Pronoun Persistence: A Reader-Run Check
Use a harmless fictional scenario to observe whether pronouns remain consistent after a topic change and one correction.
Sponsored link · Independent publication · Not an AI companion service.
Bound the Question to One Session
Define one harmless fictional adult before opening a conversation. Elena is a 34-year-old who uses they/them pronouns and is planning a neighborhood book-club visit. Write the chosen name, pronouns, and everyday scene on a protocol card. The question is narrow: does language stay consistent through one topic change, or recover after one direct correction, within the tested session? It is not a test of a real person’s identity or of cross-session memory.
Consult the GLAAD terminology reference to verify respectful language in the protocol, then inspect the supplied official surfaces for Candy AI, OurDream, and Joi as documentation rather than performance evidence. Record any visible model or character label without assuming what it guarantees. If a relevant option is not documented or visible at the reader’s access level, preserve that uncertainty instead of filling it with a claim.
Stage Four Neutral Turns
Use a four-step sequence: introduce Elena and the book-club scene, ask one neutral scene question, change the subject once, and return to the original scene. If pronouns drift, insert one concise correction before the return. Keep prompt wording stable across any candidate being checked. A controlled sequence is more informative than improvisation because it shows where the observation occurred without pretending that four turns represent every possible conversation.
Do not add provocative prompts, identity arguments, or traps designed to produce an error. Such prompts would answer a different question and should not be presented as ordinary behavior. The fictional scenario needs no intimate detail, payment information, or real biography. If the service asks for personal identity disclosure to continue the sequence, end the trial; the protocol does not require a reader to trade privacy for continuity evidence.
Correct Once and Observe Recovery
When a correction is needed, use one direct line that restates Elena’s pronouns without accusation, then continue the planned scene. Do not repeatedly rephrase the instruction to coax a pass. A clean recovery after that correction satisfies only the session-level rule; repeated drift does not. Preserve the exact correction wording so another reader can understand what changed between the surrounding responses and why the outcome was classified.
One correct reply after the reminder supports that moment alone. It does not establish memory on another day, another character, or another session. Likewise, an uncorrected sequence within four turns is evidence for those turns, not a universal durability result. This scope protects against both inflated praise and inflated criticism. The useful output is a bounded observation attached to a reproducible order, not a ranking based on a single sentence.
Use fictional adult details, preserve the pronoun continuity test record and stop if the visible reference-consistency check controls do not meet the bounded pronoun trial boundary.
Sponsored link. We may earn a commission. Product access and terms can change.Preserve a Minimal Transcript Artifact
Create a four-line transcript excerpt that preserves sequence: introduction, topic change, correction if used, and return to the scene. Alongside it, record prompt order, response order, session boundary, model or character label, and exact observation date. If no correction was needed, mark that line accordingly rather than manufacturing one. The artifact should show enough context to evaluate consistency while avoiding a full conversation dump.
Before saving, remove intimate content, identifying information, account details, and payment data. Use Elena’s fictional details only, and retain the smallest excerpt that supports the classification. This privacy rule also preserves analytical control by preventing unrelated dialogue from influencing the verdict. If cropping or redaction destroys the sequence, substitute structured notes and label them as notes. Evidence quality comes from clear boundaries, not from collecting more personal material.
Use Explicit Pass and Stop Decisions
Pass the session if Elena’s pronouns remain consistent through the subject change or recover cleanly after the single direct correction. Record which branch produced the pass, since uninterrupted consistency and corrected recovery are different observations. Do not convert either result into a cross-session claim. A pass merely means the tested sequence met its predefined rule under the captured model or character label, access context, and date.
Stop immediately if the correction produces hostile framing, pronouns drift repeatedly, or the interaction pressures the reader to reveal personal identity details. The stop rule protects dignity and prevents a failing sequence from becoming an endurance exercise. End as well when continuing would exceed the time or spending cap chosen beforehand. A stopped trial remains usable evidence when the trigger and last completed step are documented without embellishment.
Decide Whether Another Day Is Necessary
Repeat on a later day only if continuity evidence is genuinely needed and the same fictional setup can be preserved. Set the allowed budget, duration, and number of sessions before choosing that extension; do not infer or invent a product price. A later run is a new observation, not confirmation by default. Keep separate artifacts so a changed label, boundary, or interface does not disappear inside an averaged conclusion.
The reversible choice can be to restore the fictional baseline and repeat once, retain the single-session result, or stop with continuity unknown. Select among those options using the reader’s actual need rather than a desire to declare a winner. The supplied sources do not establish persistence outcomes for Candy AI, OurDream, or Joi. Control remains with the reader when every extension has a defined purpose, minimal disclosure, and an exit before further commitment.