The Selection Asymmetry
Why discretionary audit selection cannot price compliance risk, and what cross-ledger reconciliation must replace it with
A taxpayer controls the internal consistency of everything it declares. What it does not control is the consistency of that declaration with its own filings under a different tax type, on a different cycle. Detection lives in that gap. This study asks why selection never looks there.
Declared against corroborated
Each ledger reconciles to itself. The gap between what one return declares and what another implies is the entire compliance position — and it is computed nowhere.
Overview
Audit case selection determines which taxpayers, in which periods, for which tax types, will be examined. In every administration the compliance population exceeds audit capacity, so selection is the mechanism that allocates finite verification against an effectively unbounded population.
It should therefore be the most careful decision the administration makes. In practice it is among the least instrumented — made on turnover size, officer familiarity, or a complaint. The failure is not one of law, audit resourcing, or enforcement will. It is a failure of reconciliation.
The distinguishing feature of that failure is that the quantities to be compared are authored by the same taxpayer for different regulators and held by different parts of the administration. A return reconciles because the taxpayer made it reconcile. Neither figure is wrong. Neither is missing. They are simply never brought together — and so the difference between them, which is the entire compliance position, exists nowhere.
The central claim
The revenue at risk in a self-assessment administration is, in the main, not aggressively planned revenue. It is unexamined revenue — under-corroboration and cross-tax inconsistency absorbed silently into normal operations because nothing systematically sets one ledger against another.
Diagnostic framework
Process reconstruction of the as-is selection cycle, failure-mode decomposition against a problem tree, and a design prescription derived by inverting the root condition.
Key diagnostic insights
- 01 Every quantity the administration tests is one the taxpayer supplied, checked against another the taxpayer supplied. Assertions cannot be tested against themselves, however sophisticated the statistics applied to them.
- 02 CIT, GST and SWT are examined by different units, against different systems, on different cycles. No officer holds the joint view, and the inconsistency is a property of the pair, not of either return.
- 03 Audit outcomes are recorded in the case file and never returned to selection. The administration's most valuable training signal — its own sustained adjustments — is generated and discarded every cycle.
The root condition
A self-assessed return is treated as a documentary submission to be reconciled, rather than as a set of testable propositions about real economic activity on which the administration already holds independent evidence — generated by the taxpayer itself, under a different obligation.
The reference design
The prescription is organised around a single invariant, from which the architecture follows.
The invariant
Every selection decision must be the output of a test in which at least one term was authored by the taxpayer for a different regulator, a different cycle, and a different purpose — or by the population, which the taxpayer does not control. The declared value is the object of the test, never an input to it.
The three-record model
Declared · D
What the taxpayer asserted
The lodged position for the tax type in issue, exactly as filed. It is the object of the test and never an input to it.
Corroborating · R
What outside evidence supports
The same taxpayer's declarations under the other tax types, plus the sector-and-size population distribution. This is the term that carries detection strength.
Difference · Δ
The case
Δ = (Bᴿ − Bᴰ)⁺ + I, retained as a record rather than derived on request — with its sign, magnitude, evidence channel and classification.
Models order the queue; rules decide the finding
Six model classes carry the design, each consuming the joint taxpayer view. None asserts a finding — a model may order what to look at first, but the finding is always a stated rule applied to a stated figure.
The division of labour
Deterministic rules decide what is wrong. Models decide what to look at first. An administration that inverts this — letting a classifier assert the adjustment and using rules to explain it afterwards — has built a system that cannot survive its first serious objection, however well it performs on held-out data.
What should be measured
An implementation should be evaluated against outcomes rather than delivery milestones. Each indicator below is observable, and each degrades visibly when the design is failing.
| Indicator | What it measures | Healthy trend |
|---|---|---|
| Cross-ledger reach | The share of scored positions tested against a second tax type, as against those never corroborated. | Rises toward the whole population; the never-corroborated share should approach zero — under the manual process it was never measured at all. |
| No-adjustment rate | The share of examined files that close without an adjustment. | Falls, then stabilises. A rate that stays high means selection is still examining the compliant. |
| Sustained proportion | The share of raised adjustments that survive objection intact. | High and steady. A falling figure indicates the evidentiary posture is weaker than the queue. |
| Unexamined quantum | Declared position uncorroborated by any independent channel. | Falls as reach rises. Exposure previously unreachable is now identified and sized. |
| Cycle interval | Elapsed time from period close to a ranked, justified selection. | Short and stable. Latency that stops falling while volume rises means the queue outpaces capacity. |
Study reference & lineage
Unit of analysis: The selection and examination of a single taxpayer–tax-type–period position across corporate income tax, value-added or goods-and-services tax, and wage withholding.
Method: Process reconstruction, failure-mode decomposition, reference design derived by root inversion.
Applicability: Self-assessment administrations operating an integrated tax administration system alongside an indirect-tax monitoring system.
This study advances a general argument about administrative design. It does not describe, endorse, or evaluate any commercially available system, nor make any assertion about any identified taxpayer. All quantitative characterisations of method performance derive from controlled evaluation on historical return data.
© 2026 The Selection Asymmetry — concept study prepared by Sujoy Maitra.