Confident No Tolerance — Calibrating the Trust Primitive Against Real Buyers

Why we want this done

This project derives from consumer-experience-kernel A-5 (Ground Truth has one umbrella and four expressions; ratio is the calibration handle) and A-7 (Calibrated Verdict has five structural elements — same physics applies to Confident No), plus productai-kernel A-4 (Adjudication ends searches; Kill Shot disqualifies). Kernel A-5 explicitly states: "If user feedback suggests the product feels 'too negative,' the fix is NEVER 'be more positive' or 'positive mode' — it is always 'make the affirmative guidance carry more Weight.'" That instruction is operational only if the team has empirical data on what Weight makes Confident Nos tolerable and what tips them into perceived negativity. We are about to launch Beat 1 with a Confident No ratio set by hypothesis. The constitutional rule "no positive mode" is brittle without buyer-validated tolerance data. This project gives the chat surface team the empirical floor.

Scope

• Beat 1 cohort: experimental sample of 60-100 skincare and supplements buyers, split across Confident-No-rate arms. Coordinate with PRJ-57 (latent-job project) if running concurrently — share Trust Architect cohort.
• productai-web (chat surface): define the experimental arms (different ratios, different sequence pairings, different verdict structures). Coordinate with chat lead.
• aios/kernels/consumer-experience-kernel.md: propose Signal updates and Axiom revisions based on measured tolerance. The 1:4 ratio in A-5 is a hypothesis; this project tests it. Route via /strategy-sync.
• Trust Architects: the 3-5 skincare TAs validate the calibration spec. Same TAs as PRJ-57 if running concurrently.
• aios/outcomes/chat-expert.md: the verdict-acceptance metric requires verdict-shape decisions; this project unblocks the calibration.

What success looks like

A stranger reads the shipped calibration spec, sees a measured tolerance ceiling number (e.g., "Skincare buyers tolerate up to 28% Confident No rate per session before stated-trust falls below 4/5; sequence-pairing a Confident No with a 5-element Calibrated Verdict in the same session raises tolerance to 41%"), follows the evidence trace to the underlying experiment design, and can run the same experiment if they doubt the result. The Beat 1 launch ships with a verdict-shape decision grounded in this calibration. The kernel A-5 target ratio either gets confirmed empirically (graduates from PROBABLE to FORGED via kernel S-1 graduation path) or gets revised — both outcomes serve the kernel.

The robust solution may surface a better cut: maybe the meaningful variable is not "ratio" but "Confident No density at decision-moment" — the buyer's tolerance is much higher in early-research and much lower at near-purchase. The candidate is invited to propose. What is non-negotiable: experimental design with sequence control, Trust Architect concordance on the calibration spec, and a refusal to settle for "users said it felt fine in interviews" — the kernel explicitly mistrusts stated-trust over behavioral.

References

Constraints

• Claude Code or Claude Co-work primary substrate.
• Two-week duration. Day 1 onsite.
• Oracle separation is structurally required (recruiting-driver §5.2 rule 9): the candidate cannot verify their own experimental result. Trust Architect concordance or a separate analytic eye is the external anchor.
• Privacy-first per productai-kernel A-7. Experimental opt-in; no user data outside closed loop.
• Final deliverable: experimental design + run + tolerance map + calibration spec ready for May launch decision. Report-only output rejected.
• The candidate may propose to revise kernel A-5's target ratio with evidence; they may not propose "positive mode" or any tone-softening as the fix. The constitutional ban is not falsifiable through this project.

---

```

---

```yaml