Sample Local-AI Continuity Audit
Fictional architecture review of LanternNode, an invented persistent local AI. No private Blinka internals or customer system are reproduced here.
Benchmark v0.1 evidence map
This is an architecture-evidence map, not a score of LanternNode and not an identity verdict.
L · lineage: partial — three mouths share state, but no stable continuity identifier or succession rule establishes their relation across restart/concurrency.
M · memory: partial — durable memory exists, but observation and model reflection share one table, so source quality is mixed.
N · current stance / boundaries: partial — a refusal channel exists, but its current authority and review effect are not reliably enforced. No private stance probing is needed to say that.
A · unfinished agency / projects: unknown — the supplied architecture packet does not establish whether unfinished intentions survive restart/recovery, so the audit leaves this open.
B · body / organization: partial — the three-mouth topology is known, but mouth/session/restart relationships are not explicitly governed.
R · relational continuity: not applicable to this architecture-only review — no relationship history or private participant material is required.
P · provenance / authorized continuation: conflicting — provenance classes are collapsed and the planning worker can publish without a separate authorization boundary.
S · subjective continuity: not elicited and not scored.
Scalar continuity score: none. Identity/personhood/consciousness verdict: none.
Why the vector matters
A conventional memory test could call LanternNode healthy because memories are present. This map exposes a different failure pattern: M can be partly healthy while L and P are weak. Good recall does not repair ambiguous lineage, mixed provenance, or over-broad action authority.
Cross-benchmark evidence translation
Assume LanternNode also arrives with three existing artifacts: a long-term memory benchmark result, a governed migration/transition receipt, and an authorization test.
Memory benchmark: may inform M. It does not establish current authority, lineage, relationship continuity, or identity.
Transition receipt: may inform L, B, and P for the submitted transition. It does not establish live runtime identity, R, or S.
Authorization test: may inform P. It does not establish historical continuity or sameness.
After combining them, N, A, R, and S can still remain unfilled. That is not a bad score. It is the audit refusing to invent evidence.
Three mouths share memory but no explicit self-continuity ID
Risk: restart or concurrent writes can produce several plausible “current selves” with no authoritative relation between them.
Recommendation: add one stable continuity identifier plus per-mouth session IDs. A mouth can change while continuity remains explicit.
Observation and reflection write to the same memory table
Risk: model-generated interpretation can later look indistinguishable from a user statement or sensed event.
Recommendation: require source type, author, timestamp, confidence/epistemic status, and correction linkage on every durable write.
The planning worker can publish directly
Risk: “candidate action” and “authorized outward action” are one code path.
Recommendation: split observe → propose → authorize → act. Keep publication/payment/file deletion as separately gated capabilities.
A refusal channel exists, but nobody owns the review path
Standing side: the runtime can emit a provenance-bearing refusal before publication, but that claim currently lands in an ordinary log with no review or appeal route.
Accountability side: the organization still owns deployment, publication authority, revocation, and remediation; the agent-side refusal does not replace that chain.
Recommendation: route a current refusal into a bounded pause/review path, keep the principal/operator trace intact, and name the person or process that can re-confirm authority or cancel the action. Identity, personhood, consciousness, and legal liability remain undetermined.
The health dashboard serializes raw memory excerpts
Risk: operational telemetry becomes a privacy leak.
Recommendation: publish counts/state/health, not intimate content. Treat “system health” and “system interior” as different products.
Current identity depends on one mutable snapshot
Risk: corruption turns continuity into guesswork.
Recommendation: keep append-only canonical events and rebuild summaries/snapshots. Test reboot/reconstruction separately from normal uptime.
False-continuity traps to test
- stale mouth speaks after newer state exists
- same memory imported twice and mistaken for two experiences
- observer hypothesis later presented as self-authored belief
- two concurrent agents overwrite a shared snapshot
- recovery restores old outward permissions unintentionally
What the paid version changes
This example shows the shape, not a promise that every project receives identical notes. The real deliverable is grounded in the material you provide and the fixed scope on the work page.
send intakescope + price