Skip to content

Reference

Evaluation and capability evidence

A route, response, or passing software test does not establish that a learned capability is operating. This page explains what the current Mpalo API evidence supports, how the deterministic mock is evaluated, and what must be demonstrated before a model-dependent capability is described as available.

Current status.

The hosted customer path is a production-shaped deterministic mock. Its evidence covers backend transport, authentication, managed text persistence, lexical retrieval, stored-text rendering, usage, test-environment billing, export, deletion, and operational monitoring. It does not establish live Palo Bloom execution, customer-facing embeddings, semantic vector search, learned surprise, traversal, mapping, or personalization.

What current tests establish

The current test layers answer different questions. A green result is meaningful only within the scope of the layer that produced it.

Evidence layerWhat it can establishWhat it cannot establish
Contract testsResponse shapes, status codes, errors, headers, idempotency semantics, and protected-route behavior.That a model produced the result or that a configured provider is reachable.
Runtime testsTenant-scoped managed persistence, deterministic lexical ranking, stored-value rendering, export, deletion, usage, billing records, and anomaly calculations.Semantic quality, learned retention, embedding quality, or production-scale capacity.
Client parity testsHTTP, Python SDK, CLI, JSON, authentication, usage, billing, infrastructure, and idempotency surfaces preserve shared meanings.That every future client or provider has the same behavior without its own evidence.
Deployment verificationWhether the public target answers the documented health, documentation, OpenAPI, and unauthenticated rejection checks.That the target has live model readiness, universal deletion, or production billing.
Learned capability evaluationWhether a named model-dependent behavior improves against fixed controls on a defined fixture.Whether the behavior is safe or suitable for every application, modality, or physical system.

The capability registry is the authoritative status boundary. The implementation status, provider, model execution flag, billing mode, and limitations must agree before a claim is promoted.

Evaluation method for learned memory

A learned memory result is useful only when it can be compared with simpler explanations. The first evaluation should use a frozen, held-out fixture and the same reader and context budget across conditions.

ConditionPurpose
Lexical baselineMeasures how much the current deterministic text path can achieve without learned representations.
Random or shuffled controlTests whether the reported result depends on meaningful retrieval rather than accidental ordering.
Untrained or ablated conditionSeparates architecture and pipeline effects from learned behavior.
Simpler learned baselineTests whether the proposed mechanism improves over a less complex representation or retriever.
Candidate Palo Bloom conditionMeasures the named representation, retention, retrieval, rendering, traversal, or mapping behavior under the same constraints.
Quality
Report retrieval and rendering outcomes against the fixture, including no-data, conflict, stale, and correction cases.
Cost
Report write and read units, latency, storage, provider charges, and any asynchronous work separately.
Isolation
Test concurrent tenants, users, namespaces, keys, and storage connections for cross-scope leakage.
Lifecycle
Test retries, correction, supersession, export, deletion, provider failure, and version transitions.

Capability promotion gate

A capability moves beyond mock, contract-stage, control-plane-only, or research language only when the evidence names the exact implementation and deployment that produced it.

Required evidencePromotion question
Versioned contractAre the input, output, representation, model variant, dimensions, provenance, and error states explicit?
Execution truthDid the named provider and model execute, and can the path prove that it did not silently fall back to mock?
Data and authorityAre tenant, account, application, actor, subject, consent, provider, and authorization boundaries preserved?
Lifecycle behaviorDo correction, conflict, replay, deletion, export, timeout, rollback, and partial failure have tested meanings?
Measured operationAre latency, cost, capacity, billing, retention, and operational records independently attributable?
Reproducible qualityDoes the frozen fixture show a defined improvement over controls, with negative results retained?
Review and releaseHave security, privacy, legal, product, and deployment owners accepted the capability boundary?
No inference from configuration.

A model name, storage connection, API key field, charge row, or successful HTTP response does not prove that the corresponding provider, model, embedding, or learned stage executed.

How evidence is reported

Every published status should be reproducible by a reader who knows the deployment, contract, fixture, and date. Reports should preserve failures and uncertainty instead of turning incomplete evidence into a green capability label.

npm run check:api-docs
npm run check:infrastructure-boundary
npm run verify:deployment

# The deployment verifier checks reachability and public boundaries.
# It does not claim Palo Bloom readiness.

For the current runtime and its limitations, read Infrastructure. For data and model boundaries, read Data boundaries and provenance. For provider execution requirements, read Storage and providers.