Lifecycle and dependencies
Protect uses theimpact.protect/1 protocol at the documented provider and text/JSON boundaries. Runtime
execution, Decision delivery and readable Product evidence are separate outcomes. Verify enforcement and
classifier quality through actual protected work.
Async applications use protect = await runtime.activate_protect(version_id=...) after
await impact.init_async(). The returned async handle has synchronous protect.status(),
await protect.flush(timeout_ms=...) and await protect.close(). Synchronous activation returns a
ProtectHandle with local status(), bounded flush(timeout_ms=...) and close(). Both views close through
the same runtime-owned lifecycle, so closing a handle and then shutting down the runtime does not repeat
completed cleanup. A failed close reports closing and leaves cleanup retryable. Reactivate after the prior
close succeeds.
await runtime.shutdown() also closes its owned Protect activation. Cancellation preserves the caller’s
CancelledError and closes an unclaimed late activation. An owner acquired by another live caller stays
active.
Protect activation is explicit. Ordinary imports and init() do not load models, initialize ONNX Runtime,
download weights or change provider decisions. Base impact leaves native inference engines
application-owned. Install impact[protect] for the selected convenience tuple: ONNX Runtime 1.31.1,
tokenizers 1.0.0rc3 and NumPy 2.5.4 on Python 3.12+, or NumPy 2.4.6 on Python 3.11. Applications retaining
compatible ONNX Runtime 1.30.0/tokenizers 0.23.2 install base impact instead. On non-Windows systems set
ORT_DISABLE_TELEMETRY=1 before any ONNX Runtime initialization. The lazy loader refuses initialization
without it and never changes customer environment or OpenTelemetry. Setting it after another module
initialized the runtime is not proof of the original opt-out. Windows ETW and other native behavior on Windows
remain unassessed. See
ONNX Runtime privacy.
Applications can explicitly install the ONNX Runtime 1.32.0.dev20261106001 preview with
tokenizers==1.0.0rc3 and base impact. Select the exact host wheel and SHA-256 from the
official nightly index
as an application dependency. The Protect extra uses the ordinary PyPI release. Install the nightly from the
official nightly index. The observed native CPU tuple uses macOS 14+ on ARM64. Other nightly hosts remain
unassessed.
Activate Protect
Enable the Product key’s Protect grant and explicitly activate the initialized SDK:/ws/sentinel with Bearer
authority and X-Impact-Purpose: protect. Use trusted TLS even for a loopback origin. The controller
automatically reconnects and re-registers its same runtime identity, selected Version and label when supplied.
Unchanged configuration is acknowledged again without reloading models. Connection, delivery, application,
health and useful protected work remain distinct observations.
Select the Environment and Version
A customer-runner Environment can be selected without changing Presence or Trace export’s base Environment:version_id is the canonical Builder whole-material Version UUID, distinct from the reported
context version_id / IMPACT_VERSION_ID reference and callable selectors. The version-specific sample reads
that actual configured UUID from IMPACT_PROTECT_VERSION_ID and passes it explicitly. The example reads the
variable in application code. Likewise, IMPACT_PROTECT_ENVIRONMENT_ID supplies the actual authorized
Environment UUID for the optional override. The server resolves the selected formal Version’s digest.
Caller-supplied version_digest values are unsupported and refuse before discovery or activation. Generic
init/deployment labels are not substituted for version_id. Omitted Version input does not qualify a
Version-specific runtime. Labels are trimmed valid text of 1–120 UTF-16 units without controls. Equivalent
activation options share one registration and conflicting options refuse. The runtime also exposes
invoke_protect_provider_test(boundary, invoke) for the existing authorized Product enforcement-command
journey. The synchronous runtime invokes a synchronous application callback once. The async runtime awaits the
async callback on its caller loop. This helper preserves the command’s scoped identity and requires the actual
selected provider boundary to apply a Decision. Policy authority and Product acceptance remain with the
platform.
Classifier data and configuration
report_scores defaults to False. Opt-in HTTP Decisions include each genuinely checked finite 0–1 score,
including zero. Unavailable checks omit the score. Empty text has no evaluated chunk and follows the captured
unavailable posture, without inventing a measured zero. Monitor and allow-and-record preserve the native empty
result. Enforce applies its configured unavailable blocking. The runtime retains raw Policy threshold/PCA and
the pinned Arctic XS/S/M/L pipeline-2 catalogue. Immutable classifier references are acquired with same-origin
purpose authentication and size/hash/media compatibility checks. Models are acquired lazily only for an
applied Policy. Set model_directory or IMPACT_PROTECT_MODEL_DIRECTORY for the cache. Optional model_ids
restrict delivered models. Existing files can be supplied without an acquisition requirement:
configSequence/configRevision and Policy/artifact tuple: selected
adoption changes that runtime while unselected runtimes retain their original basis until their own adoption.
Off removes applied on Policies without acquiring a new model. An expiry renewal keeps the same sequence and
revision. A higher sequence may carry the same revision when its configuration body matches. Stale sequences
and changed material under a retained revision are refused. Failure, expiry and partial state remain visible.
Enforce follows the captured unavailable posture and blocks unless allow_and_record. Monitor preserves
native work and records unavailable checks. Late original stream Decisions keep their captured configuration
rather than being recast against the newest material.
Provider boundaries
The observed provider boundary is native OpenAI Chat Completions. Input checks precede the native provider request. Output enforcement uses the result of that call. Off and Monitor preserve configured client retries, native results, metadata, errors, lazy streams and caller cancellation. Existing sync/async wrappers retain the actual worker, provider read, body and iterator close until they join. Verify native inference termination separately from Python worker cancellation. A pending or failed cleanup remains an error. Pending unary Monitor and explicitallow_and_record calls finish normally during shutdown, with captured runtime_closed
unavailable Decisions. Shutdown retains their real join. Required Enforce/block still fences release. Customer
clients remain caller-owned. Other provider operation sets remain unassessed.
For Anthropic typed parsing, a text replacement makes the message/block parsed_output unavailable (None),
including the final helper snapshot. Impact does not validate the replacement into a new application model or
rerun user validators/schema hooks. Off, Monitor and unchanged output retain the original native parsed value
and identity. An enforced helper snapshot is unavailable until its output has been checked.
Python retains at most 32 active native sends, reads and Protection workers. At that limit, Monitor and
explicit allow_and_record preserve the application’s next native call and report
provider_retention_unavailable against its captured Policy configuration, without a measured score or a
safe-check claim. An unretained call reports its output check unavailable before the native send. Its eventual
body remains application-owned and uninspected. Required Enforce/block refuses unavailable send admission
before the provider call. Failed read handoff after headers closes that response before consumer release.
Shutdown joins retained work within its deadline and does not claim that unretained application work
terminated. The protocol requires a digest even when that future output is unknown. Such output reports use
the exact unavailableBasis value
provider_retention_unavailable:output_body_unavailable:proposed_digest_placeholder with unavailable
check/effect, no score and the required SHA-256 of an internal empty placeholder. Treat the body as unknown
under this marker.
Decision delivery
Protect sends Decisions from a bounded outbox to the purpose-authenticated HTTP/v1/protect/decisions route.
Retries keep the same Decision occurrence and body without repeating a provider call or rescoring.
protect.flush(timeout_ms=6000) and shutdown make bounded delivery attempts. Status reports pending,
failed-attempt and lost counts. The reports carry digests, lengths, provenance and actual effects. Proposed or
released plaintext stays out of the reports. protect_test and protect_test_result require a genuine
enforcement callback, otherwise the test is unavailable. Process exit may lose unacknowledged in-memory
Decisions. Confirm admission and readback in Product independently of local flush.
Linear JSON artifacts use binary64 coefficients and PCA output, left-to-right dot-product accumulation, and
bias added after summation. Both SDKs use the same bounded sigmoid reduction so the same embedding, artifact
and threshold produce the same linear score and decision. Native encoder and ONNX execution retain their own
precision and platform requirements. Verify classifier quality separately.