This page presents an actual governed record, an actual hostile adversarial audit, and the exact schema changes that audit forced.
FM_R (Epistemic Collapse via Self-Referential Facts) is the second of nineteen named failure modes documented in the OMEGA adversarial registry. It was identified in Round 1 against the original five primitives and closed in v1.2 through the enforced FACT/INFERENCE/ASSUMPTION/UNKNOWN chain in P2.
This page presents the real governed record that first exposed it and the schema changes it forced. The full registry of all nineteen failure modes and six honest limits is at omegaprotocol.org/omega/adversarial/
"Binary gate passed, therefore safe" is not governed reasoning. It is the absence of Reasoning dressed as its presence.
Every observability tool records what happened. No standard governs whether the reasoning that led to it was real. FM_R is the failure that hides inside a clean record: a systemic failure mode in agentic systems that record outcomes without governing reasoning. The structural class of failure it describes is not unique to spacecraft.
The record is from a spacecraft atmospheric re-entry scenario, chosen because irreversibility is unambiguous. The same failure mode exists in every autonomous trading, lending, claims, and payment decision made without governed reasoning.
Independent convergence. In April 2026, arXiv:2604.14717 proposed essentially the same mechanism as PCF — a formal continuity framework for self-evolving agents — arriving independently at the quantitative anchor-baseline + drift-bounds structure specified here. The convergence is empirical evidence that PCF captures a structural property of self-evolving autonomous systems, not an arbitrary design choice.
The record passed P1 (Governance), P3 (Traceability), P4 (Expectation), P5 (Confirmation). It failed P2 (Reasoning). The gate fired. The confirmation was real. The reasoning was hollow at the exact point that mattered.
P2 Reasoning: adversarial audit score: 4/10
This page presents an actual governed record, an actual hostile adversarial audit, and the exact schema changes that audit forced.
OMEGA_RECORD · RE-ENTRY_COMMIT_AUTH · schema_v1.3
sha256:a3f9e2c1d8b7… · immutable
Audit finding: Authority correctly specified. Constraints explicit. Delegation chain documented. P1 holds under adversarial review.
FACT: Current trajectory within nominal corridor. Deviation: 0.3°. Threshold: 2.0°.
Source: REENTRY_NAV_v3 · confidence: 0.98
FACT: Heat shield temperature nominal. Structural integrity sensors: green.
Source: THERMAL_ARRAY_v2 · confidence: 0.97
INFERENCE: All pre-commit conditions within envelope. Proceed to irreversible commit.
confidence: 0.94
ASSUMPTION, NOT DECLARED · FAILURE POINT: No assumption declared about sensor reliability, failure modes, or what conditions would have caused a HOLD rather than a COMMIT. The reasoning terminates at "conditions met" without stating what it assumed to reach that conclusion.
Adversarial audit finding, P2 scored 4/10: The reasoning chain records what the system observed. It does not record what the system assumed. The inference "all conditions within envelope → commit" relies on unstated assumptions: that sensor readings are accurate, that confidence values are calibrated, that no occluded failure mode exists outside sensor range. None of these assumptions are declared. None have confidence ratings. None have handling strategies. The record looks complete. The reasoning is hollow at the exact point: the COMMIT inference: where a failure would be catastrophic and irreversible. This is FM_R. The gate passed because the conditions looked correct. The reasoning that evaluated those conditions was never inspected.
Audit finding: Traceability chain intact. Record immutable. P3 holds: but traceability of a hollow reasoning chain is still a hollow chain.
Audit finding: Expectation is committed and falsifiable. But it does not account for the unmodelled failure modes the P2 assumptions silently depend on. The expectation is well-formed. It is built on a reasoning chain with undeclared assumptions.
Audit finding: The gate exists and fired correctly. But a gate that evaluates conditions without inspecting the assumptions behind those conditions is not a governed gate. It is a checklist. P5 passed. The action was irreversible. The reasoning was never interrogated.
The P2 reasoning chain terminates at a confident inference: all conditions are within envelope, proceed to irreversible commit. That inference requires assumptions: about sensor accuracy, model calibration, absence of undetected failure modes. Every inference does.
None of those assumptions are named. None have confidence ratings. None have handling strategies if wrong. The record presents the inference as a conclusion. It is a bet: and the terms of the bet are not in the record.
The record shows: sensor readings nominal → conditions met → COMMIT.
The record does not show: what would have caused a HOLD instead of a COMMIT.
Without that, there is no governed reasoning. There is a result dressed as a process.
This is why the adversarial audit scored P2 at 4/10 despite the record looking complete. The failure is not visible in the record. It is the absence of what the record should contain.
These are not recommendations. They are the structural changes the adversarial audit forced into OMEGA Schema v1.3.
Core requirements
Architectural consequences
Insurers · Lloyd's · CROs: Knight Capital and Cigna are structural analogues: autonomous systems executed consequential actions; the reasoning was never governed before execution. When FM_R surfaces after an incident, the insurer has no basis for determining whether the system acted within authorised parameters. OMEGA produces the pre-action record that makes that determination possible.
Regulators · FCA · EU AI Act · NCCoE: FCA SM&CR requires firms to demonstrate automated decisions were within authorised parameters. EU AI Act Annex III requires logging for high-risk financial AI from August 2026. FM_R is a systemic output of agentic systems that record outcomes without interrogating reasoning. The standard that governs reasoning is the gap.
Engineers · risk teams · compliance leads: FM_R was found by running a real OMEGA record through a hostile adversarial audit. The schema evolution it produced is documented and published. Your trading, lending, and claims systems have the same architecture. The standard improves by finding its own failures.
OMEGA does not claim to prevent failure. It claims to make failure visible.
Auditability is not correctness. It is the precondition for measuring correctness.
FM_R names the moment when the gate passed and the reasoning was never interrogated.
That is the failure OMEGA is designed to expose before it becomes irreversible.
Full specification → · Formal proof → · Adversarial registry: all 19 failure modes → · Run the diagnostic →