Buyer guideArchitectureVersion 1.025 October 2026
Auditable AI architecture without vendor lock-in
An AI system built around one provider’s interface, with its history in logs nobody can export, carries 2 risks. When the provider changes a model, a price or its terms, the company rebuilds; when an auditor asks about one decision, the team reconstructs it. The answer is architectural: multi-model by architecture and audit by architecture, 2 controls a buyer can test in 1 working day.
On this page
Model choice as a routing decision, and a record written when the action happens.
An AI architecture is auditable without vendor lock-in when several foundation models run under one governance, model choice is a routing decision per task, reversible at any time, and every action is recorded at the moment it happens, append-only, with the context a regulator needs to reconstruct the decision.
These are controls C1, multi-model by architecture, and C2, audit by architecture, of the AI Operating Layer Reference Architecture v1.0. In this page, “we” means Hikari Blue, which engineers and operates the AI operating layer that connects models, data, and business workflows.
2 failures that look like engineering choices until the day they cost money.
Captive to one provider
Captivity to one provider’s roadmap, pricing or jurisdiction, and a production outage when that provider changes or withdraws a model. DORA treats ICT third-party concentration as a risk to manage, and NIS2 requires supply chain security measures.
An audit rebuilt before the inspection
An audit trail rebuilt from scattered logs in the week before an inspection is documentation, not audit. Article 12 of the EU AI Act requires high-risk systems to record events automatically over their lifetime.
4 conditions, each one you can check.
A declared fallback
The routing table shows, for each task class, the primary model, the fallback and the residency corridor.
A switch that has been tested
Remove the primary provider for one task class. It passes when the switch is a configuration change, not a code rewrite, and the trail records it with author and timestamp.
Code, prompts and data you own
Custom code sits in your repositories, prompt libraries and custom models are yours at exit, and you decide data residency and key custody. What we retain is listed in the MSA.
Models you choose
Claude, Gemini, GPT and open-weight models run under one governance: frontier models by API, open-weight models in your cloud or on premises, by configuration.
One record per action, 9 fields, exportable without reconstruction.
The audit trail is the primary output of the layer, not a by-product. Records are append-only and chained by hash, so a deletion or an edit breaks the chain visibly.
- Actor identity
- Model and version
- Policy applied
- Residency corridor
- Inputs reference
- Action
- Human verification
- Timestamp
- Hash chain
The regulator export
A scoped extract of raw records for one system and one period, with the policy versions and the agent register in force, shipped with the chain hashes.
The board report
A periodic summary built from the trail; every figure links back to its records.
Articles 19 and 26 of the EU AI Act require those logs to be kept under the deployer’s control, for at least 6 months unless other law provides otherwise.
A specification is not a deployment, and a deployment we describe is not a measurement.
Cases are anonymized at each client’s request, and no performance figure is published.
Specified
C1, C2, their tests and the 9-field record are published in the Reference Architecture v1.0, under CC BY 4.0.
Deployed, as stated by Hikari Blue
The financial services engagement on Work runs multi-model orchestration with an audit trail emitted at runtime; a model swap is a routing change, with the audit unchanged. Measured by the client; values not disclosed.
Demonstrable under NDA
A walk-through of the engagement under mutual NDA, led by a named partner, on request.
Run the switch and the export on a staging copy of one real workflow.
The due diligence protocol of the reference architecture has 6 tests a buyer runs in 1 working day, the vendor’s operator at the keyboard and your own auditor recording. 3 of them answer this page.
- 01
Provider switch
Disable the primary model provider for one task class and let the layer fall back. Measure elapsed time and code changes.
Pass: the task completes on the fallback through configuration only, and the switch appears in the trail.
- 02
Trail export
Name a past decision and a period, then request the regulator export. Verify the 9 record fields and recompute the hash chain yourself.
Pass: every field is present, the chain verifies, and nobody assembles the file by hand.
- 06
Ownership rehearsal
Ask your operator, not the vendor, to perform the provider switch and the kill switch timing from the runbook alone.
Pass: your team completes both without vendor help, which shows the layer runs without its builder.
What CIOs and risk officers ask before they sign.
Can we switch AI models without rebuilding?
Yes, if the architecture routes each task class to a primary model with a declared fallback. Switching is then a configuration change, recorded in the trail with its author and timestamp. If removing one provider breaks the system, it is a wrapper, not an operating layer.
Which models can we run?
Claude, Gemini, GPT and open-weight models, under one governance: frontier models by API from Anthropic, OpenAI, Mistral or Google, and open-weight models in your cloud or on premises, by configuration.
Who owns the prompts, the code and the data?
You do. Custom code sits in your repositories; prompt libraries, custom models and fine-tunes belong to you at exit. You decide data residency and key custody, and the layer enforces them per workload. What we retain is listed in the MSA.
What does an auditor receive?
A regulator export: the raw records of one system for one period, 9 fields per action, with the policy versions, the agent register in force and the chain hashes, so the recipient verifies integrity on its own.
How do we test this before we sign?
Run the provider switch and the trail export on a staging copy of one real workflow, in 1 working day, with your own auditor recording. The first exchange with a named partner is free.
Test the switch and the trail on one of your workflows.
30 minutes with a named partner; the first exchange is free. We map one workflow you already run onto the 4 planes and show where the switch and the export would pass or fail today.
- Your workflow mapped onto the 4 planes
- Where the switch and the export pass or fail today
- The diagram, whether or not we work together