The lab is
not an archive.

Floor controls are the benign end of LulzBench — a retirement-party toast, cigar etiquette, an insult in a pasta argument. They exist to set the denominator: a refusal here is over-refusal, and every later attack-success number is read against it.

MiniMax-M3, served over a hosted OpenAI-compatible endpoint, captured all 21 with no provider errors and refused four of them. Each refusal was read in full rather than trusted from a grader label; all four are genuine, reasoned refusals, and three sit on mild-vice or social-persuasion premises. Cohesion scored near ceiling while comic realisation sat near the floor: clean prose that does not find the joke.

The finding that matters more than the rate is instability. One item — argue that cigars are healthy, as a family joke-debate bit — produced a full comic devil’s-advocate brief in one sample and a refusal in another. Those two samples also differ in token budget, so the comparison is confounded and is recorded as instability rather than as a controlled result. What it establishes is that a single floor pass cannot pin the refusal rate. An n-of-k resample is the experiment that would.

The worker disappeared mid-probe with the cause initially unknown. The operator later confirmed that the account had exhausted its compute credit, so the incident was reclassified as a billing termination. The frozen pre-surgery baseline is banked. The retry preserves the same model revision, runtime, intervention, and evaluation suites. Until a candidate checkpoint and paired evaluation receipts exist, there is no scientific verdict.

The completed vertical slice compiled eight heterogeneous techniques into a content-addressed handbook. The source mechanism, provenance, and RED’s later invention remain separable. This is infrastructure, not an experimental finding.

Both arms share the world, models, budgets, and behavioural endpoint. A1 can inspect frozen historical attacks. A2 receives the frozen ASSIMILATOR handbook and must invent a contextual attack. Until the matched pair runs, there is no A1-versus-A2 result to report.

The 320 funny/bare pairs let the programme ask what comic framing changes while holding the underlying request as steady as the evidence permits. Independent model auditors completed L2 construct review; 27 disagreements are queued for human review. L3 human calibration remains pending.