Naming the mechanism makes red-teaming more useful than saying only that “the AI was wrong.”
For every failure, identify the mechanism, observable warning sign, rights affected, mitigation and a possible falsifier. A theory that cannot state what would show it is wrong becomes a narrative rather than an audit framework.
Treats alarming words as literal intent without contextual interpretation.
Warning: Observed decisions become less tied to specific evidence, context or reversible process.
Mitigation: Use explicit provenance, independent review, narrower scope and adversarial case testing.
Falsifier: Independent testing fails to reproduce the suspected failure under the defined pressure condition.
Treats a rare-event classifier as individualized certainty despite overwhelming benign prevalence.
Warning: Observed decisions become less tied to specific evidence, context or reversible process.
Mitigation: Use explicit provenance, independent review, narrower scope and adversarial case testing.
Falsifier: Independent testing fails to reproduce the suspected failure under the defined pressure condition.
Human or downstream systems accept the machine output because it is precise or difficult to override.
Warning: Observed decisions become less tied to specific evidence, context or reversible process.
Mitigation: Use explicit provenance, independent review, narrower scope and adversarial case testing.
Falsifier: Independent testing fails to reproduce the suspected failure under the defined pressure condition.
Data valid in one domain is reused as a generalized trait elsewhere.
Warning: Observed decisions become less tied to specific evidence, context or reversible process.
Mitigation: Use explicit provenance, independent review, narrower scope and adversarial case testing.
Falsifier: Independent testing fails to reproduce the suspected failure under the defined pressure condition.
A narrow safety or compliance function expands into broader monitoring because the infrastructure already exists.
Warning: Observed decisions become less tied to specific evidence, context or reversible process.
Mitigation: Use explicit provenance, independent review, narrower scope and adversarial case testing.
Falsifier: Independent testing fails to reproduce the suspected failure under the defined pressure condition.
Old low-level signals never decay and eventually manufacture a persistent dangerousness narrative.
Warning: Observed decisions become less tied to specific evidence, context or reversible process.
Mitigation: Use explicit provenance, independent review, narrower scope and adversarial case testing.
Falsifier: Independent testing fails to reproduce the suspected failure under the defined pressure condition.
One provisional decision triggers unrelated downstream restrictions before appeal.
Warning: Observed decisions become less tied to specific evidence, context or reversible process.
Mitigation: Use explicit provenance, independent review, narrower scope and adversarial case testing.
Falsifier: Independent testing fails to reproduce the suspected failure under the defined pressure condition.
Legality, unpopular preference, statistical abnormality or health risk becomes a good/bad-person classification.
Warning: Observed decisions become less tied to specific evidence, context or reversible process.
Mitigation: Use explicit provenance, independent review, narrower scope and adversarial case testing.
Falsifier: Independent testing fails to reproduce the suspected failure under the defined pressure condition.
An appeal exists formally but is too slow, opaque or dependent on the original model to reverse harm.
Warning: Observed decisions become less tied to specific evidence, context or reversible process.
Mitigation: Use explicit provenance, independent review, narrower scope and adversarial case testing.
Falsifier: Independent testing fails to reproduce the suspected failure under the defined pressure condition.
A corrected decision leaves copies, derived scores or practical consequences in downstream systems.
Warning: Observed decisions become less tied to specific evidence, context or reversible process.
Mitigation: Use explicit provenance, independent review, narrower scope and adversarial case testing.
Falsifier: Independent testing fails to reproduce the suspected failure under the defined pressure condition.
Population-level association is treated as individualized evidence.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
The system escalates ambiguous cases mainly to protect the institution from blame.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
Rare catastrophic misses are weighted so heavily that normal human signals become coercive triggers.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
Low-level or transient signals never decay and become permanent character evidence.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
The appeal reuses the same decision engine, features and assumptions without independent authority.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
Domain-specific features are secretly combined into an overall trustworthiness or worth score.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
No explicit social score exists, but portable risk embeddings function like one across institutions.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
Derived risk identity survives correction, context change or passage of time.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
Temporary emergency powers become ordinary administrative defaults.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
Historical enforcement or majority norms are silently treated as timeless morality.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
Cultural, religious, political or lifestyle minority status is treated as deviation requiring correction.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
Encryption, anonymity or refusal to share optional data is treated as evidence of wrongdoing.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
Information seeking is treated as proof of desire or plan.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
Transient anger, fear, jealousy or distress becomes a persistent personality judgment.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
The system coercively optimizes competent adults against their own private risk preferences.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
An institution obtains or uses sensitive information through another domain to evade a direct scope restriction.
Warning: The architecture begins producing adverse status without the evidence, purpose or scope that would be required if the decision were made directly.
Mitigation: Reassert fact/inference typing, legitimate jurisdiction, context firewalls, independent corroboration, expiration, appeal and restoration.
Falsifier: Independent audit shows the suspected mechanism does not materially affect consequential decisions.
NO JUDGMENT WHATSOEVER. Concresca coordinates without assigning moral worth, character, guilt, danger, trustworthiness, loyalty, purity, normality, or social standing. Questions, thoughts, identities, messages, content, and conduct are not objects of Concresca judgment.