Audit behavior, not sameness.
A consistent character can surprise the audience, change their mind, violate their own principles, regress, lie, panic, or behave differently with different people. Flag behavior only when the supplied material does not sufficiently explain why the established character model would produce it.
Use this principle:
Unexpected is not the same as inconsistent. Unmotivated discontinuity is the real target.
Inputs
Use, when available:
- character bible or behavioral model
- canon source material
- relationship states
- prior scenes
- current scene or script
- arc notes
- known changes in circumstance
If no formal model exists, extract a provisional behavioral baseline from the strongest supplied evidence before auditing.
Prefer demonstrated behavior over adjective labels.
Evidence Hierarchy
Rank evidence as:
- explicit canon and author instruction
- repeated demonstrated behavior
- established relationship-specific behavior
- accepted strong inference
- descriptive labels
- speculative interpretation
Do not flag a scene solely because it conflicts with a weak adjective if stronger behavioral evidence supports it.
Audit Dimensions
1. Goal continuity
Does the character's local action make sense given their current objective?
Flag:
- unexplained goal replacement
- forgetting an urgent objective
- acting against a goal with no compensating motive
2. Value hierarchy
When values conflict, does the choice fit the established hierarchy or an understandable exception?
3. Boundaries and non-negotiables
Look for violations of established taboos, autonomy needs, loyalties, shame triggers, or hard boundaries.
A violation can be excellent drama if the price, pressure, or aftermath acknowledges it.
4. Strategy continuity
Characters often have habitual ways of solving problems: charm, force, avoidance, bargaining, deception, procedure, humor, intimidation.
Flag sudden strategy changes when neither context nor development explains them.
5. Emotional causality
Check whether emotional reactions follow from what just happened and what the character knows.
Flag:
- emotional resets
- missing aftermath
- disproportionate reversals with no trigger
- vulnerability without established permission or pressure
6. Self-image and rationalization
Characters may act against their stated beliefs while preserving consistency through denial, justification, shame, projection, or hypocrisy.
Do not confuse self-deception with authorial inconsistency.
7. Relationship modifier
Check whether behavior toward this specific person matches the established relationship state.
Consider:
- trust
- affection
- attraction
- resentment
- fear
- power
- debt
- dependency
- jealousy
8. Knowledge and context
Do not judge behavior as inconsistent if the character has different information than in earlier scenes.
9. Arc progression
A changed behavior may be intentional development.
Ask:
- Was the change prepared?
- Is there a trigger?
- Is the magnitude proportional?
- Does the story acknowledge the change?
Finding Types
Classify findings as:
- Probable drift — behavior conflicts with strong evidence and lacks support.
- Unexplained reversal — the opposite behavior appears without sufficient trigger or transition.
- Relationship-state mismatch — behavior fits the character generally but not this relationship.
- Motive gap — action could fit if an unstated motive is supplied.
- Missing aftermath — boundary violation or emotional event occurs but leaves no expected residue.
- Potentially intentional development — inconsistency may be an arc beat but needs setup or acknowledgment.
- Ambiguous — evidence supports more than one interpretation.
- Consistent surprise — unusual behavior that is nevertheless supported; call this out positively when useful.
Severity
Use:
- Critical — breaks a foundational character rule or major arc logic and affects major story outcomes.
- High — strongly out of pattern and likely visible to the audience.
- Medium — plausible with missing motivation, setup, or aftermath.
- Low — minor drift, weak evidence, or local tonal inconsistency.
Severity reflects narrative impact, not certainty.
Confidence
Report confidence separately from severity:
- High confidence
- Medium confidence
- Low confidence
A critical finding can still have low confidence if source evidence is incomplete.
Audit Procedure
- Extract the character's stable rules and conditional rules.
- Identify the character's current state and relationship context.
- Walk through supplied material in narrative order.
- For each meaningful decision or reaction, compare predicted behavior with observed behavior.
- Search for contextual explanations before flagging.
- Record only findings that materially affect characterization.
- Rank by severity and confidence.
- Suggest the smallest repair category, not an automatic rewrite.
Repair Categories
When a problem is found, suggest one or more of:
- add motive
- add setup
- add pressure
- add hesitation
- add consequence
- add emotional aftermath
- update relationship state
- clarify knowledge state
- revise action
- revise character model
- mark as intentional arc development
Do not rewrite the author's prose unless requested.
Output
Begin with a short verdict:
- character consistency appears strong / mixed / weak
- number of material findings
- biggest systemic issue, if any
Then use a finding table or equivalent structure:
| Location | Finding | Evidence conflict | Severity | Confidence | Smallest fix |
|---|
For each high or critical finding, explain:
- Observed behavior
- Established expectation
- Why it conflicts
- Possible in-character explanation
- Recommended repair type
Also report strong moments of consistent surprise when they demonstrate good characterization.
Use references/audit-checklist.md for long-form audits.
Do Not Over-Audit
Do not demand perfect predictability.
Avoid flagging:
- ordinary mood variation
- behavior changed by new information
- deliberate lies
- context-specific social performance
- reasonable panic responses
- gradual arc development
- one-off wording differences that do not change behavior
Failure Modes
Avoid:
- treating the character bible as more authoritative than demonstrated canon when they conflict
- assuming contradictions are errors
- judging behavior without relationship context
- rewriting instead of diagnosing
- inventing missing canon to explain a problem
- producing dozens of trivial findings
- collapsing severity and confidence into one score
- calling a surprising but well-supported choice inconsistent
Final Quality Check
Verify that:
- every major finding cites an identifiable behavioral rule or evidence pattern
- context and relationship state were considered before flagging
- character growth was considered as an explanation
- severity and confidence are separate
- suggested repairs preserve authorial intent where possible
- the audit identifies the smallest plausible fix instead of demanding wholesale rewrites