# OiMy NPS Report v10 — Post Bridge Fix
**Date:** 2026-06-02  
**Evaluator:** Grok 4.3 (in-character user simulation + NPS scoring)  
**Method:** Grok plays a specific parent persona across 3 turns, then scores as that user — not as a technical evaluator

---

## Summary

| Metric | v9 (before) | v10 (after bridge fix) |
|--------|-------------|----------------------|
| **NPS** | **-40** | **+30** |
| Avg score | 3.3/10 | **6.1/10** |
| Promoters (≥7) | 1 | **5** |
| Passives (5-6) | 1 | 3 |
| Detractors (≤4) | 8 | **2** |

**+70 NPS point swing** from fixing the bridge mode (max_tokens: 120→400 on reasoning model).

---

## Results by Scenario

| Scenario | NPS | User feeling | What would improve it |
|----------|-----|-------------|----------------------|
| life_admin | **8/10** | Validated — "gets how bills sneak up on you" | Follow-up checklist/reminder |
| speech_delay | **7/10** | Seen, less crazy for worrying | Concrete next action right away |
| screen_teen | **7/10** | Less crazy — "maybe he's not just shutting me out" | Actual scripts to try |
| parenting_tantrum | **7/10** | Less like a failure — "sees how much this costs me" | One small thing to actually try |
| family_bonding | **7/10** | Less alone, seen | (not captured — last scenario) |
| ot_sensory | 6/10 | Seen, less crazy about suspecting sensory | One concrete next step for tomorrow |
| life_direction | 6/10 | Validating but incomplete | Actual next steps, not reflection |
| money_fights | 6/10 | Less alone with "the knot in my chest" | Finish the thought on what to say to son |
| wellness_burnout | 4/10 | (captured mid-run) | — |
| elder_dementia | **3/10** | Heard, then stuck — "echoing back my words" | One or two practical ideas |

---

## What's Working (Voice Quality)

- **Zero expert-speak violations** across all 10 scenarios — the voice fix from June 1 is holding
- **Peer-voice moments that landed:** "Classic Mark" (money fights), "That guilt's a liar" (burnout), "That grunt wasn't rejection" (screen teen)
- **Bridge message now works:** screen_teen Turn 2 shows bridge_hold working ("hold on, i want to think this through properly") — 7/10 still because Turn 3 delivered
- **Life-admin performed best (8/10):** Specific negotiation scripts, concrete steps, Oi's voice felt casual and helpful

---

## Root Cause of Detractors

**Elder dementia (3/10):** Response is repeating the same 48-char fragment on Turns 2 and 3 ("twice already this week. you said 'at the store..."). This is a DeepSeek caching/repetition bug — the model is echoing rather than advancing. Need to investigate why this specific scenario gets stuck in repetition.

**Wellness burnout (4/10):** Captured mid-run, but pattern visible: emotional validation without actionable follow-through.

---

## Universal #1 Feedback: "Now what?"

Every single NPS score below 8 shared the same improvement: **"Give me one concrete thing to actually do, not just acknowledgment."**

The emotional opening is working (peer voice, meets user where they are). The gap is in the transition from "I hear you" → "here's what to try." Oi needs a cleaner handoff from empathy to action.

This is especially true for:
- DeepSeek-handled skills (short messages go through DeepSeek, which validates well but doesn't always escalate to Kimi with the concrete framework)
- Skills where the user expects specific next steps (OT, speech, parenting discipline, life direction)

---

## Bridge Performance

| | v9 | v10 |
|--|--|--|
| Bridge message on Turn 1 | 39 chars (generic fallback) | 162 chars (specific, peer voice) |
| Example | "On it — give me a moment" | "six months of trying, hitting that exact two-week wall every time. 'hate waking up early' and 'zero accountability' — i can feel the weight of that combo already." |
| NPS impact | Detractor (0-3) | Passive/Promoter (6-7) |

**Root cause of original failure:** DeepSeek V4 Flash is a reasoning model — it uses `reasoning_tokens` internally before generating content. With max_tokens=60 or 120, ALL tokens went to reasoning, leaving 0 for the response content. Fixed by setting max_tokens=400.

---

## Key Bugs Found During Testing

1. **Elder dementia repetition bug (3/10):** DeepSeek returns the same 48-char fragment repeatedly in certain scenarios. Likely a caching artifact when system prompt + conversation repeats the exact same context. Needs investigation.

2. **OT/speech responses truncate at 45-155 chars on some turns:** Expert archetype may be hitting a token limit or having RAG injection issues on specific turn patterns. The 589-char turn on OT works fine, but Turn 1 (45 chars) is clearly truncated.

---

## Next Steps to Reach NPS ≥ 40

1. **Fix "now what?" gap** — After emotional validation, Oi needs to offer ONE specific action. This is a prompt engineering fix, not a routing fix. Opus 4.8 should write a "action transition" section for the DeepSeek system prompt.

2. **Fix elder dementia repetition** — Investigate why the engine echoes the same fragment. May be a `recent_messages` loop where the 48-char response gets stored and retrieved as context.

3. **Fix OT/speech first-turn truncation** — Check max_tokens and RAG injection timing in expert archetype for Turn 1.

4. **Add proactive outreach** — Several users (life_direction, wellness_burnout) expressed feeling "like nothing would actually change." Proactive follow-up (Task 1 already implemented) addresses this directly.