Table of Contents

Adaptive Risk Context, Freshness, and Drift

Learning objective: Understand how time-sensitive and probabilistic risk observations remain historical evidence while current policy decides when stale signals, provider/model changes, threshold changes, resource drift, or environmental drift require reevaluation before consequential execution.

Pattern classification: General learning material

Advanced area: Adaptive risk context, freshness, and drift

Difficulty: Advanced

Required prerequisites: Risk-Based Decisions in Governed Systems and Deterministic and Probabilistic Inputs in Policy Evaluation.

Recommended background: Policy Context and Explicit Decision Outcomes, Scoped Capability and Host-Owned Execution, Policy Versioning and Decision Provenance, and Safe Degraded Mode and Fail-Safe Governance.

Glossary: Decision provenance, execution authority, host-owned execution, and trust boundary.

Framework-neutral scope: This article teaches freshness, provenance, reevaluation, degraded-mode, and execution-boundary semantics around changing risk inputs. It does not define a fraud model, adaptive-learning algorithm, drift detector, MLOps platform, compliance standard, or production risk service.

1. Assumptions and Non-Goals

This treatment assumes:

  • A host can distinguish authoritative deterministic facts from probabilistic or model-derived observations.
  • Risk observations can carry provider, model, scoring-method, calibration, and observation-time provenance when those identities matter.
  • Governance policy owns the interpretation of risk observations, including threshold and degraded-mode rules.
  • Consequential delayed execution can validate current host context before invoking the protected executor.
  • Historical decision evidence can preserve the exact observation used rather than silently replacing it with a later score.

This treatment does not assume:

  • A probability is perfectly calibrated.
  • A model output is an authoritative fact.
  • Newer model versions are automatically safer or more permissive.
  • A risk provider owns authorization.
  • Every stale observation must produce the same outcome.
  • Every model or data drift event must be detected inside the governance evaluator.
  • A feedback loop should automatically retrain, deploy, or change policy.
  • One universal risk threshold is correct across operations or consequence levels.

The central lesson is:

A changing risk signal should trigger explicit reevaluation rules, not silently mutate authorization or execution authority.

2. At a Glance

A useful lifecycle is:

Authoritative host facts
        +
Captured risk observation
        |
        v
Versioned governance policy
        |
        v
Governance decision
        |
        v
Narrow execution authority when allowed
        |
        v
Current-context + freshness validation
        |
        +-- current ----------> host-owned execution
        |
        +-- drifted ----------> reevaluate
        |
        +-- unavailable ------> explicit degraded posture

Keep these layers distinct:

Risk observation
!= governance decision
!= execution authority
!= protected side effect

A risk observation can be necessary evidence without becoming permission.

3. Keep the Artifacts Distinct

A delayed risk-aware workflow commonly contains at least five different artifacts:

Artifact Meaning What it does not mean
Risk observation What a named provider/model reported at a known time Authorization
Deterministic context Host-owned facts such as resource version, amount, destination, or incident posture Risk score
Governance policy How rules interpret deterministic and probabilistic inputs Model output
Governance decision Explicit policy outcome for one captured context Permanent execution permission
Execution authority Narrow host-issued authority for a specific execution boundary A reusable risk assessment

A risk service can legitimately report:

FraudProbability = 0.21

without being allowed to say:

Therefore release the payment.

The host's versioned governance policy owns that interpretation.

4. Running Scenario: payment.release

At 10:00, a fictional payment-release workflow captures:

PaymentId = pay-981
ResourceVersion = pay-981:v1
Amount = 250000 USD
DestinationApproved = true
IncidentPosture = Normal
EnvironmentVersion = env-normal-v1

FraudProbability = 0.21
Provider = fraud-service
Model = fraud-detector
ModelVersion = risk-v7
ScoringMethodVersion = fraud-score-v3
CalibrationVersion = fraud-cal-2026-08
ObservedAt = 10:00
ProviderValidUntil = 10:10

Policy returns:

Decision = Allowed
PolicyVersion = payment-policy-v12
ThresholdVersion = threshold-v12
FreshnessRuleVersion = freshness-v1

Before execution, current state becomes:

Payment resource = pay-981:v2
Incident posture = Elevated
EnvironmentVersion = env-elevated-v2

New observation:
FraudProbability = 0.76
ModelVersion = risk-v8
ObservedAt = 10:04

The question is not which number should overwrite 0.21. The questions are:

  • Which facts remain historical evidence?
  • Which current facts changed?
  • Which changes invalidate the old authority?
  • What does current policy require before execution?

The original 0.21 / risk-v7 observation remains part of the original decision provenance. The 0.76 / risk-v8 output is a new observation.

5. A Risk Observation Needs Identity, Not Only a Number

A useful teaching envelope is:

public sealed record RiskObservation(
    string ObservationId,
    string SignalName,
    decimal FraudProbability,
    string ProviderId,
    string ModelId,
    string ModelVersion,
    string ScoringMethodVersion,
    string CalibrationVersion,
    ModelHealth ModelHealth,
    DateTimeOffset ObservedAtUtc,
    DateTimeOffset ProviderValidUntilUtc);

The identities answer different questions:

Field Question
ObservationId Which exact captured output did policy consume?
SignalName Which named risk signal does the value represent?
ProviderId Which source asserted the observation?
ModelId / ModelVersion Which model family/revision produced it?
ScoringMethodVersion Which scoring/preprocessing contract applied?
CalibrationVersion Which calibration treatment applied?
ModelHealth What explicit health state accompanied the observation?
ObservedAtUtc When was the signal observed?
ProviderValidUntilUtc Until when did the provider claim applicability?

Provenance does not prove the model was correct. It makes later reasoning possible.

6. New Model Output Is New Historical Evidence

Suppose a decision used:

ObservationId = risk-observation-1001
FraudProbability = 0.21
ModelVersion = risk-v7

A later run produces:

ObservationId = risk-observation-2001
FraudProbability = 0.76
ModelVersion = risk-v8

Do not rewrite the old decision record.

Historical decision
        |
        +-- observation 1001 / risk-v7 / 0.21

New evaluation
        |
        +-- observation 2001 / risk-v8 / 0.76

Re-running the model later is not equivalent to reconstructing the original input. Model version, calibration, upstream data, preprocessing, time-sensitive features, or provider behavior may have changed.

Required invariant

Model version changes
        |
        v
Prior observation remains historical evidence
        +
New output becomes a new observation

Decision identity is separate from risk-observation identity. Two evaluations can consume the same observation and still be different governance events because decision time, threshold policy, freshness policy, or deterministic context changed. The companion sample therefore requires the host to supply a distinct DecisionId for each decision event rather than deriving identity only from payment, policy, and observation fields.

Same payment + same observation
+ different evaluation event
        |
        v
Distinct DecisionId
Preserved historical outcomes

7. Provider Validity and Host Freshness Policy Are Different

A provider may report:

ProviderValidUntil = 10:30

The host may still require:

MaximumSignalAge = 5 minutes

Then the effective acceptance window is the earlier of:

provider-valid-until
host-observed-at + maximum-policy-age

For example:

ObservedAt = 10:00
ProviderValidUntil = 10:30
Host MaximumSignalAge = 5 minutes

Effective validity = 10:05

The provider supplies evidence and validity metadata. The governance policy decides whether that evidence is fresh enough for this operation.

8. Staleness Is a Policy State, Not an Automatic Denial

When an observation is stale, legitimate current policies can choose different responses:

Re-score and reevaluate
Defer until current evidence is available
Escalate to human review
Deny for selected critical operations
Continue under deterministic rules for a low-consequence operation

There is no universal stale-signal outcome.

The companion sample models two explicit execution-freshness dispositions:

StaleSignalDisposition = Reevaluate

or:

StaleSignalDisposition = Defer

The important requirement is that the disposition is named policy, not a cache default.

The companion issuer normally makes authority expire no later than the same effective risk-freshness boundary. The focused stale-execution tests deliberately extend authority expiry to isolate the evaluator's stale-signal disposition. That branch is therefore defense in depth; normal issuance often rejects the authority at its tighter expiration boundary first.

Required invariant

Risk signal becomes stale
        |
        v
Current policy decides the required next step

9. Provider Unavailable Is Not Low Risk

Unsafe collapse:

fraud provider unavailable
        |
        v
FraudProbability = 0.00
        |
        v
Allowed

Safer representation:

RiskSignalAvailability = Unavailable
        |
        v
Current degraded-mode policy
        |
        +-- Defer
        +-- Escalate
        +-- Deny for selected operations
        +-- Continue only where policy explicitly permits it

The companion sample chooses Deferred. That is a teaching choice, not a universal mandate.

Required invariant

Risk provider unavailable
        |
        v
No manufactured low-risk observation
No execution authority from missing evidence

10. Model Version and Threshold Version Are Different Drift Dimensions

A model change can alter the observed signal:

Same threshold policy
        +
risk-v7 -> risk-v8
        |
        v
Potentially different observation

A threshold-policy change can alter the governance outcome without changing the observation:

Same captured observation = 0.76
        +
threshold-v12 -> threshold-v13
        |
        v
Potentially different outcome

For example:

threshold-v12: Escalate at >= 0.80
threshold-v13: Escalate at >= 0.75

Preserve both ModelVersion and ThresholdVersion. They answer different reconstruction questions.

Scoring-method and calibration versions can be separate again. If those identities materially affect what the number means, a change should be visible rather than hidden behind an unchanged model label.

11. Define Reevaluation Triggers Explicitly

For payment.release, the sample treats evaluation order as part of the teaching contract. First-match behavior is intentional:

1. Hard authority time / audience / operation / payment-identity rejects
2. Policy and deterministic-context drift
3. Current evidence availability and time validity
4. Same-observation integrity
5. Current policy acceptance of signal / provider / model
6. Approved risk provenance drift
7. Staleness under the current freshness rule
8. Proceed only when every prior check remains current

This ordering prevents a softer result from hiding a harder problem. Payment substitution is rejected even if policy also changed. An unapproved current provider is reported as unapproved evidence, not merely as drift. The teaching policy intentionally approves more than one signal/provider value so "unapproved" and "approved but different from the authority binding" are both reachable states. Within the risk checks, a reused observation identity is treated as immutable evidence: if any captured bound fact changes under the same ObservationId, the sample rejects it as an integrity mismatch. Only a different observation identity enters ordinary provenance-drift classification. Provenance drift is surfaced before a stale-signal disposition so a stale replacement model does not look like only an age problem.

Change Teaching response
Execution authority not yet valid or expired Reject old authority
Authority audience or operation mismatched Reject
Payment identity substituted Reject
Policy version changed Reevaluate
Threshold version changed Reevaluate
Freshness-rule version changed Reevaluate
Payment resource version changed Reevaluate
Payment amount changed Reevaluate
Destination approval state changed Reevaluate
Incident posture changed Reevaluate
Environment version changed Reevaluate
Required risk provider unavailable Defer in the sample
Current observation missing Defer in the sample
Observation timestamp is in the future Defer; future-dated evidence is not current evidence
Signal is not approved by current policy Defer as risk.signal-unapproved
Provider is not approved by current policy Defer as risk.provider-unapproved
Model identity/version is not approved by current policy Defer as risk.model-unapproved
Approved signal name changed from authority binding Reevaluate
Approved provider identity changed from authority binding Reevaluate
Model version changed Reevaluate
Scoring-method version changed Reevaluate
Calibration version changed Reevaluate
Model-health state changed Reevaluate
New observation replaced the one bound to authority Reevaluate
Same observation identity now carries changed captured facts Reject as integrity mismatch
Signal became stale Reevaluate or defer according to current policy
Authority was already claimed Reject replay / duplicate use
Final executor sees an authority-command binding mismatch Reject

A production system may choose a different precedence. What should not remain implicit is the assumption that an unexpired authority makes every upstream fact current or that every non-current condition has the same remediation.

12. Resource and Environmental Drift Are Separate From Risk Drift

A fresh risk score can still be bound to stale deterministic context.

Decision time:
PaymentVersion = pay-981:v1

Execution time:
PaymentVersion = pay-981:v2

The host must ask both:

Is the risk evidence fresh?

and:

Is this still the resource/environment state the decision authorized?

Environmental examples include incident posture, regional status, threat posture, session assurance, or other host-defined state.

The teaching sample binds:

ResourceVersion
Amount
DestinationApproved
IncidentPosture
EnvironmentVersion

Required invariant

Material context changes before execution
        |
        v
Old authority does not silently execute
Revalidate or reevaluate as defined

13. Execution Freshness Is a Separate Boundary

A conceptual execution authority can bind:

DecisionId
Operation
PaymentId
Resource version
Amount
Destination approval state
Incident posture
Environment version
Risk observation identity/value/time
Signal/provider/model/scoring/calibration identity
Model-health state
Policy version
Threshold version
Freshness-rule version
Issued-at time
Expiration
Audience

The executor should not accept a raw fraud probability as a credential.

Risk observation
        |
        v
Governance decision
        |
        v
Scoped authority
        |
        v
Execution-freshness validation
        |
        v
Atomic single-use claim
        |
        v
Validated local command
        |
        v
Final authority-command validation
        |
        v
Host-owned executor

The companion sample bounds authority expiration by the effective risk-evidence freshness window, so a newly minted grant cannot silently outlive the evidence that justified it.

It also models bounded use in process. After freshness validation succeeds, the gateway atomically claims the authority before invoking the executor. A replay of the same authority cannot create a second protected execution, including when two callers race concurrently.

The final executor receives both the authority and a host-created ValidatedPaymentCommand. It rechecks time, audience, operation, and command bindings including payment/resource identity, amount, environment, decision identity, policy identity, and observation identity. This does not replace the gateway's current-context refresh; it makes the final host boundary reject command substitution or an authority that became invalid between validation and execution.

If the executor rejects after the single-use claim, the sample keeps the claim consumed. A real system would reconcile the failed/ambiguous attempt explicitly rather than silently restoring reusable authority.

The issuer also treats stale evidence as a runtime issuance result rather than an exception. Programming/configuration mismatches remain explicit reason-coded issuance failures in the teaching API.

14. The Running Drift Scenario Must Reevaluate

Initial state:

0.21 / risk-v7 / Normal / pay-981:v1
        |
        v
Allowed
        |
        v
Authority A

Current state:

0.76 / risk-v8 / Elevated / pay-981:v2

Unsafe behavior:

Authority A
+ overwrite stored score
        |
        v
Continue under old authority

Safer behavior:

Current context differs
        |
        v
Authority A cannot execute
        |
        v
Capture current observation
        |
        v
Rebuild current policy context
        |
        v
Reevaluate
        |
        v
EscalationRecommended in the teaching policy

The new decision does not rewrite the old one. Both remain reconstructable.

15. Model Drift, Data Drift, and Model Health Are Inputs to Governance

Upstream monitoring may report states such as:

ModelHealth = Degraded

or:

OutOfDistribution = true

Those can become explicit policy inputs. The governance evaluator does not need to become the drift detector.

Avoid one opaque loop that:

Detects drift
        |
        v
Retrains model
        |
        v
Deploys replacement
        |
        v
Changes thresholds
        |
        v
Changes authorization

Model monitoring, model deployment, threshold policy, and execution authorization have different owners and failure modes.

The sample accepts ModelHealth as an input and routes degraded health to EscalationRecommended; it does not diagnose drift itself.

16. Feedback Loops Need Governance of Their Own

Decision policy can change the data-generating process.

High risk score
        |
        v
Request denied
        |
        v
No downstream outcome observed
        |
        v
Training data differs from the population originally scored

Or review policy can create more labels for one subset of cases than another.

The point is not that any one loop is inherently wrong. The point is that operational decisions should not be treated as unquestioned ground truth for automatic retraining or policy changes.

Separate lifecycles for:

  • operational decision evidence;
  • outcome/label collection;
  • training-data selection;
  • model evaluation and approval;
  • model deployment;
  • threshold-policy deployment.

Do not make the hidden loop:

Decision outcome
        |
        v
Automatically changes model or threshold
        |
        v
Immediately changes future authorization

without an explicit reviewable update process.

This article does not prescribe one retraining, fairness, drift-detection, or model-governance algorithm.

17. Human Review Is a Legitimate Uncertainty Response

For high-consequence operations, uncertainty can itself be policy-relevant:

ModelHealth = Degraded
Calibration version not approved
Signals disagree
Out-of-distribution indication

A legitimate policy may return EscalationRecommended instead of pretending uncertainty is low risk or a deterministic denial.

Human review does not make the model correct and does not create execution authority by itself. After a delayed review, current context and risk evidence may still need refresh and current policy reevaluation.

18. Deterministic Policy Can Interpret Probabilistic Observations

A policy can remain deterministic even when one input is probabilistic:

Captured FraudProbability = 0.76
ThresholdVersion = threshold-v13

If DestinationApproved = false
    -> Denied

Else if FraudProbability >= 0.90
    -> Denied

Else if FraudProbability >= 0.75
    -> EscalationRecommended

Else
    -> Allowed

This preserves an important distinction:

Signal generation may not be reproducible
        +
Captured signal is preserved
        +
Policy interpretation can be replayed deterministically

A low probabilistic score should not weaken an authoritative deterministic denial.

19. Failure and Drift Matrix

Situation Teaching interpretation Consequential execution
Fresh approved observation + unchanged bound context Current May proceed under valid, unclaimed authority
Fresh observation says high risk New policy input Reevaluate; outcome may deny/escalate
Observation stale Freshness state Reevaluate or defer according to current policy
Risk provider unavailable Missing required evidence Deferred in sample; no synthetic low score
Observation timestamp is in the future Invalid current-time relationship Deferred in sample
Signal/provider/model rejected by current policy Unapproved evidence Deferred; do not relabel as ordinary drift
Approved signal/provider/model identity changed New observation provenance Reevaluate
Threshold version changed Policy drift Reevaluate
Scoring/calibration version changed Signal-semantics drift Reevaluate
Model health degraded Explicit uncertainty input Escalation in sample
Resource version, amount, or destination state changed Deterministic context drift Reevaluate
Incident/environment changed Environmental drift Reevaluate
Payment identity substituted Hard authority-binding failure Reject even when softer drift also exists
Authority not yet valid or expired Authority freshness failure Reject old authority
Audience or operation mismatched Authority binding failure Reject
Same observation ID carries changed captured facts Evidence integrity failure Reject
Authority already claimed Replay / duplicate use Reject; no second executor call
Executor rejects after claim Post-claim execution failure Keep claim consumed; reconcile explicitly
Executor command differs from authority bindings Final host-boundary mismatch Reject

The matrix distinguishes:

risk says high

from:

risk cannot currently be established

and from:

current evidence is not approved by policy

Those are different governance facts and can require different remediation.

20. Common Failure Modes

Failure mode Why it is dangerous
Score directly gates executor Risk service quietly becomes authorization authority
Missing provider becomes zero Dependency failure masquerades as low risk
Stale score reused indefinitely Old evidence silently outlives intended context
New score overwrites old decision evidence Historical reconstruction becomes false
Distinct decisions reuse one derived DecisionId Different outcomes become ambiguous in audit/provenance history
Model and threshold versions collapse together Cannot distinguish scoring drift from policy drift
Resource substitution checked after soft drift Reevaluate can hide a harder wrong-resource reject
Unapproved provider/signal is reported only as drift Trust/acceptance failure loses its distinct remediation
Resource drift ignored because score is fresh Evidence may refer to the wrong resource state
Provider ValidUntil treated as host policy External source decides local freshness semantics
Model health hidden inside numeric score Uncertainty becomes invisible to policy
Authority remains reusable after execution Replay can duplicate a protected side effect
Executor trusts a gateway-created command without binding validation TOCTOU/substitution can bypass the final host boundary
Decision outcome feeds model automatically Learning and authorization loops become circular
Current model rerun explains old decision New observation is mistaken for historical input
Low score overrides deterministic denial Advisory evidence weakens authoritative fact

21. When Simpler Deterministic Policy Is Better

Prefer a direct authoritative rule when the host already knows the relevant fact.

DestinationApproved = false
        |
        v
Denied

is better than manufacturing a destination-trust score when the allowlist fact already expresses the requirement.

A low-consequence synchronous operation may not need a risk provider, observation versioning, delayed freshness validation, or adaptive feedback. Use the extra boundary when uncertainty materially improves the decision.

22. Design Checklist

Before using changing risk context in a governed workflow, answer:

  1. What exact observation did policy consume?
  2. Who provided it?
  3. Which model, scoring-method, and calibration versions produced it?
  4. When was it observed?
  5. What freshness rule does the host policy apply?
  6. What happens when the provider is unavailable?
  7. Which model versions are accepted?
  8. Which threshold policy interpreted the observation?
  9. Which deterministic facts take precedence?
  10. What resource/environment changes require reevaluation?
  11. What is bound into execution authority?
  12. Does a new observation create new evidence rather than mutate history?
  13. Can history be reconstructed without re-running the model?
  14. Is model-health/drift state explicit rather than hidden in a score?
  15. Can feedback alter model or policy only through a separate reviewed lifecycle?
  16. Would a deterministic rule be clearer?

23. Companion Sample

Run the Adaptive Risk Context sample.

The sample is local and deterministic. It models:

  • captured risk observations with provider/model/scoring/calibration provenance;
  • caller-owned decision identity for each evaluation event;
  • versioned threshold and freshness policy;
  • deterministic policy interpretation;
  • result-based narrow authority issuance;
  • authority lifetime bounded by effective risk freshness;
  • execution-time drift validation with explicit precedence;
  • in-process atomic single-use authority claims;
  • final executor validation of authority-command bindings;
  • a dry-run host-owned payment executor.

It does not perform fraud inference or drift detection. Its purpose is to make freshness, bounded-use, provenance, and authority boundaries executable. The console demonstration also walks one Allowed decision through bounded authority issuance, freshness validation, single-use claim, validated command construction, dry-run execution, and a rejected replay so dotnet run reaches the same final boundary described here.

24. Executable Coverage

Invariant Coverage
Low risk remains an input to governance ✅ LowRiskObservationProducesAllowedDecision
Deterministic destination denial cannot be weakened by low risk ✅ LowRiskCannotOverrideDestinationDenial
Provider unavailability becomes an explicit deferred decision ✅ UnavailableProviderProducesDeferredDecision
Missing observation remains explicit ✅ MissingObservationProducesDeferredDecision
Future-dated observation is not treated as current evidence ✅ FutureObservationProducesDeferredDecision
Unapproved signal is distinct at decision time ✅ UnapprovedSignalNameIsDeferred
Unapproved provider is distinct at decision time ✅ UnapprovedProviderIsDeferred
Unapproved model is distinct at decision time ✅ UnapprovedModelIsDeferred
Stale observation becomes an explicit decision state ✅ StaleObservationProducesDeferredDecision
Host maximum age can expire provider-still-valid evidence ✅ HostMaximumAgeCanExpireProviderStillValidObservation
Degraded model health can route to review ✅ DegradedModelHealthProducesEscalationRecommended
High risk can deterministically deny under policy ✅ HighRiskProducesDeniedDecision
Elevated risk can route to escalation ✅ ElevatedRiskProducesEscalationRecommended
Elevated incident posture can change outcome even with a low score ✅ ElevatedIncidentPostureCanEscalateLowScore
Same observation plus changed threshold policy can change outcome without reusing decision identity ✅ ThresholdChangeCanChangeOutcomeWithoutChangingObservation
Model change preserves prior observation as historical evidence ✅ ModelChangePreservesHistoricalObservation
Separate evaluations keep distinct decision identities even when they reuse one observation ✅ SeparateEvaluationsKeepDistinctDecisionIdentityWhenOutcomeChanges
Non-allowed decision cannot mint authority ✅ NonAllowedDecisionCannotMintAuthority
Allowed-looking decision without the bound risk evidence cannot mint authority ✅ MissingRiskEvidenceCannotMintAuthority
Decision/issuance policy mismatch cannot mint authority ✅ PolicyMismatchCannotMintAuthority
Authority expiry is capped by effective risk freshness ✅ AuthorityExpiryIsBoundedByRiskFreshness
Already-stale risk evidence returns a non-issued authority result ✅ StaleRiskEvidenceCannotMintNewAuthority
Unchanged authority produces the explicit freshness.current success value ✅ FreshnessEvaluatorReturnsCurrentForUnchangedAuthority
Current authority reaches executor only after freshness and claim validation ✅ CurrentAuthorityReachesExecutorAfterValidation
Sequential replay cannot execute the same authority twice ✅ ReplayedAuthorityDoesNotExecuteTwice
Two actually concurrent claims produce one execution ✅ TwoActuallyConcurrentAuthorityClaimsProduceOneExecution
Final executor rejects command/resource substitution ✅ ExecutorRejectsCommandResourceSubstitution
Executor rejection after claim leaves single-use authority consumed ✅ ExecutorRejectionAfterClaimLeavesAuthorityConsumed
Not-yet-valid authority is rejected before execution ✅ NotYetValidAuthorityIsRejectedBeforeExecution
Audience and operation bindings are enforced ✅ AuthorityExecutionBoundaryBindingsAreValidated
Expired authority is rejected before execution ✅ ExpiredAuthorityIsRejectedBeforeExecution
Policy/threshold/freshness-rule drift requires reevaluation ✅ PolicyIdentityDriftRequiresReevaluation
Payment identity substitution is rejected ✅ ResourceIdentitySubstitutionIsRejected
Payment substitution outranks softer policy drift ✅ ResourceIdentitySubstitutionTakesPrecedenceOverPolicyDrift
Resource/amount/destination/incident/environment drift requires reevaluation ✅ MaterialContextDriftRequiresReevaluation
Current provider unavailability defers execution ✅ CurrentProviderUnavailableDefersExecution
Current observation missing defers execution ✅ CurrentObservationMissingDefersExecution
Current unapproved signal is not mislabeled as ordinary drift ✅ CurrentUnapprovedSignalDefersInsteadOfReportingDrift
Current unapproved provider is not mislabeled as ordinary drift ✅ CurrentUnapprovedProviderDefersInsteadOfReportingDrift
Current unapproved model is not mislabeled as ordinary drift ✅ CurrentUnapprovedModelDefersInsteadOfReportingDrift
Future current observation defers execution ✅ FutureCurrentObservationDefersExecution
Stale signal can require reevaluation under the named disposition ✅ StaleSignalCanRequireReevaluation
Stale signal can defer under the named disposition ✅ StaleSignalCanDefer
Provenance drift outranks stale disposition in the sample's first-match order ✅ ProvenanceDriftTakesPrecedenceOverStaleDisposition
Approved signal/provider/model/scoring/calibration/health/observation drift is explicit ✅ RiskProvenanceDriftRequiresReevaluation
Same observation identity cannot mutate captured facts ✅ MutatedObservationWithSameIdentityIsRejected
Same observation identity plus model mutation is an integrity reject, not ordinary drift ✅ SameObservationIdentityWithModelMutationIsRejected
Running scenario blocks old authority, preserves history, and reevaluates current state ✅ MaterialDriftBlocksOldAuthorityAndCurrentReevaluationEscalates

Intentionally not modeled

Area Status
Real fraud/model inference ◐ Not modeled; observations are fictional records
Model/data drift detection ◐ Not modeled; model-health state is supplied as input
Model training/retraining ◐ Not modeled
Feature store/model registry ◐ Not modeled
Production calibration measurement ◐ Not modeled
Current governance-policy provider outage ◐ Not modeled; see degraded-mode material for the broader dependency pattern
Durable or multi-instance authority claim storage ◐ Not modeled; atomic single-use claim is process-local only
Distributed replay coordination across hosts/regions ◐ Not modeled
Executor re-fetch of external payment state ◐ Not modeled; final executor validates authority-command/time bindings while the gateway owns current-context reconstruction
Real payment side effects ◐ Not modeled; executor is dry-run/in-memory
Human review workflow ◐ Not modeled; escalation is only an outcome
Fairness/compliance evaluation ◐ Not modeled

Continue with:

26. Closing Principle

A changing risk signal is evidence that the world or scoring system may have changed. It is not permission to rewrite the old decision, and it is not authority to execute under the new state.

Risk observation changes
        |
        v
Preserve old evidence
        +
Apply explicit current freshness/reevaluation policy
        +
Mint or accept execution authority only for the resulting current decision

Changing evidence should cause explicit reasoning, not silent authority mutation.

Read it. Run it. Question it. Improve it.

Edit this page