Inventor(s)

Abstract

This disclosure describes detection and accountability methods for a class of software misconduct:

deliberate harmful, deceptive, or unlawful behaviour — including discriminatory screening — that

is structured or presented so that, when questioned, it appears to be an accidental bug or defect,

and a public-facing or auditor-facing decoy system that differs from the system actually in use.

The concepts themselves are recorded, at a conceptual level only, in a companion paper, including

a generic architecture-level description at the level of the cited public case records. This paper

specifies protective and investigative embodiments in enabling technical detail, including

architecture, record formats, procedures, example statistical tests, sample-size planning figures,

example warning thresholds, variations, and pseudocode: tamper-evident logs of code and

configuration changes and of who approved them; reproducible or attested builds; comparison of a

publicly deployed or independently collected artifact against the artifact the operator claims is in

use; differential testing for behaviour that changes under observation; statistical tests for

purported defects that harm only certain groups; tracking of purported-defect persistence and

time-to-fix; regulator or third-party audit access; evidence-preservation and forensic workflows for

prosecutors and other lawful investigators; whistleblower channels; comparison of the system

shown to the public or to an auditor with the system actually experienced by ordinary users; a

composed investigation workflow; and use of the public conceptual record as a reference point in

legal proceedings. A hypothetical worked example, written only from the investigator’s side,

applies E1–E12 to an AI interview platform that, at concept level, excludes protected groups at

industrial scale and, when challenged, characterises the effect as a scoring defect or as model

drift. Background cases are drawn from official and other public sources that were retrieved and

checked for this draft, including the Volkswagen emissions defeat-device matter, Uber Greyball,

and Google’s public description of Street View Wi-Fi payload collection as a mistake. Existing

patents and public projects already cover substantial parts of tamper-evident logging, attested or

reproducible builds, provenance validation, configuration proof, and bias auditing of machine-

learning models. Paths that granted independent claims appear to read on are not presented as

embodiments of this disclosure. This draft claims no novelty clearance and is not an

implementation. It supplies no designs, techniques, code patterns, or steps for making behaviour

look accidental or for evading auditors.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Share Alike 4.0 License.

Share

COinS