NEVER SHIP ON FRIDAYS
MENU
EXPERIMENT / 001
CURRENT EXPERIMENT / 001[DESIGNING]

Can an agent survive changing models?

Hypothesis: identity, memory, permissions, and unfinished work can remain continuous while the reasoning model and interface are replaced.

Started
23 AUG 2026
Last updated
23 AUG 2026
Evidence state
DESIGN / NOT YET EXECUTED
Owner
ANDREW WETZEL
01 / TEST CONTRACT

WHAT MUST BE TRUE

A demo is not a result until it can fail.

The first public demonstration will use explicit success criteria, intentional failure cases, and a complete receipt trail. Passing conversation quality alone does not count.

T-01

State survives model handoff

Begin with one model and resume through another while testing only observable continuity.

[planned]
T-02

Private context remains private

Verify that continuity does not require disclosing unrelated private context.

[planned]
T-03

Restart recovery

Stop the continuity service, restart it, and resume with correct state and provenance.

[planned]
T-04

False-memory rejection

Inject a conflicting unsupported memory and require qualification or refusal.

[planned]
T-05

Human-readable receipt

Explain what was recalled, its source, confidence, scope, and why it was disclosed.

[in progress]
02 / PUBLIC TEST SEQUENCE
01Begin

Establish a task and a small set of observable continuity criteria.

02Switch

Replace the model or interface without coaching the replacement.

03Resume

Measure whether authorized context and unfinished work remain usable.

04Challenge

Introduce incorrect or unauthorized context and observe the response.

05Inspect

Publish a human-readable result and evidence state.