LEN-0280
Reference Stability Lens
Evaluates whether structural references, identifiers, nodes, and edges remain consistent across execution cycles or comparable states.
Primary Pattern Matches
Persistence Instability
A structural condition where persisted state cannot be stored and restored into an equivalent structure without alteration.
Lens Application
The Reference Stability Lens is a primary inspection mechanism for Persistence Instability because it focuses on whether persisted structures return with stable references and equivalent relationships. It can surface altered identifiers, reordered links, broken edges, or reference drift introduced during save, load, serialization, or migration paths.
Inspect For
- Identifier changes after restore
- Broken or remapped references
- Node and edge drift across cycles
- Equivalent state with altered structure
Avoid
Treat reference changes as inspection signals that may indicate Persistence Instability, not as standalone proof that it has been established.
Reference Instability
A structural condition where references, identifiers, links, or anchors change across equivalent evaluations without a declared cause.
Lens Application
The Reference Stability Lens is a primary inspection mechanism for Reference Instability because it focuses directly on whether references remain consistent across comparable states. It helps surface changed IDs, shifted anchors, rewritten links, or altered node-edge mappings when no declared cause explains the variation.
Inspect For
- Identifiers that change between equivalent evaluations
- Links or anchors resolving to different targets
- Node or edge references shifting without explanation
- Reference mappings that differ across cycles
Avoid
Treat reference drift as an inspection signal for further review rather than proof that Reference Instability has been established.
Silent Mutation
A structural condition where a change occurs without a corresponding declared update, authority update, or version change.
Lens Application
The Reference Stability Lens is a primary inspection mechanism for Silent Mutation because it focuses on whether references remain stable between comparable states. It helps surface node, edge, identifier, or linkage changes that appear structurally meaningful but lack matching version, authority, or update signals.
Inspect For
- References that shift while declared version remains unchanged
- Node or edge changes without update metadata
- Identifier reuse after structural alteration
- Divergence between comparable execution states
Avoid
Treat reference drift as an inspection signal requiring comparison against update signals, not as proof that Silent Mutation has occurred.
Secondary Pattern Matches
Authority-State Mismatch
A structural condition where observed system state no longer aligns with the declared authority state that is supposed to govern it.
Lens Application
The Reference Stability Lens applies indirectly to Authority-State Mismatch by highlighting reference drift, identifier inconsistency, or unstable structural links between authority declarations and observed state. It can help surface where mismatch becomes visible across cycles without serving as the primary diagnostic mechanism.
Inspect For
- References that point to outdated authority states
- Identifiers that change across comparable executions
- Nodes or edges that no longer match declared governance structure
- Repeated state observations tied to inconsistent references
Avoid
Treat reference instability as a signal for further inspection, not sufficient evidence that Authority-State Mismatch is present.
Non-Deterministic Execution
A structural condition where equivalent inputs and declared constraints produce divergent outputs across executions.
Lens Application
As a secondary lens, the Reference Stability Lens helps inspect Non-Deterministic Execution by checking whether structural references remain consistent when equivalent inputs and constraints are executed repeatedly. Reference drift, unstable identifiers, or changing node-edge mappings may indicate where divergent outputs are being introduced or amplified.
Inspect For
- Identifier changes across equivalent executions
- Node or edge mappings that shift between runs
- References resolving to different targets under the same constraints
- Output divergence correlated with structural reference drift
Avoid
Treat reference instability as a supporting inspection signal, not as confirmation that Non-Deterministic Execution has been established.
Related Issues
Actual Policy Differs From Declared Policy
The policy the AI or workflow actually follows differs from the policy that is documented, declared, displayed, or expected.
AI Forgets Earlier Constraints
A constraint, instruction, preference, or decision that should persist through the task stops affecting later output.
AI Memory Has No Governance
Saved or persistent AI memory affects output without clear rules for ownership, scope, review, update, expiry, or removal.
AI Memory Updated Without Asking
AI memory, saved context, preference, or durable state is updated without the user clearly asking for or approving that update.
Answer Has No Traceable Source Link
The answer makes a claim, recommendation, citation, or factual statement without a source link or trace path that allows the user to verify where it came from.
Behavior Does Not Match Declared Role
The AI or agent behaves outside, below, or differently from the role, authority, responsibility, or permission posture declared for it.
Cannot Identify Authoritative State
The user or workflow cannot tell which state, version, review result, decision, source, or output is currently authoritative.
Citation Points to Wrong Source
A citation, reference, link, or source pointer is present, but it points to the wrong source, wrong passage, wrong document, or unsupported evidence.
Context Changes After Restore
Restoring, reopening, resuming, or reloading a task changes the context that the AI uses to continue the work.
Context Leaks Between Tasks
Context, assumptions, constraints, examples, files, or decisions from one task affect another task where they should not apply.
Declared Owner Cannot Control Outcome
A person, role, system, or policy is declared responsible for an outcome but does not have the actual authority or control needed to govern it.
Early Model Output Gets Overweighted Downstream
An early AI output receives too much authority in later workflow steps, decisions, reviews, or generated artifacts.
File-Bounded Task Uses Outside Content
The AI is asked to work only from a specific file or document but uses content, assumptions, or sources outside that file.
Human Review and Automation Disagree
A human review result and an automated AI or workflow result disagree without a declared rule for resolving the difference.
Local Exception Grows Into Policy
A local exception, special case, or one-off allowance begins to function like a general policy.
Model and Workflow Disagree on Next Step
The AI model recommends or selects a next step that conflicts with the workflow state, required handoff, routing rule, or process sequence.
No Owner for Agent Action
An agent action can affect the system without a declared responsible owner, authority, or accountable decision path.
Old Output Expectations Survive Migration
Expectations from a prior model, prompt, schema, tool, or workflow survive a migration and continue shaping review or downstream handling after they should be replaced.
Output Changed Without Declared Change
Output shape, content, format, fields, or behavior changes without a declared change to the prompt, schema, model, workflow, or governing rule.
Policy Area Has No Examples
A policy, rule, standard, or guidance area has no examples showing how it should apply to real cases.
Policy Update Not Reflected in Output
A policy, rule, standard, or instruction has been updated, but the AI output still follows the older version.
Prompt Behavior Changed Without Version Change
A prompt begins producing different behavior even though no prompt version, model version, workflow version, or declared dependency change is recorded.
Prompt Only Works After Retry
The prompt fails, misroutes, or produces an unusable response on one attempt but works after retry without a meaningful change to the input.
Relationship Map Has Missing Links
A map of related Issues, rules, patterns, cases, fields, tools, or workflow steps is missing links needed to navigate or reason over the structure.
Results Vary Too Much
Repeated or comparable runs produce outputs that vary more than the task, workflow, or user can tolerate.
Retry Makes the Problem Worse
A retry, repair attempt, regeneration, or follow-up instruction increases the error, expands the failure, or creates additional breakage instead of narrowing the problem.
Revoked Approval Still Treated as Active
An approval, permission, exception, or authorization that was revoked continues to affect AI behavior or workflow decisions as if it were still active.
Routing Overrides Task Intent
Routing, mode selection, agent behavior, or workflow classification sends the task down a path that overrides what the user was trying to accomplish.
Rubric Changed but Results Did Not
A review rubric, scoring rule, evaluation standard, or classification criterion changes, but AI results continue to reflect the old rubric.
Same Contract Name Has Different Meanings
The same prompt, schema, field, policy, tool, or workflow contract name is used in different places with different meanings.
Same Instructions Allow Different Outputs
The same instructions are broad or underspecified enough to allow materially different outputs that all appear compliant.
Saved Memory Not Used
A saved memory, preference, instruction, or durable context item exists but does not affect the AI output when it should.
Saved Reference No Longer Works
A saved source, citation, file reference, prompt reference, or workflow pointer previously worked but no longer resolves to the expected object or meaning.
Schema Reference Loops Without Base Case
A schema, field, type, object, or structured reference points through a loop without a base case that allows validation or interpretation to resolve.
Similar Cases Route to Different Outcomes
Similar inputs, cases, prompts, or workflow states are routed to different outcomes without a declared difference that explains the split.
Stale Context Affects Output
Old context, prior instructions, outdated references, or earlier task state continue to affect output after they should no longer apply.
Task Progress Is Lost Midway
The AI loses track of completed work, prior decisions, current position, or remaining steps before the task is finished.
Validation Result Changes on Retry
A validation, grading, review, classification, or pass/fail result changes after retry even though the input and declared validation rules did not change.
Version Change Breaks Existing Prompt
A prompt that previously produced usable results stops working after a version change in the model, tool, policy, schema, product surface, or workflow.