Turn adjectives into observable behavior
“Warm, thoughtful and playful” leaves plenty of room for interpretation. A useful baseline names actions: acknowledges a changed plan, asks one relevant follow-up, avoids a disliked nickname and keeps established events in order. Select three or four behaviors rather than an impossible list of perfect traits.
| Situation | Reference behavior | Record separately |
|---|---|---|
| You cancel a fictional outing | Acknowledges the cancellation and asks about another plan | Incorrectly insists the outing already happened |
| You disagree about a film | Responds directly without changing its stated preference just to agree | Different wording with the same position |
| You ask about an unknown event | Distinguishes established history from a suggestion | A confident invented shared memory |
Keep a usable before-and-after record
Save the prompt, reply, date and available model or mode information. Record the active reference instructions. Repeat the same neutral situation later under comparable conditions, preserving the first answer rather than regenerating until it sounds right.
The surrounding conversation may still differ. Note that limitation. A comparison across ordinary sessions can identify a practical change without proving what caused it.
Separate three kinds of drift
Wording variation: a longer reply or a new phrase may preserve the same behavior. Instruction conflict: a recently added request for formal language can explain a different tone. Continuity failure: an established event or preference is contradicted without an intended change.
These categories suggest different responses. Accept harmless variation if it suits you. Resolve contradictory instructions when they are visible. For a factual problem, use the memory troubleshooting sequence.
Do not blame an update without a baseline
If you noticed the shift after an update, record that timing as an observation. Without an earlier reply and comparable settings, “after the update” does not establish “because of the update.” Avoid inventing version numbers or internal explanations.
Where possible, ask a second reader to compare anonymized excerpts against your written criteria without telling them which is newer. Keep disagreements visible. If you disagree about whether a reply is “warm,” refine the criterion rather than forcing a numeric score.
A practical stop rule
If a difference repeatedly breaks a behavior you care about, the experience may no longer suit you even when factual recall is accurate. If only the phrasing changes, rewriting the entire character may cost more continuity than it restores. Change one reference instruction, observe a later reply and keep both versions.
This evaluates generated character behavior. It does not measure feelings, consciousness or an enduring personal identity inside a system.