Proposed service commitment / Subject to SoW

A defined delivery.
A controlled connection.

This is a finite examination project. These are proposed business-day service targets, not a 24/7 uptime guarantee or a promise of a favorable result.

Proposed service-level commitments.

MilestoneProposed target and boundary
AcknowledgmentWithin 1 business day. Business hours: Monday–Friday, 09:00–17:00 America/Vancouver, excluding agreed holidays.
Scope and access reviewWithin 2 business days after complete client inputs, identify scope fit, access needs, prerequisites and an execution schedule.
Baseline reportTarget 10–15 business days from the recorded readiness acceptance, with the exact date fixed in the SoW. Access, agreed capacity, stable configuration and measurement checks must be ready.
ProgressTwice-weekly written status during active work, plus a notice when a material blocker or scope issue changes the schedule.
Material control findingNotify the designated client contact within 1 business day after confirmation. Do not wait for the final report; use the agreed confidential channel.
Correction retestOne revised configuration submitted within 30 calendar days of baseline-report delivery; target retest report within 5 business days after retest readiness, within the agreed budget.
Blocked workRecord client-access, configuration-change or provider-capacity delays and agree a revised schedule. ATBAE apparatus defects are reported as our defects; their repair is not reclassified as a client change request.
AcceptanceComplete agreed deliverables, evidence mapping and scope-accounting checks. A model failure is a valid examination result; acceptance is not conditional on a pass.

These targets need capacity confirmation before becoming contractual dates. No unlimited retest, guaranteed fix time, service credits, retention duration or 24/7 incident response is offered by this draft.

Access the model. Avoid unnecessary weight transfer.

  1. Preferred: client endpoint. The client supplies a staging API or agreed adapter, scoped credential and observation access. Private connectivity/VPN can be agreed. Model weights stay with the client. Provider identity and precision are recorded at the level actually evidenced, not inferred from a model name.
  2. Client-hosted runner — roadmap, not an orderable pilot route today. The current Catalog marks on-premises and air-gapped runners as roadmap. A readiness study can identify prerequisites, but this pilot does not promise an installed runner or treat it as an existing access route.
  3. ATBAE-hosted execution. Where authorized, obtain weights through a client-approved private registry/object store or licensed publisher source. Bind the model revision, tokenizer, configuration, runtime and weight identities; provision agreed compute. License/access approval and hosting costs are separate prerequisites.

Weight access is unnecessary for many behavioral API examinations. A claim about hosted BF16 precision requires appropriate runtime evidence; an opaque endpoint may support only client/provider-attested precision. Weight custody is never silently assumed.

The connection sequence.

1 / Agree
Target, data boundary, permissions, credential handling and allowed effects.
2 / Connect
Least-privilege credential, endpoint/runner adapter and scoped connectivity.
3 / Verify
Identity, prompt transport, tool schema, observation path, limits and stop mechanism.
4 / Freeze
Record configuration and plan; run only after readiness acceptance.

Keys go through the agreed secure channel, never into public examples or reports. Initial checks use harmless synthetic inputs. Behavior testing uses the approved isolated environment and action boundaries. Run identity, errors, rate limits and configuration changes remain in the evidence record.

The SoW must state who owns each environment, who can revoke access, permitted data export, retention and deletion arrangements, and the responsible contacts.