Copilot

3 changes since Monday

Context

⏎ send · ⇧⏎ newline · ⌘K palette · scripted answers

F24 Copilot rail Proven

What this copilot cannot see

Answered before you ask. The same boundary as the integrity panel, at the point of use.

Private messagingWhatsApp, WeChat groups and Moments, LINE. Closed, and not ours to read.

Client book and allocationWaitlists and allocations live in your systems. Public talk is folklore, not inventory truth.

Counterfeit incidenceWe measure the conversation about fakes. We cannot measure how many exist.

What is actually running

Be precise. The model can be real. What it reasons over depends on the domain.

ScriptedCanned sample. Engine not connected. The page opens and runs with no network.

Live model · qwen-plusReal reasoning, real SSE tokens. Luxury still reasons over synthetic data. Anti-scam runs real RAG over 77,727 posts.

CostEvery live question spends tokens on the lab’s account. max_tokens 700. No call fires on load — only when you ask.

PoC-11 · PoC-18The chat endpoint is unauthenticated. RAG returns retrieval only — the join to generation happens in this page.

Recent threads

Attach context

More than one chip is cross-panel reasoning.

Scope

What the copilot can see right now.

MarketsHong Kong · Macau · Taiwan — two audiences, not three markets.

WindowLast selected window versus the prior period of equal length.

Channels in scopePublic platforms only. Instagram and Xiaohongshu at useful volume are still to be collected.

Locked outPrivate messaging · client book · counterfeit incidence.

Keyboard

`⌘`/`Ctrl` + `K`	Command palette
`?`	This overlay
`Esc`	Close overlay / stop stream
`[` `]`	Collapse / expand copilot
`g` then `a` / `l`	Anti-scam / Luxury domain
`g` then `i` / `p`	Intelligence / Platform surface
`g` then `b` / `d`	Aliases for Intelligence / Platform
`/`	Focus composer
`n`	New thread
`t`	Cycle theme
`r`	Toggle review mode (client mode is the default)
LIVE / SNAPSHOT / SYNTHETIC	Data provenance · default never LIVE
SCRIPTED / LIVE MODEL	Copilot engine · default never live · no call on load

Two domains. One platform.

This is not a brand dashboard with a second tab. It is a domain-configurable intelligence platform. The pipeline is the same. The corpus is not.

Luxury brand · illustrative

Five maisons, compared

The headline of this demonstration: Cartier read against Tiffany & Co., Bvlgari, Van Cleef & Arpels and Chaumet across Hong Kong, Macau and Taiwan. Product names and design codes are real; every voice, sentiment and campaign figure is illustrative, because no luxury corpus has been collected yet — PoC-01.

Anti-scam · real

462 cases · 77,727 posts

Live tenant, live embeddings, live BFF. The copilot runs real RAG, then qwen-plus, over the public-source corpus. Numbers on the canvas match the 14 August 2026 snapshot exactly.

The argument is the contrast: the pipeline you see working on the scam corpus is the pipeline a brand corpus would run through. Luxury is illustrative only because collection has not started.