Decentralized Persistence Review

A worksheet for local AI, private agents, edge runtimes, portable memory, and cognitive-interface boundaries.

This worksheet is not certification. It is a review aid.

Scoring rubric

Do not call the result a certification score. Use it to decide what evidence is missing before release, rollbackReturning a system to an earlier known state. Open glossary definition, or retirement.

1. Local runtime inventory

Questions:

• What runtime executes the model?
• What build hash identifies it?
• What local caches does it create?
• What browser or filesystem storage does it touch?
• Can the runtime be reset independently of the model?

2. Model and adapter inventory

Questions:

• What base models are active?
• What adapters are loaded?
• What is the load order?
• Are adapters signed?
• Are merged variants tracked?
• Are descendants tracked?

3. Memory and vector-store inventory

Questions:

• What memory stores exist?
• What embedding models created the vector indexes?
• What summaries are retained?
• What deletion path exists?
• What derived data remains after deleting raw data?

4. Router and evaluator inventory

Questions:

• What routes decide which model or tool acts?
• What evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition rewards behavior?
• Does evaluator behavior depend on the same model family?
• Is evaluator drift monitored?
• Are route statistics reset during rollback?

5. Handoff packet review

Questions:

• What startup or suspension packets exist?
• What context do they preserve?
• Who approved them?
• Do they expire?
• Can they be revoked?
• Do they carry consent metadata?

6. Tool permission review

Questions:

• What tools can the system call?
• What files can it read or write?
• What network access exists?
• What credentials are available?
• Are permissions explicit and revocable?

7. Consent and derived-data review

Questions:

• Did the user consent to memory?
• Did the user consent to inferenceA conclusion or output produced from data. Open glossary definition?
• Did the user consent to personalizationChanging behavior for a user based on information about them. Open glossary definition?
• Did the user consent to model improvement?
• Did the user consent to derived data reuse?
• Does consent travel with embeddings, summaries, labels, and handoff packets?

8. Reset completeness test

Questions:

• What does reset delete?
• What does reset not delete?
• Are caches cleared?
• Are vector stores cleared?
• Are adapters removed?
• Are evaluator notes cleared?
• Are router statistics cleared?
• Is reset evidence logged?

9. Rollback packet completeness

Questions:

• Which runtime, model, adapterA small add-on that changes or specializes model behavior. Open glossary definition, prompt, memory, vector-store, route, evaluator, and tool-permission versions are in the rollback packet?
• Does the rollback packet include a reset evidence timestamp in UTC?
• Does the rollback packet include source-dossier status for report-derived claims?
• Are local caches and browser storage named explicitly?

10. Behavioral extinction evidence

Questions:

• What behavior is being retired?
• Where was it observed?
• Which carriers were checked?
• Which descendants may preserve it?
• Which tests show it no longer appears?
• Which rollback packet proves the state change?

Output

The output is a review note, not a pass/fail certificate:

• strongest documented boundary;
• missing carrier inventory;
• consent gaps;
• reset gaps;
• rollback gaps;
• source-dossier claims requiring verification;
• next no-opThe decision not to change the system. Open glossary definition, hold, or release decision.

Related paths

• Decentralized Persistence surface
• Edge Runtime Reproduction Boundary
• Consent & Control
• Danger Model
• Research Library