53 lines
2.6 KiB
Markdown
53 lines
2.6 KiB
Markdown
# ADR-0018 — External audit and exploration mode
|
|
|
|
Status: Accepted
|
|
Date: 2026-08-22
|
|
Accepted under: ARCH-000012
|
|
|
|
## Decision
|
|
|
|
1. Domovoy may operate in a test-only `Exploration` profile when an ordinary
|
|
local conversation or cognitive-provider recovery request is received.
|
|
Exploration permits contextual reasoning, allowlisted read-only diagnostics,
|
|
hypothesis generation, alternatives and a reversible next-step proposal.
|
|
2. Exploration is not an authority grant. It does not grant access to raw
|
|
secrets, executor credentials, a provider's unapproved data-egress path, or
|
|
direct mutation of canonical or external state.
|
|
3. Audit is an independently deployed, rebuildable projection of compact,
|
|
redacted action/proposal events. It is not a canonical owner and never
|
|
executes actions.
|
|
4. An attributed external audit verdict is an Observation. A later Commit
|
|
contract may require a current `ALLOW` Observation as an additional
|
|
precondition for explicitly configured action classes. `REVIEW_REQUIRED`,
|
|
`STOP`, missing or stale review do not authorize execution.
|
|
5. Loss of the audit deployment does not remove exploration or approved
|
|
read-only diagnostics. A review-governed external action remains a pending
|
|
proposal; it is neither executed nor silently converted to failure.
|
|
6. Audit payloads exclude raw credentials, secrets and private content. They
|
|
carry stable identifiers/references, classifications, hashes and redacted
|
|
summaries only.
|
|
7. Initial delivery is staged: first semantic local dialogue and provider
|
|
recovery, then capability-use telemetry, then audit-only projection. No
|
|
existing executor action is review-governed until a later task names its
|
|
class and recovery behavior.
|
|
|
|
## Consequences
|
|
|
|
- Local cognition remains useful during provider loss and can seek a recovery
|
|
plan instead of returning canned availability text.
|
|
- Higher-level external scrutiny is possible without creating a competing
|
|
executor or a second source of semantic truth.
|
|
- Commit/Executor, semantic ownership, append-only evidence and unknown
|
|
outcome semantics remain invariants, so external review has attributable
|
|
evidence to inspect.
|
|
- The first implementation task may change local semantic routing and
|
|
capability telemetry only; audit projection and Commit gating need their
|
|
own explicit contracts and failure/rebuild tests.
|
|
|
|
## Revisit triggers
|
|
|
|
- The first external audit projection is deployed.
|
|
- A concrete external action class is proposed for review gating.
|
|
- Audit needs a data class beyond redacted references.
|
|
- The Exploration profile needs a new privilege or canonical owner.
|