The Economics of Unnecessary Work

1. The wrong unit of optimisation

Requests per second and isolated microbenchmarks are useful for diagnosis, but they do not define the value of a system. An organisation ultimately pays for ingest, storage, network transfers, compute, engineering time and failures. A technically elegant filter can lose economically if it requires more operational intervention than the cost it removes.

A meaningful baseline starts with a real workload and a cost boundary: which records or operations generate billable units, which are required by retention policy, and which can safely be transformed, delayed or rejected. The denominator must be explicit. A percentage reduction in synthetic input volume is not a percentage reduction in a customer's invoice.

2. Classify before you eliminate

An upstream admission decision should be based on policy and evidence value, not the appealing idea that 'noise' is disposable. Some rare events are the only usable incident evidence. Eligible reduction therefore needs a negative control: what the system must retain even under overload.

The most useful evaluation compares unchanged baseline, policy-aware candidate and adversarial exceptions. Record false rejection, false admission, deferred processing, residual overhead and required evidence coverage. Only then evaluate throughput, tails and sustained load.

3. Follow the data path

Work often survives a filter and reappears as copies, serialization, memory allocation and retries. This is why admission, memory pressure and copy accounting are related but not interchangeable questions. T.A.R.T.A.R.U.S. explores repeat-work avoidance; C.E.R.B.E.R.U.S. probes memory cost; C.Y.C.L.O.P.S. exposes serialization and data movement; LOGVERITY frames the accounting and retention decision.

These public demonstrators are bounded interactive models. They are not customer invoices, verified deployment savings or a universal speed-up. Their purpose is to make the variables visible before a funded workload-specific trial.

4. An acceptance test a buyer can actually use

Freeze the baseline and candidate versions. Agree on eligible and protected work. Capture representative traffic or approved synthetic equivalents. Measure cost-bearing units with controls for correctness, sampling bias and operational overhead. Include tail behaviour and incidents, not only medians.

The decision is commercial as well as technical: net verified savings minus integration, maintenance and policy risk. A negative result is useful; it can prevent an uneconomic rollout. Licensing should follow the measured value and the customer's actual operational scope.

5. Evidence boundary

The public record gives the problem statement, test design and linked demonstrators. Replication-sensitive implementation, client data, exact acceptance configuration and unpublished evidence remain in a controlled diligence process.