Status: Durable task/run/event and usage evidence exists for the live
personal partner preview. Mature receipts, public traces, datasets,
experiments, model comparison, and routing-learning governance are
unavailable.
Current durable records
Praxa’s task substrate stores internal task and task-event records required for
durability, billing authority, retries, and recovery. The public API projects a
strict customer-safe subset:
- run identity, lifecycle status, safe timestamps, and links;
- sequenced durable lifecycle events;
- verified text only when stored result, action digest, and verification
evidence pass the projector’s checks;
- generic customer-safe failure information;
- tenant-scoped usage derived from completed provider and settled retail rows.
It does not expose prompts, private reasoning, raw tool arguments, provider
payloads, credentials, executor leases, internal billing references, provider
cost, or margin.
Exact browser-action approval projection is source-ready pending production
activation. It exposes only the validated instruction/steps, target host,
approval reference, and matching action digest.
Retention
No public retention window for runs, events, webhook deliveries, or future
receipts is committed in this documentation. Do not infer one from database
existence. Contractual retention/deletion, backup, legal hold, and regional
residency require an explicit policy before enterprise procurement.
Behavior-corpus pseudonymization
Praxa’s implemented platform-corpus normalizer replaces common timestamp, IPv4,
email, UUID, key-shaped-token, and hostname patterns with stable safe values
inside one normalization batch. Semantic comparison can replace streamed text
with a length marker; exact comparison retains scrubbed text.
This is a testing safeguard, not a universal de-identification claim. Pattern
scrubbing cannot prove that every retained free-form string is anonymous, and a
normalized behavior corpus is not the same product as a customer dataset or
experiment workspace.
Learning and training
Per-tenant routing learning, cross-tenant aggregated metadata, opt-out controls,
consent/training labels, and evaluation-gated learned-policy updates are design
directions, not live public execution guarantees. Current model routing does
not establish those data-governance properties by implication.
Before activation, Praxa must document and verify:
- exact source fields and retention;
- tenant partitioning and deletion;
- aggregation thresholds and privacy review;
- opt-out authority and enforcement;
- training/diagnostic consent semantics from first write;
- evaluation, approval, rollout, and rollback for learned changes.
See also
Last modified on August 14, 2026
