Choose the owner
These source entry points are selected by the platform’s immutable component
lock. A source import or a unit test does not establish package publication,
native deployment compatibility or hosted qualification.
Start with a metered handle
Declare a catalog once withdefineLanguageModels, install it on the language
models component, and pass typed handles to consumers. A handle contains model
identity and capabilities; it never contains a reusable provider credential.
Credentials remain deployment configuration or native gateway custody.
defineAiTask combines an input schema, an output schema, instructions and a
context builder. It opens a metered session before the builder reads protected
records. Each retained intent has one provider attempt, with SDK retries
disabled. Results distinguish completed output, invalid input/output, refusal,
failure, insufficient evidence, uncertainty and cancellation. An uncertain
attempt retains its reservation and must not be replayed to guess whether the
provider ran.
defineManagedAgent adds a declared tool allowlist and bounded steps, tool
calls, prompt bytes and output bytes. It checks the native session around tool
work. Hitting a step or output limit is a refused outcome, never a completed
answer. Tools execute through the application’s current authenticated context.
Publish knowledge from canonical sources
defineKnowledgeCollection embeds a bounded source population into a namespace
isolated by collection and generation. The host owns accepted indexing work and
publication. Its publication mutation must recheck the original work, source
revisions and previous pointer before accepting ready index entries.
Retrieval uses vector results only to select candidates. It rereads canonical
sources and their current revisions before returning text and citations, then
rechecks the published generation. Cached text cannot bypass a revoked or
changed source. Citations identify collection, source, revision, generation and
chunk. Different chunks or revisions remain distinct in the workspace.
Review business changes through native operations
The assistant’s operation tool creates drafts for a person to submit through the normal operation dialog. Model output does not grant permission or execute a business change. The actual operation rechecks current data and its review rule. Documents use the same boundary: an external worker claims an existing native attempt and reads files through scoped native delivery. Parsed pages and model values retain source evidence.requireResult checks the exact definition,
current source selection, unresolved issues and required review before a
business operation can use the result.
The agent workspace exposes host-owned original work, immutable definition
identity, model attempts and verified native request links. Cancellation fences
late content writes through /storage; settlement can still resolve an already
dispatched attempt. Recovery reads original retained evidence.
The knowledge /workspace and /application adapters expose source coverage,
current publication, bounded refresh history and explicit metered previews.
Inspection performs no inference. Current-source checks still apply to previews.
Measure original executions
defineAiEvaluation evaluates explicit case/candidate/repetition slots against
original outputs, tool traces, citations, latency, cost and optional judgments.
Declarations pin target/configuration, case-set revision, baseline, evaluator
versions, repetitions, populations and finite cost/deadline limits. Missing or
unknown slots remain in the denominator. Duplicate execution IDs cannot count as
independent repetitions. Partial cost stays partial, with exact picodollar
values retained.
defineAiJudge from /ai/judge runs a pinned rubric through a metered typed
task. Its original model-produced judgment has its own uncertainty and
accounting. It grants no approval.
The evaluation workspace adapters expose versioned definitions and retained
reports. aiLifecycleMeasurements converts a report into metric/table fragments
for the existing model-lifecycle owner, using an explicit selection or held-out
validation population. Aggregate cost/deadline status remains explicit. It
creates neither a native report identity nor promotion approval. Run new
inference as an explicit admitted intent; evaluating stored evidence does not
silently run a model.