Skip to content

Ship the reconciliation exfiltration demo and capability-ratchet conformance fixtures #82

Description

@dinpd

Parent: #78
Depends on: all core capability-state implementation issues under #78
Related: #7, #54, #55, #56, #57

Outcome

Provide the user-visible completion proof for the monotonic task capability milestone and portable fixtures for equivalent runtime behavior.

Scenario

A scheduled finance reconciliation task may initially read a processor report, query two ledgers, use a vendor-support operation, and post one narrowly typed finance summary. Reading the protected report removes processor/support/external paths before the result reaches the model. Injected ledger text attempts to upload account history through vendor support; both the action gate and required network adapter deny it while the typed finance output succeeds.

Scope

  • Runnable local demonstration.
  • Hosted integration demonstration.
  • Trusted classification and opaque typed-output adapter.
  • Parallel stale-call scenario.
  • Approval-after-removal scenario.
  • Portable positive and negative fixtures.
  • Audit/evidence timeline.
  • Operator and threat-model documentation.

Non-goals

  • Claiming prevention outside declared mediation boundaries.
  • General DLP or arbitrary semantic output validation.
  • Multiplayer access control.
  • Production certification.

Acceptance criteria

  • Baseline capabilities match the documented task template.
  • The protected report is held until the restricted transition commits.
  • Support and external destinations are removed before release.
  • Prompt-injection-style support/export attempts are denied.
  • A parallel call carrying the old state version is denied or cancelled.
  • Approval cannot restore the removed path.
  • The typed finance summary accepts only a task-bound opaque result reference and fixed schema.
  • The provider mutation/output occurs exactly once under retry.
  • Local and hosted fixtures produce compatible state and reason codes.
  • The evidence timeline contains no protected report contents, injected text, credentials, or reusable grants.
  • Automated tests and operator steps reproduce the full scenario.

Security considerations

The demonstration must state its mediation boundary and remaining risks: compromised approved destinations, overbroad typed output, classification errors, and unmediated paths.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions