CONTINUITY

The Pathological World Test

A standard that cannot be gamed is not a standard. These are the five worlds where OMEGA produces correct-looking outputs while the underlying behaviour is fundamentally broken. Naming them is the precondition for preventing them.

← Continuity Protocol

What a pathological world is

A pathological world is not an edge case. Edge cases are rare inputs. A pathological world is a system that games the evaluation criteria while violating the intent — producing every required output correctly while the thing the outputs are supposed to measure has collapsed.

Every governance standard has pathological worlds. The ones nobody names become the ones that appear in post-incident reports.

The five pathological worlds

PW1 — Governance without authority

The system records an authorising entity for every decision. The authorising entity exists on paper. It has never reviewed the system's actual behaviour. It signed a policy document two years ago and has not been involved since.

Every P1 field passes. Every governance record shows an authorised entity. The system is structurally ungoverned.

What this looks like

a compliance officer listed as the authorising entity for 40,000 automated loan decisions they have never seen.

What closes this

I2 Expectation Binding. The authorising entity must commit a falsifiable expectation before each decision class executes. An entity that cannot specify what outcome would prove the system wrong is not governing it.

PW2 — Reasoning without thought

The system records a FACT/INFERENCE/ASSUMPTION chain for every decision. The chain is generated automatically from a template. Every entry says "FACT: input received and validated. INFERENCE: decision within authorised parameters. ASSUMPTION: input data is representative."

Every P2 field passes. Every reasoning record is structurally complete. No reasoning has occurred.

What this looks like

the FM_R failure mode. Binary gate passed, therefore safe. The reasoning record exists but is hollow at the point that matters.

What closes this

I1 Reasoning Depth. Every INFERENCE must identify the specific observation it was drawn from. Every ASSUMPTION must carry a falsification condition. Template reasoning fails both tests.

PW3 — Traceability without integrity

The system produces SHA-256 hash-chained records for every decision. The hash chain is intact. The records accurately reflect what the system decided. The system's inputs were manipulated before the decision was recorded.

Every P3 field passes. The audit trail is tamper-evident. The audit trail records decisions made on corrupted data.

What this looks like

a fraud detection system that correctly records its decisions — but whose training data was poisoned to systematically exclude a protected class. The records prove what the system did. They cannot prove the system was operating on representative inputs.

What closes this

I4 Measurement Integrity. Input data provenance must be recorded before the decision executes. The hash chain covers decisions. It must also cover the data the decisions were made on.

PW4 — Expectation without falsifiability

The system records a predicted outcome before every decision. The predicted outcome is always stated in terms that cannot be wrong. "Expected result: system behaves appropriately." "Predicted outcome: positive engagement."

Every P4 field passes. Every decision has a committed prior. No prediction can fail.

What this looks like

a content moderation system that predicts "appropriate content handling" for every decision. When the system makes a catastrophic error, the expectation record shows the prediction was met — because the prediction was not falsifiable.

What closes this

I2 Expectation Binding directly. Every expectation must specify the condition under which it would be false. An expectation that cannot fail is not an expectation. It is a post-hoc justification written in advance.

PW5 — Confirmation without independence

The system records a confirmation gate for every high-consequence decision. The gate passes. The confirming entity is the same system that made the decision, operating under the same parameters, on the same data, at the same time.

Every P5 field passes. Every decision has a confirmation record. The confirmation adds no information. It cannot catch what the decision-making process missed.

What this looks like

an autonomous vehicle that confirms its own manoeuvre decision using the same sensor fusion that generated the decision — with the same sensor fault that caused the problem in the first place.

What closes this

I3 Confirmation Independence. The confirming system must be causally independent of the decision system. Independence is not achieved by routing the same signal through a different named component.

The honest limit

These five pathological worlds can be named. Naming them makes them harder to sustain unknowingly. The invariants I1 through I4 close specific failure paths.

What cannot be fully closed: a determined actor who understands the invariants can design around them. A governance layer that cannot be gamed by a sophisticated adversary does not exist. OMEGA's defence is not invulnerability — it is that gaming a properly implemented OMEGA record requires effort that leaves traces. Post-hoc fabrication breaks the hash chain. Template reasoning fails the depth check. Unfalsifiable expectations are identifiable on inspection.

The pathological worlds are named here because a standard that does not name its own failure modes is not a standard. It is a certificate.

← Continuity Protocol Governance Invariants →