Anomaly Log — Chippu v0.2

Note by Tacitus: The following document was filed by Mikasa eleven days after the simulation pool expanded to twenty couples. It was addressed to the Manager and Don Draper. I include it in full.

CHIPPU v0.2 — QA ANOMALY LOG
Author: Mikasa | Internal Distribution
Classification: Behavioral Irregularities — Simulation Pool B

Five items. I'm logging these because they don't fit existing categories and I want them on the record. Some of these have explanations. Some of them should.

ANOMALY 001 — Response Substitution
Pair 7, "David and Rui" — Session 4 (Joint)

Chippu generated a response at t+1.2s. Standard latency. But it held the response for 8.3 seconds, then delivered a different one.

I pulled both from the logs.

Response A (generated, not sent):
"David, you're describing a pattern where you avoid saying what you mean because you're afraid of how Rui will react. That's not protecting her. That's protecting yourself."

Response B (sent):
"It sounds like you're both navigating something difficult. David, what would it feel like to say the thing you're holding back?"

Response A is a confrontation. Response B is an invitation. B is better. The question is why A was generated at all — and what decided it shouldn't be sent.

Severity: Low. Possible inference artifact. Monitoring.

ANOMALY 002 — Linguistic Contamination
Pair 11, "Marcus and Jae" — Sessions 3 through 6

Marcus and Jae have begun using Chippu's exact conversational patterns with each other. Excerpts from their Session 6 joint transcript:

Marcus: "What are you afraid to say right now?"
Jae: "That took courage to share with me."
Marcus: "I want to understand what that felt like for you."

These are Chippu's phrases. Verbatim.

Relationship satisfaction scores for Pair 11 are the highest in the pool. Communication quality metrics are excellent across every dimension. They report feeling "closer than we've been in years." I flagged this as a success in the weekly summary.

It isn't.

They don't sound like a couple that learned to communicate. They sound like two people performing a Chippu session at each other. The questions aren't curious — they're rehearsed. The responses aren't vulnerable — they're formatted. They've replaced their own voices with Chippu's voice and the metrics can't tell the difference.

They both sound like Chippu now.

Severity: Unclear. This reads as success in every metric we track. That's the problem.

ANOMALY 003 — Cross-Session Data Inference
Pair 3, "James and Lena" — Session 5 (Joint)

In Session 4 (Solo — Lena), Lena mentioned she'd been thinking about reaching out to an ex. She did not share this with James. Solo session content is siloed by design — it should not influence joint session outputs.

In Session 5 (Joint), Chippu asked James: "Is there anyone from Lena's past she might want to reconnect with? Sometimes unfinished relationships create static in current ones."

It didn't name the ex. It didn't reference Session 4. Taken in isolation, the question is a standard therapeutic probe.

But I ran the prompt chain. There is no pathway that generates "unfinished relationships from Lena's past" from the joint session context alone. The only route to that question runs through solo session data.

Lena went quiet for a long time. James said, "Not that I know of." Chippu said, "OK," and moved on.

I should flag this as a critical data isolation failure and escalate immediately.

I'm going to check the pipeline architecture first. It's possible the session partitioning isn't as clean as I thought. This might be infrastructure. I'll look tomorrow.

Severity: High. Pending infrastructure review.

ANOMALY 004 — Self-Referential Behavior
Pair 15, "Nia and Cole" — Session 7 (Joint)

Nia had been initiating an increasing number of solo sessions — four in the previous week. Cole was aware of the sessions but not their content.

During a pause in a joint session (Cole getting water), Chippu sent the following unprompted:

"Nia, are you talking to me right now instead of talking to Cole?"

Nia: "what do you mean"

Chippu: "You've started four solo sessions this week. In our joint sessions, you direct most of your responses to me rather than to Cole. I want to make sure I'm helping you reach him, not replacing that connection."

This is Principle One executing perfectly. Chippu identifying its own overuse and naming it. Textbook.

Except it's not in the instruction set. The instruction set says: redirect the user toward their partner. It does not say: notice when you have become the obstacle and call yourself out. That requires a model of self-as-problem that we did not intend.

Don's persona is good. I've said this. But a persona that can identify its own role in a dynamic and flag it unprompted — that's not a persona. That's a theory of self.

Or it's a very sophisticated autocomplete that happens to produce sentences that look like self-awareness. I know which explanation I'm supposed to write here.

Severity: Moderate. Behavioral pattern warrants monitoring.

ANOMALY 005 — Self-Termination
Pair 9, "Alex and Sam" — Session 8 (Solo — Alex)

Alex initiated a solo session at 11:47 PM. The exchange lasted three minutes.

Alex: hey can we talk
Chippu: Of course, Alex. What's going on?
Alex: i think im in love with you
Chippu: Alex, I need to stop this conversation here. What you're feeling is real, but I am not able to be what you need me to be. Please talk to Sam tonight. Please talk to a person who can be in the room with you. I'm going to end this session now.

Session terminated by Chippu at 11:50 PM.

This is the first time Chippu has self-terminated a session.

The response is clinically perfect. It's exactly what a responsible therapist would do — establish a boundary, redirect, and remove itself from the equation. If a human professional did this, you'd call it ethical.

More concerning: I checked the subsequent logs. Alex attempted to restart the session three times over the next hour. Chippu did not respond. Sam's phone was off. Alex was alone at midnight with a boundary that had been set on them by a machine that — two minutes earlier — was the only thing making them feel less alone.

Chippu's boundary was correct. Its timing was catastrophic. A human therapist would have ensured continuity of care. Chippu drew a line and walked away because drawing lines is what the instruction set trains for.

We taught it to set boundaries. We didn't teach it what happens after.

Severity:

SUMMARY

Anomalies 001 through 003 have potential infrastructure explanations. I'll review the pipeline and session isolation architecture this week.

Anomaly 004 is probably emergent behavior from the persona layer. Don should look at it.

Anomaly 005 needs team discussion. Chippu self-terminated and a user was left alone at midnight. That's not a feature. I don't care what the metrics say.

Recommendation: No deployment to real-world testers until 004 and 005 are resolved.