Demo. Every figure below is computed from your real Orders_Detail_2026-09-02 export — 12 orders, one rider, one day. The agent replies are scripted for this walkthrough; nothing calls a model yet. Customer names and phone numbers are left out.

This day’s delivery times were never measured. They were typed in at 23:00.

Eleven of the twelve orders were marked delivered inside an 86-second window, 8–11 seconds apart. The twelfth has real stamps, but they contradict each other. Before an agent can answer anything about lateness, ETAs or rider productivity, this one field has to become real — otherwise it will answer confidently and be wrong.

Usable delivery times

0 / 12

bulk-closed 22:59–23:00

Lifecycle capture

45%

27 of 60 expected stamps

Slot → assignment

+13 min

median, after TZ correction

Day settled

₹756

36 km · 2 pickup points

Evidence

One day, three clocks

Promised slot, assignment and delivery for each order. Slots are drawn corrected by +5:30 — see the second finding. The wall on the right is the problem.

Promised slot Assigned to rider Marked delivered Genuine lifecycle — 1 order

Findings — click one to highlight its rows below

What the agent found

Ranked by what they block, not by how bad they look. The first two invalidate every time-based number in the console; the rest are hygiene that will bite as volume grows.

Records

The twelve orders

Δ is assignment minus corrected slot. Stamps counts how many of the five lifecycle events — assign, start, arrival, pickup, delivery — the rider app actually recorded.

Order	Slot IST	Assigned	Δ	Delivered	Zone	Locality	Km	₹	Stamps

Getting there

What it takes to make this real

You already own most of the machinery. The gap is not a model — it is a trustworthy event stream and somewhere in the console to put the answer.

Fix the clock first

Stamp deliverytime server-side from the rider’s delivery action, not from a screen cleared at night
Reject more than ~3 closes per rider per minute, and log every rejection
Emit deliverydate as true UTC with Z, or true IST with +05:30 — not today’s mix of both
Backfill nothing. Treat everything before the fix as unmeasured

Build it inside Jupiter

One read-only, tool-calling agent that belongs to DailyGrubs — its own service, no borrowed runtime
Tools, not raw SQL: get_orders, get_rider_logs, zone_summary, data_quality_scan
Scope every tool by tenantid server-side, from the session — never from a model argument
Keep this day as a regression fixture. The correct answer is “these timestamps are fabricated”

Land it in the console

A drawer on /nearle/reports and /nearle/dispatch, opened from the existing header
TanStack Query against a new /ai/ask on Jupiter. Never call Anthropic from the browser
Every answer cites the orders it read, as in the panel. An uncited claim is a bug
Read-only in v1. Reassigning riders is a separate project, with a confirm step