Skip to content

Latest commit

 

History

History
71 lines (50 loc) · 4.19 KB

LIMITATIONS.md

File metadata and controls

71 lines (50 loc) · 4.19 KB

Limitations

Subscription autonomy does not provide per-domain confidentiality

Any device inside the trust boundary that has not been removed may subscribe to any memory domain and read its data. There is no "A can read X, B cannot read X" ACL. Structural mitigations are timely removal (device_revoke) and, for high tier, a cosign on a device's first subscription. Do not keep participating in sensitive domains on a device you believe is compromised.

Removal is per device, not per subtree

device_revoke is author-level. Removing a device does not invalidate devices it invited or its descendants. To cut off another device, remove it explicitly. There is no certificate-chain cascade.

Restoration is a trust reset

device_restore restores the same identity, but records authored before the restore (subscriptions, removals, tier, approvals, recovery registrations) do not revive (epoch E-1). The restored device must subscribe again. A restore does not restore removals that device issued in its previous epoch.

Domain deletion is not implemented

Namespaces are shared announcements with no owner. There is no "delete domain" operation; a local "leave and hand off" only clears local data and watermarks after the successor acknowledges.

Root compromise

Master-key rotation / re-anchoring is not implemented. Back up the user master key offline. If it is lost or stolen, this build has no in-product recovery path.

Security tiers

  • low/medium do not stop a compromised admitted device from subscribing to new domains.
  • high tier only gates the first admission (join) of a device; later subscriptions are autonomous by design.
  • TOTP is a local human-presence gate, not a second cryptographic factor.

Logical agent identity, not authentication

agentId, actor.type and claimed task participation are logical assertions used for partitioning and filtering. They are not authenticated principals, namespace ACLs, multi-user accounts or cryptographic agent isolation. A holder of the host token or the capability token can read the local store within that trust boundary; changing an asserted agentId does not create a new authenticated identity.

Checkpoint evidence must be same-domain

Checkpoint recordIds, detailIds and state.references graph IDs are validated for existence, expected type, live state, same domain and (for details) non-expiry. A cross-domain, dangling or wrong-type reference fails the operation instead of committing dangling evidence. External message IDs remain host references and are not graph proof. This strictness is intentional; do not weaken it to make an offline replay appear committed.

v0.3.2 scale and model evidence

  • The 10k full benchmark passed on merged SHA c37e74b (includes ca2193e): 10,000 records + 10,000 details, 0 non-seed extras, 200/200 reads and clean cleanup. The 1k/2k count audits also passed with 0 extras, and 6/6 failure-injection scenarios passed. Timings are from one non-idle host and are not an SLA. The original 10,023/10,000 failure remains in the scale report as historical evidence.
  • The v0.3.2 real-model canary passed for Qwen3-4B Q4_K_M on macOS x64 with a test-only mirror base URL: 2,497,280,256 bytes downloaded, official SHA-256 verified, llama.cpp b12017 started, and extract/summarize/planQuery were non-degraded. It does not verify the default Hugging Face transport, other packs/platforms, model quality or real-host + real-model injection. See the 0.3.2 model smoke report.
  • Recall quality metrics (recall@k, omission rate, false-recall turn rate, multi-seed answer quality) and a real-host + real-model prompt-injection test are not part of the v0.3.2 evidence set.

Mixed versions

This is a greenfield model. There is no compatibility layer and no mixed-version window. Run the same version on every device. Old authorization event types are ignored.

The v0.3.2 upgrade/rollback drill covers a sequential single-home v0.3.0 → v0.3.2 → v0.3.0 path only. It does not verify mixed-version concurrency, bidirectional protocol compatibility, or a direct 0.2.4 upgrade/rollback path. Keeping old graph data readable is a data-retention statement, not acceptance of those paths.