the world is humbler than any of them, and it fits in a sentence. If your system corrects errors, it must correct them in both directions.
That is the whole pattern. Bidirectional correction. And the story of how American healthcare learned it, at a cost running into hundreds of millions of dollars, is one of the most instructive case studies in AI system design this decade.
The one-way machine
The setting: US health insurers are paid according to risk scores computed from members’ recorded diagnoses. More documented illness, higher monthly payments. To make sure no illness went undocumented, insurers built retrospective review systems, human teams and later AI pipelines that re-read years of clinical notes hunting for diagnoses that were supported by the record but never submitted for payment.
As engineering, these systems were genuinely impressive: natural language processing over millions of unstructured documents, entity extraction, mapping to payment-relevant code sets. As error correction, they had one architectural property that would eventually attract the Department of Justice. They only corrected in the profitable direction. Missed diagnoses were found and added. Unsupported diagnoses, conditions in the record that did not actually meet documentation standards, were not flagged for removal, because removal reduced revenue.
Federal prosecutors made that asymmetry the centre of their case. In March 2026, a major insurer settled for 117.7 million dollars over chart-review programmes that, in the government’s framing, functioned as one-way ratchets: thousands of additions, effectively zero deletions. Around the same time, federal auditors reported that 81 to 91 percent of sampled high-risk diagnosis codes at three audited plans lacked adequate supporting documentation. The one-way machines had been running for years, and the errors they were structurally blind to had compounded quietly the whole time.
Why asymmetry is a system smell
The lesson generalises far beyond healthcare, because one-way correction is everywhere once you look for it. Recommendation systems that add engagement-boosting content but never prune the manipulative kind. Fraud models tuned to minimise false negatives while false positives quietly ruin legitimate users. Data-quality pipelines that enrich records but never retire stale ones. Any feedback system whose corrections align suspiciously well with the operator’s incentives.
The deep problem is that asymmetric correction does not look like an error. Every individual action the system takes can be locally justified; the additions really were supported, in the healthcare case, often enough. The failure lives in the distribution of actions, visible only when someone asks the question auditors eventually asked: across all your corrections, which direction do they point? A system that answers “always toward my owner’s benefit” has disclosed its true objective function, whatever its documentation says.
For engineers, that suggests a concrete review heuristic. For any pipeline that modifies consequential data, instrument the direction of its modifications and publish the ratio. A healthy correction system produces a defensible mix. A ratio near 100:0 is not a performance metric. It is a confession waiting for a subpoena.
The bidirectional rebuild
The healthcare industry’s response is a live demonstration of retrofitting symmetry into production systems. Modern review platforms now run what practitioners call two-way chart review, in which the same AI pass that surfaces missed diagnoses also flags recorded ones lacking evidence, with both streams routed through human validation and full audit trails. The engineering details are worth studying: every suggested change, addition or deletion, ships with the source evidence, the rule it satisfies or fails, and the reviewer’s recorded decision, so the system’s behaviour can be reconstructed inference by inference years later.
Notice what symmetry cost and bought. It cost short-term revenue; deletions are negative dollars by design. It bought survivability: plans running bidirectional review enter federal audits with error rates they already know, evidence already organised, and a correction history that demonstrates good faith at the distribution level, which is precisely where prosecutors now look. Several vendors report that the deletion capability, the anti-feature nobody wanted to demo five years ago, is now the first thing sophisticated buyers ask to see.
The pattern, portable
Strip the healthcare specifics and the design rule travels to any AI system whose outputs carry consequence. Make correction bidirectional by architecture, not policy; if deletions require a different team or a manual process, the asymmetry will reassert itself under deadline pressure. Instrument the direction of changes as a first-class metric. Treat “the system only ever helps us” not as a success story but as the specific failure mode regulators have learned to recognise. And when the incentives of the operator and the accuracy of the data diverge, assume the eventual audience for your logs is adversarial.
Elegant patterns earn conference talks. This one earns something better: systems that survive the day someone hostile reads their history. Two directions. That is the entire pattern. The industry that ignored it paid nine figures for the reminder.
