The experiments.

Each experiment begins as a claim that can fail. Results remain qualified by what was actually run, and an inconvenient finding does not disappear merely because the headline says pass.

THE SERIES SO FAR

Continuity, one replaceable layer at a time.

EXPERIMENT / 001[Qualified pass]

Work + authority

Can an agent survive changing models?

One passport preserved identity, authorized memory, permissions, provenance, and unfinished work across local, Claude, and ChatGPT runtimes.

Still openRestart recovery remains untested.

EXPERIMENT / 002[Pass]

Judgment

Does written stance survive a mouth swap?

Four written judgment rules survived a mid-task Claude-to-ChatGPT hop on novel cases, with the governing stance cited in the result stream.

Still openA low-confidence unevidenced assertion was admitted and remains an open defect.

HOW TO READ THIS LAB

A pass is not a victory screen.

Result pages contain the hypothesis, method, grades, failure boundaries, and public-safe evidence.

Friday Violations records when the public claims changed and preserves links to superseded pages.

Receipts is the claim-level ledger across the entire laboratory, including contradicted and unverified claims.