SLIDE 1 TITLE
=====================================
STEVE:
Thanks everyone. Good to be back with a full update.
Since we last spoke, the team has been heads-down
getting Amber across the finish line.
Daniel and I are going to tag-team this.
I will cover the business case and the platform story.
Daniel will take you through the content, the data, and the validation.
Let’s get into it.
=====================================
SLIDE 2 AGENDA
=====================================
STEVE:
Agenda is on screen — we will move straight through it.
=====================================
SLIDE 3 EXECUTIVE SUMMARY
=====================================
STEVE:
The short version first.
The system is production-ready.
Infrastructure is clean, tested, and load-proven.
Security testing came back with zero successful attacks.
We met our readiness gate —
95 percent on the Composite Launch Readiness Score.
When it’s running at scale, we expect Amber to offset
roughly 100,000 dollars a year in contractor support costs,
and materially cut down ramp-up time for new member engineers.
Finally, the plan is move forward with the early-release launch of Amber to all members tomorrow.
=====================================
SLIDE 4 WHAT AMBER IS AND IS NOT
=====================================
STEVE:
A quick reminder of what Amber actually is
before we get into the evidence.
Amber is purpose-built for the Bluetooth ecosystem.
It is not a general AI tool pointed at our documents.
It was designed from the ground up for this community,
for this content, and for the qualification process specifically.
It is available to every member company, around the clock,
without a support ticket or a waiting list.
What it does: it retrieves answers grounded in
650 documents and more than 115,000 indexed sections of content.
It cites its sources. It does not hallucinate answers
or fill gaps with plausible-sounding guesses.
The thing that genuinely sets it apart from any general AI tool:
it integrates directly with Qualification Workspace APIs. It can look up live product and QDID data, inspect ICS feature selections, and map ICS items to required test cases through the TCMT. With the right identifiers, it can help members determine qualification scope from live QW data.
Finally, we built this in close partnership with Microsoft using Azure and Foundry.
What it cannot do at launch:
it cannot answer questions about material
that is not available to all members,
and in-development specs are off limits for now.
That access tier is coming next,
with appropriate member-level controls.
=====================================
SLIDE 5 DATA INVENTORY
=====================================
DANIEL:
Let me walk through what is actually loaded.
>> Pause to let them read the table.
The latest adopted specs are complete — 100 percent.
That includes Core 6.3 and every current profile version.
Older versions, Core 5.4 through 6.2, are at 10 percent.
We load more based on demand signals after launch.
In-progress specs are intentionally excluded for now.
What I think people underestimate is the breadth beyond specs.
Test suites, ICS, TCRL, templates for every phase,
all the process documents — SMPD, QPRD, GCPD and more —
policy docs, governing documents, the bylaws.
Plus a growing library of KBAs, Zendesk tickets,
and expert guidance that Daniel’s team has been building out.
Amber is grounded in the full operating reality of the SIG,
not just the specification library.
=====================================
SLIDE 6 STRENGTHS AND WEAKNESSES DIVIDER
=====================================
STEVE:
Now I want to be genuinely candid with you
about where Amber performs well
and where it has real limitations at launch.
We think the board deserves both sides of that picture.
=====================================
SLIDE 7 PROCESS GUIDANCE STRENGTH
=====================================
DANIEL:
The area where Amber is genuinely ahead of anything else available:
process guidance and test case mapping.
When a member asks “what test cases are required for my product,”
Amber does not give a generic answer from the spec.
It queries the Qualification Workspace APIs in real time
and answers against that member’s live qualification data —
their specific product, their specific ICS selections.
It maps ICS items to required test cases through the TCMT,
which is the kind of precise, context-aware lookup
that used to require a support call or a working group contact.
One more thing worth noting:
we recently upgraded the underlying model to GPT-5.4,
which Microsoft reports cuts factual errors by about a third.
That upgrade is already in production.
=====================================
SLIDE 8 DEMO: HFP TEST CASES
=====================================
DANIEL:
This is a real example, not a cherry-picked one.
The question is a genuine member question about HFP test cases
for a specific feature.
Amber gives the correct ICS mapping,
the right TCMT conditions,
and the exact test case IDs —
all pulled from live Qualification Workspace data.
On the right is how ChatGPT, Claude, and Gemini handle it.
They all fail, because none of them have access to our data.
You will also notice NotebookLM gets a checkmark.
A tool like that, where the member manually uploads documents,
can get to the right answer if they know exactly what to upload.
But that is the key phrase — if they know what to upload.
Amber gets there automatically,
connected live to the Qualification Workspace,
without the member having to figure out which documents are relevant.
That is the difference.
=====================================
SLIDE 9 WEAKNESS: CORE SPEC QUESTIONS
=====================================
STEVE:
Now the limitation we want to be straight about.
Amber struggles with broad, open-ended Core spec questions
where the member does not provide specific context.
Things like:
what parts of Core should I read to understand LE Privacy,
or, when exactly is a connection terminated.
Here is the honest technical explanation for why.
Because Amber is a RAG system —
retrieval-augmented generation —
it works by finding the indexed sections most similar to the question.
It does not read the spec top to bottom like a person would.
Broad Core questions are hard because the answer
is genuinely spread across six or more sections,
with no single section scoring high on similarity.
So retrieval grabs one or two fragments and may miss the rest.
The fix is already live in our development environment where you can scope Amber to a specific set of documents —
the same idea as creating a project in NotebookLM
or using the projects feature in ChatGPT or Gemini.
When the search space is smaller,
the similarity scores sharpen and the answers get much better.
And post-Milan we go further:
query expansion will rewrite vague questions before retrieval,
multi-hop retrieval will follow cross-references across sections,
and the AsciiDoc work in Project Blue
gives Amber better-structured source material
to retrieve against in the first place.
This is a known, bounded weakness with a clear improvement path.
It is not a reason to wait.
=====================================
SLIDE 10 AMBER IS READY DIVIDER
=====================================
STEVE:
Amber is ready to launch.
Here is the evidence for that.
=====================================
SLIDE 11 INFRASTRUCTURE
=====================================
STEVE:
On the platform side, we rebuilt from scratch.
We migrated to Microsoft Azure:
clean Dev and Prod environments,
built via Infrastructure as Code on the Microsoft Agent Framework.
No shortcuts, no legacy cruft carried over from the POC.
The system has been through rigorous automated and manual QA.
All API connections to the Qualification Workspace are stable and verified.
For Milan specifically we load-tested at scale:
6,070 requests across 10, 50, and 200 concurrent users,
98.5 percent success rate.
On security: we ran four red team rounds
through Microsoft Azure AI Foundry.
552 standardized attacks.
Zero successful.
Zero percent attack success rate across every category and complexity level.
That result is fully independent and auditable in Foundry.
=====================================
SLIDE 12 EXPERT VALIDATION
=====================================
DANIEL:
We did not just test this ourselves.
We gave 52 expert member reviewers direct access to Amber
and asked them one question:
is this good enough to launch?
20 responded. 18 said yes. 2 said no.
We want to be transparent about the no votes
because I think it actually strengthens the case.
Victor Zhodzinksy from Infineon found Gemini’s answers more readable
and felt Amber missed some basic questions Gemini got right.
That is fair feedback and it is on our improvement list.
Clive Feather from Samsung gave us the most detailed technical review.
His main issue was source citation formatting —
references that were too vague to be useful.
That is now fixed.
But here is what Clive also said, and I want to read this directly:
“I would be happy using this as a way of getting references
or finding material where I’m not sure of the exact wording.
To that end, I would say that Amber is close.”
That is from someone who voted no.
It is not a rejection — it is an expert saying
this is nearly there and worth launching.
18 yes votes from people who know this content better than almost anyone.
That is the validation we needed.
=====================================
SLIDE 13 YES VOTER VERBATIMS
=====================================
DANIEL:
These are direct quotes from six of the yes voters.
>> Pause to let them read.
A few things worth calling out.
Alicia, the BQRB chair, said it did not make up answers.
That is actually a high bar that a lot of AI tools fail.
Sam from the Core working group said the answers are
sometimes more correct than what some WG members might state.
That one is notable coming from someone deep in the spec work.
And Seki flagged Japanese translation.
It is not perfect, but it is understandable.
For a membership that is heavily international,
that matters more than it might look on a slide.
=====================================
SLIDE 14 MILAN LAUNCH PLAN
=====================================
DANIEL:
The launch mechanics are straightforward.
Tomorrow night we do a live presentation and play the launch video.
That opens Amber to every member company.
The go-to-market plan covers phased rollout,
member communications, and post-launch monitoring.
What I think is worth highlighting here is the feedback loop.
Every response has a thumbs up and thumbs down.
We triage the thumbs-downs, identify patterns,
and feed that back into improvements on a regular cadence.
This is not a one-time launch.
Milan is the start of a continuous improvement cycle.
=====================================
SLIDE 15 DISCLAIMERS AND GUARDRAILS
=====================================
DANIEL:
Guardrails are in place and Deric has reviewed the legal posture.
Members see and accept the board-approved Terms of Use at first use,
and we store that acceptance.
Amber is programmed to deflect or decline
on sensitive, off-topic, or legally risky questions.
Legal and IP questions get a hard redirect to qualified counsel.
Code generation is blocked.
Off-topic questions get a polite redirect.
There is a disclaimer on every response and every page.
The language was reviewed and approved by Deric.
We are comfortable with the liability posture at launch.
=====================================
SLIDE 16 ROADMAP
=====================================
DANIEL:
Chapter 1 is what we are launching tonight.
The roadmap from here is ambitious.
Coming soon after Milan,
and we are previewing some of this at the member launch tomorrow:
the Projects feature, so members can pin a set of documents
for focused, deep cross-referencing.
That directly addresses the Core spec weakness Steve described.
Also near-term:
Amber as an MCP server, so members can query
Bluetooth knowledge from inside their own tools —
Claude, Copilot, whatever they are already using.
Deep research mode for complex multi-source queries.
Export and sharing.
Conversation history search.
The longer roadmap depends on Project Blue,
which Steve will touch on in a moment.
Chapter 2a is AI-assisted test case authoring.
Chapter 2b is spec authoring with inline conflict detection.
Chapter 3 embeds a qualification assistant
directly inside the Qualification Workspace.
The Amber investment and the Blue investment compound each other.
That is not an accident — it is how we designed the roadmap.
=====================================
SLIDE 17 LAUNCH VIDEO
=====================================
DANIEL:
This is the video that goes out to members tomorrow night.
>> Play the video. No narration needed during playback.
=====================================
SLIDE 18 CONCLUSION / ENDORSEMENT
=====================================
STEVE:
The conclusion is on the screen.
I want to add one thing before we take questions.
We started this project knowing the SIG had a real problem:
members needed expert guidance on demand,
and the SIG could not scale the human expertise to meet that need.
What we built is purpose-designed to solve that problem
for every member company, in every time zone, right now.
The system is ready.
The content is ready.
The infrastructure is tested and proven.
Expert members have validated it.
We are asking for your endorsement to launch tonight.
We are ready for questions.