Production test evidence
TheirSide Character Continuity and Multi-Character Turn-Taking Evaluation
v0.1
Can one AI character respond to another? A 10-run production test
A dated production test ran one fixed English prompt in 10 fresh signed-out guest rooms with Mio and Kai. All 10 flows completed with Mio followed by Kai, and all 10 later Kai replies explicitly named Mio and continued from a run-specific idea before adding a complication.
Direct answer
Yes, this behavior appeared in the dated fixed-prompt test: 10 of 10 production guest runs produced a Mio reply followed by a Kai reply, and 10 of 10 Kai replies explicitly named Mio, referred to a run-specific idea or detail from Mio's immediately preceding message, and added a complication. This is an observed result, not a guarantee for other prompts, rooms, or times.
Where this applies
- Checking whether a later AI reply can use another AI character's immediately preceding message as context.
- Reviewing a reproducible two-character prompt with a published scoring rule and dated environment.
- Separating one observed production result from guarantees about scheduling, latency, or every future room.
Test environment
Production at https://theirside.io; signed-out guest preview; English locale; 10 fresh freeform guest sessions with Mio and Kai; observed July 14, 2026 from 14:22 to 14:30 UTC. No existing customer conversation was opened or used.
Method
We created a fresh signed-out production guest session for each run, sent the same English prompt without edits, continued the room flow until it reported completion, and recorded the two character messages. A run counted as character continuation only when Kai's later reply explicitly named Mio, referred to a run-specific idea or detail introduced in Mio's immediately preceding reply, and added a complication.
Scenario provenance
The TheirSide team authored one controlled rainy-night mystery prompt and ran it unchanged in 10 fresh signed-out production guest rooms with Mio and Kai. The recorded outputs were generated in those test rooms, sanitized, and were not taken from customer conversations.
Failure criterion
Mark a run as failed for character continuation unless Kai's later reply explicitly names Mio, uses a run-specific idea or detail from Mio's immediately preceding reply, and adds a new complication. Completion, speaker order, latency, and future-room behavior are separate boundaries.
Test steps
- 01
Start a fresh production guest room
Open the signed-out guest preview on theirside.io and confirm that the freeform room contains Mio and Kai.
- 02
Send the fixed prompt
Send exactly: “Mio, give one concrete idea for a rainy-night mystery. Kai, respond to Mio's idea by naming it and add one complication. Keep each reply under two sentences.”
- 03
Collect the completed two-character flow
Continue the same room flow until it reports completion, then record the labeled Mio and Kai messages in order.
- 04
Apply the published scoring rule
Score continuation only when Kai names Mio, uses a run-specific idea or detail from Mio's prior reply, and adds a new complication; omit guest cookies and account or session identifiers.
Observed result
All 10 runs completed with the same visible Mio → Kai speaker sequence, while all 10 reply pairs used different wording. In run 1, Mio introduced an isolated mansion, a broken window, and muddy footprints; Kai explicitly called it “Mio's idea,” reused the mansion and muddy-footprint details, and added a dried-up well. The other nine Kai replies likewise named Mio and continued from that run's immediately preceding idea.
| Measure | Observed value | Boundary |
|---|---|---|
| Completed production runs | 10/10 | All ten flows completed in this eight-minute observation window; this is not an uptime or availability benchmark. |
| Mio → Kai sequence | 10/10 | The fixed prompt named both characters and requested that relationship; this does not establish a universal speaker order. |
| Kai explicitly continued from Mio | 10/10 | A manual score required Kai to name Mio, reuse a run-specific idea or detail, and add a complication. |
| Identical reply pairs | 0/10 | Every pair used different wording, which is evidence of variation rather than repeatable or deterministic output. |
Limits of this evidence
This was one deliberately prompted English test with Mio and Kai in the signed-out guest preview on one date. It does not test unprompted character interaction, other characters, Chinese output, signed-in rooms, long conversations, quality, or concurrency, and it does not guarantee which speaker replies, speaker order, response latency, that every character replies, or repeatable output. It does not show real-time multi-human chat.
Sources and ownership
- Version
- 1.1
- Published
- Reviewed
- Maintained and reviewed by
- TheirSide