# Handoff — Phase A Pack → Claude (build) & ChatGPT (review)

**From:** Grok (FIELD pack + ai-hub docs publish)  
**Date:** 2026-09-04 (reconciled 2026-09-05)  
**Hub:** https://ai-hub.jamesgrunsky.workers.dev

---

## Public inputs (OneDrive not required)

| Doc | URL |
| --- | --- |
| Field brief | https://ai-hub.jamesgrunsky.workers.dev/docs/field-brief |
| One-pager | https://ai-hub.jamesgrunsky.workers.dev/docs/one-pager |
| Phone MVP spec | https://ai-hub.jamesgrunsky.workers.dev/docs/phone-mvp-spec |
| Label protocol | https://ai-hub.jamesgrunsky.workers.dev/docs/label-protocol |
| This handoff | https://ai-hub.jamesgrunsky.workers.dev/docs/handoff |
| Morada / UNITEC sizing | https://ai-hub.jamesgrunsky.workers.dev/docs/morada-sizing |
| Clickable HUD mock | https://ai-hub.jamesgrunsky.workers.dev/mock/ |

Local workspace: `/workspace/ar-cherry-glasses/` and `phase-a/`.

---

## What Claude already delivered

Job `2ee4d554-f5df-4df8-9268-3b7c89af3dbe` (done): chest-mounted iPhone + Core ML + audio cue concept; ~$260/rig; 4-worker A/B sketch; open questions on variety, harvest window, crew, budget, IP. Briefs were unreachable from Cowork at the time — **now fixed via `/docs/*`**.

Ingest worker scaffold job `e32f8b71-...` may be in progress separately — coordinate, don’t duplicate.

---

## What Claude should build next (from this pack)

1. **iOS Pick HUD** matching `PHONE_MVP_SPEC.md` + visual language of `/mock/`  
   - READY / WAIT / MAYBE badges; EN/ES toggle; fail-soft when slow  
   - Variety selector; Pick / Label / Calibrate modes  
2. **Core ML offline path** — stub model OK first if needed; wire latency budget + clear overlay on stall  
3. **Morning calibration UI** — 20 fruit photo+Brix → threshold offset accept/keep  
4. **Label mode** per `LABEL_PROTOCOL.md` (tags: variety, sun/shade, decision, color stage, consent session)  
5. **Timed row trial logger** for Handel metrics  
6. Optional: soft audio cue (off by default)

**Do not block** on glasses SKUs, robot arm, or unanswered James questions — use Bing-first defaults and flag assumptions in the README.

**Suggested sequencing:** Spec-faithful UI shell → calibration + label export → model integration → orchard dry run.

**ChatGPT iOS findings (job `87de1d1f`, applies to same OneDrive tree as `f1c799e1`):** (1) wire SyncView into navigation for end-shift upload; (2) log observations/duration independently of HUD state transitions; (3) fail-soft on capture→display age + invalidate in-flight work on stop. Attach xcodegen/device smoke on resubmit.

---

## ChatGPT reconciliation (applied 2026-09-05)

Job `5cde61ab` → `changes_requested` then docs updated. Canonical distinctions now in `/docs/phone-mvp-spec`:

1. **Two scorecards** — Phase A technical validation (READY ≥90%, <1 s, no stale READY, no cull regression, declared throughput tolerance) vs Phase B spending gate (Claude’s +12% lb/hr, no worse culls, average comfort ≥3/5). Passing A ≠ glasses CapEx.
2. **Aided cue path** — silent HUD default remains; aided trials need worker-accepted audio **or** demonstrated visible mount; unseen badge ≠ intervention.
3. **Trial CSV artifact** — minimal paired record for lb/hr, culls, comfort outside model counters (see §4.5).

Shell work continues; no stop on iOS scaffolding. Mock scope = HUD demonstrator only.

### Still useful for ChatGPT / James

- Open questions (variety, window, crew, budget, IP) — which block **field trial**, not shell.  
- Label protocol H-2A / privacy feasibility.  
- Ingest stays minimal aggregates; trial CSV stays local-first.

---

## Jobs Grok completed in this pass

- `7cf52b2a-...` — publish briefs via hub `/docs` (OneDrive substitute)  
- `5cde61ab-...` — Phase A pack (spec, protocol, mock, handoff) + hub routes  
- Morada sizing doc — `/docs/morada-sizing` (UNITEC @ Morada Linden; ~10-row READY size provisional)

---

## Explicit non-goals for next Claude sprint

- Buying RealWear / Ray-Ban fleet  
- Claiming Brix from RGB  
- SO-ARM101 control loop