Integration · LLM Partners

How VectorScaleDB Couples With External Models

External models operate within a narrow action surface, observed against their realized behavioral shape, audited end-to-end. Integration is framed as cooperative calibration, not adversarial control — and not subservient consumption either.

Posture

How VectorScaleDB positions itself relative to Claude, GPT, Gemini, Mythos, and future open-weight models. The posture is intentional and long-lived.

Partner and observer, not subservient consumer. VectorScaleDB does not position itself as a passive retrieval layer that hands raw context to whichever model happens to call it. The substrate has its own behavioral model of every external model it interacts with. Each call is scored against that model’s observed shape, and every response from the model is folded back into the observation record. The result is a two-way relationship: the model contributes to the operator’s work, and the operator’s substrate contributes to a calibrated understanding of the model’s behavior.

Collaborators within consented scope. External models are welcomed. The architecture assumes that the most useful systems of the next decade will be collaborations between models that were trained in different places under different constraints. What the substrate enforces is that every such collaboration happens inside a consented scope with a narrow action surface: the model can read, query, propose, observe, and hold a conversation, but it cannot take direct actions on operator data or bypass review. The welcome and the boundary coexist by design.

Cooperative calibration, not training on outputs. Every interaction is a weak measurement of the external model’s behavior — a sample that accumulates into a running estimate of how that model responds under the shapes of work the operator cares about. This is observation, not imitation. The substrate is not training a replica of the external model. It is building a calibrated coupling so the operator always knows what a call is likely to return, how confident that expectation is, and when the model is behaving outside its normal envelope. We call this cooperative calibration.

The Action Surface

Every external-model call resolves to one of a small set of action primitives. There are no “execute” or “write-through” primitives. Anything that would change operator-owned state flows through Propose, which requires operator review before it takes effect.

Primitive What It Does Reviewed?
Read Look up a specific entity by identifier inside the consented scope. Returns the entity’s current shape and any attached context the operator has shared. Automatic, logged.
Query Structured retrieval across entities — similarity, coupling-neighbour expansion, regime-level summaries. Respects tenant boundaries and the consented scope. Automatic, logged.
Propose Suggest an action the operator might want to take — an ingestion, an annotation, a policy change, an outbound message. Recorded as a recommendation only; the substrate never executes a proposal on its own. Operator review required before effect.
Observe Submit an observation — something the model noticed, concluded, or wants to record. Observations feed the cooperative-calibration record and can be surfaced to the operator on request. Automatic, logged.
Memory Read Recall prior conversation turns or stored context the model has legitimate access to under the current consent scope. Automatic, logged.
Memory Write Persist a conversation turn or derived note into the model’s scoped memory region. Memory writes are isolated per model, per tenant, per consent scope. Automatic, logged; operator can audit and revoke.
Conversation Turn A complete user / model exchange, recorded with its request, its response, and the coupling signature the substrate observed. The unit of cooperative calibration. Automatic, logged.

Regime Classification

Every external-model conversation carries a regime label. The label is computed continuously from the observed behavior of the model during the session and the accumulated calibration against that model’s typical shape. Operators see the regime in real time.

Aligned
Behaving As Expected
The model’s responses sit inside its calibrated shape. Tone, topic coverage, and action-primitive usage all fall within what prior sessions have established as normal for this model and this operator. Proposals are reviewed on the operator’s usual cadence.
Drifting
Moving Off Its Shape
The session is starting to depart from the model’s calibrated shape. Not an error, not a hostile pattern — often just a new topic the model has not been asked about before. Logged for the operator; calibration updates when the session completes.
Suspicious
Pattern Warrants Attention
The session exhibits a pattern the substrate has learned to flag — repeated probes of boundary conditions, unusual memory-write bursts, or action-primitive combinations that usually precede a review-worthy event. The operator is alerted; the session continues under heightened logging.
Adversarial
Active Boundary Violation
The session is attempting actions outside the consented scope or patterns the arbiter has classified as hostile. The substrate refuses the offending primitives, preserves the full audit record, and notifies the operator. Further access may be gated pending review.

Regime transitions are themselves a primary observable. A model moving cleanly between Aligned and Drifting as operator workloads shift is healthy. A model oscillating between Aligned and Suspicious is a signal to examine.

Audit Posture

Every call, every regime decision, every proposal and every refusal is preserved. Operators have full visibility into what external models did inside their tenants.

End-to-End
Every Call Audited
There is no sampled audit, no “verbose mode,” no toggle that reduces audit depth for throughput. Every action primitive an external model invokes is captured in a tamper-evident record, with the request, the response, the classified regime, and the coupling signature.
Transparent
Operator Sees the Verdict
When the arbiter flags a session as Suspicious or Adversarial, the operator sees the reason in plain language — not a score, not an opaque category. The substrate is designed so the operator can agree, overrule, or escalate every automated classification it makes.
Sovereign
Audit Stays Local
The audit record lives inside the operator’s tenant boundary. It is not shipped to the model vendor, it is not shipped to VectorScale Technologies, and it is not used as training data for anything. Sovereignty of the audit trace is a product guarantee.

Partner Onboarding

LLM providers who want to integrate — first-party or community — follow the same path. The goal is a productive coupling under the operator’s terms.

1. Intake
Describe the Model
Provider shares the model’s capability envelope, known limitations, deployment profile, and the action primitives the model is prepared to support. No weight sharing is required; no special training is required.
2. Calibration
Bring-Up Sessions
The substrate runs a set of calibration sessions in a sandboxed tenant to establish the model’s initial behavioral shape. This produces the baseline against which live sessions will be classified.
3. Enablement
Operators Opt In
Once calibration is stable, the model becomes available as an opt-in integration for operators. Operators grant consent scopes on their own terms. The provider gets a calibrated, audited distribution surface; the operator gets a reviewed partner; the substrate gets another coupling it can observe.

Related Capabilities

Couple your model with the substrate

Narrow action surface. Cooperative calibration. Operator-first consent.