What decision you are buying
A qualification engagement starts from the decision it must support: publish a claim,
commit to a capacity envelope, buy hardware, ship a release, deploy offline. Scope, workload,
versions and pass conditions are fixed before the first run, so the result is an evidence-backed
answer to that decision — whichever way it goes.
Negative results are first-class deliverables. A verdict that a claim does not reproduce, or
that a system fails its SLO at the intended load, is delivered with the same rigor as a pass:
localized cause, reproduction recipe, and safe wording for what the system actually does.
The fee is the same in both cases.
Services
-
Vendor Claim Evidence Pack
Is one specific public claim about a system true under stated conditions?
- Scope
- One exact claim, a bounded set of qualification cells, frozen versions and workload, agreed pass/fail conditions before any run.
- Deliverable
- Evidence-backed verdict with raw artifacts, limitations, and reproduction recipe.
-
Workload Capacity Qualification
What load does this exact system sustain under our workload and SLO?
- Scope
- Concurrency sweep and goodput measurement against a frozen workload trace with a customer-defined SLO; failures kept in the denominator.
- Deliverable
- Operating envelope: recommended range, pass-with-limits range, SLO-fail boundary, known failure modes.
-
Runtime/Model Portability Sprint
Can our exact runtime/model be made to work on Strix Halo, GB10, RTX or Apple Silicon — and what does it take?
- Scope
- Starts with a fixed-scope Portability Diagnostic; enablement work is a separate, clearly labeled engineering engagement.
- Deliverable
- Diagnosis with exact failure points, then (optionally) patches/flags/build recipes labeled as commissioned engineering.
-
Long-Context Reliability Qualification
How much context is actually useful on this system — not just how much fits in memory?
- Scope
- Depth ladder (8K→128K) with quality gates at each level, distinguishing configured / loadable / completed / quality-preserving / recommended context.
- Deliverable
- Useful-context verdict per depth with quality evidence, latency profile, and failure documentation.
-
Procurement Decision Pack
Which of 2–4 candidate systems should we buy for our workload?
- Scope
- Same frozen workload and SLO across candidates; comparison limited to what the comparison class allows.
- Deliverable
- Side-by-side evidence with explicit limits, not a universal score.
-
Air-Gapped Readiness Check
Will this stack install and operate in an isolated network segment?
- Scope
- Technical controls only: offline install, egress behavior, secrets, backup/restore, rollback, logging. Not a compliance opinion.
- Deliverable
- Verified technical-control checklist with evidence per control.
Engagement modes
Every engagement runs in exactly one mode, agreed before the test plan is written. The mode
determines who controls the conclusions and how the result may be presented. Funding is
disclosed in every public report regardless of mode.
Independent commissioned test
The client funds the test; AGmind owns methodology, execution and conclusions.
- Does: pre-registered test plan, frozen versions, verdict with limitations.
- Does: disclose funding in any public report.
- Does not: allow the client to edit conclusions or methodology.
- Does not: include debugging or tuning of the system under test.
Commissioned engineering
AGmind works toward the client's goal: bring-up, patches, build recipes, tuning.
- Does: deliver working recipes with before/after evidence.
- Does: label every output as commissioned engineering.
- Does not: produce an independent verdict about the same system.
- Does not: get mixed with independent test results in one report.
Reference configuration
● planned
A maintained, versioned stack release for an exact SKU, with an acceptance suite and a
revalidation policy.
- Does: separate validation cells from enablement work in the release notes.
- Does: disclose the commercial relationship wherever results appear.
- Does not: exist as a ready product — it is demand-gated (see below).
- Does not: turn engineering output into an independence claim.
Process
-
Step 1
Scope call
We establish the decision behind the engagement, the exact claim or workload, who owns the
budget and the technical review, which versions are already frozen, and where the hardware
is. If the decision itself is not yet clear, discovery comes before a statement of work.
-
Step 2
Test plan approval
A pre-registered test plan fixes pass conditions, workload revision, SLO and quality floor
before any run. The first part of the fee is tied to test-plan approval and reservation of
the lab window.
-
Step 3
Frozen cells
Scope is expressed in qualification cells: system revision × runtime revision × model
artifact × workload revision × operating setting. Any material change after the freeze —
another BIOS, driver, container digest, model file, context bucket or backend — is a change
request, never a silent update.
-
Step 4
Execution
Runs follow the approved plan. Failures, timeouts and errors stay in the record; repeats
and endurance runs happen where the plan calls for them. Timelines pause when access,
hardware or artifacts are blocked on the client side.
-
Step 5
Factual review round
The client receives the draft and can correct factual errors — a wrong version string, a
mislabeled artifact. The client cannot edit conclusions, verdicts or methodology. One
factual review round is included in scope.
-
Step 6
Delivery
The agreed deliverables are handed over: manifests, raw and derived data, verdict,
limitations, reproduction recipe. The second part of the fee is tied to delivery. The fee
never depends on the verdict.
Exclusions
- No 24/7 operations or incident SLA. Qualification
is a bounded engagement, not an operations contract.
- No penetration testing. Security testing outside
the agreed technical controls is out of scope.
- No certification. AGmind never presents results
in certification wording. Deliverables are evidence bundles with explicit scope and
limitations.
- No hosting of client inference. AGmind qualifies
systems; it does not operate them in production for clients.
- No compliance opinions. Air-gapped readiness
checks cover technical controls only; a report is never a legal compliance statement.
- No pay-for-positive. An engagement that requires
a guaranteed favorable verdict is declined before it starts.
Disclosure options
Public
The client keeps the full private evidence bundle; AGmind publishes a redacted report with
a funding disclosure. This is the baseline option and the one that builds the public
research record.
Embargoed
The client holds the private bundle until an agreed date; publication follows after it.
Useful around launches and procurement deadlines.
Private
Only the client receives the results; AGmind may state that the work happened only if
permitted. Private-only work is priced above public work because it creates no public
research asset — the difference is stated in the proposal.
Demand-gated future products
These offerings are built only when paid demand proves the need. They are listed here so the
gate is public — they are not ready products and are not sold today.
-
AGmind Reference Stack
● planned Built when two independent prospects request it and at least one pays for setup.
-
Release Revalidation Channel
● planned Offered when clients return to re-check after BIOS/kernel/driver/runtime/model changes.
-
BenchOps
● planned Private run history and regression tracking — only after repeated paid validations create the need.
Pricing
Pricing is scoped per engagement from the frozen cell count, the required repeats and
endurance runs, and the chosen disclosure option. It is shared in the proposal, together
with the payment split and the change-request terms. Shipping, insurance, customs,
commercial licenses and cloud costs are billed separately at cost.
Start a scope conversation
Bring the decision you need to make, the exact claim or workload, and where the hardware
lives. If versions are not frozen yet, that is fine — freezing them is the first step of the
engagement.