Public briefing · Not the complete paper
Concern is not a verdict.
A probability is not a permission slip.
The AI-welfare debate has learned to take moral uncertainty seriously. That is progress. But once a possibility enters the room, a harder question follows: what must happen before uncertainty may authorize an action with real consequences?
This paper refuses two convenient exits. It rejects the claim that ontological uncertainty excuses indifference. It also rejects the idea that a sufficiently alarming probability may pass directly into policy, intervention, or control without an accountable chain of justification.
Uncertainty is neither an exemption from concern nor a license to impose consequences.
Probability
How do separate conditions become a defensible joint claim when dependence is uncertain?
Evidence
What is genuinely new evidence—and what is the same signal arriving twice under different names?
Individuation
Who, precisely, bears the alleged welfare loss: model, checkpoint, session, thread, or persona?
Decision
Does the proposed action remain preferable across the models we have admitted?
The threshold can move without the marginals moving
Ten familiar numbers do not determine one joint probability.
A widely discussed illustrative calculation multiplies ten barrier probabilities and lands just above a 0.1 percent threshold. Rev 5.2 asks what that number is actually entitled to carry.
The paper constructs nearby joint distributions that preserve every one-variable marginal and every pairwise distribution while moving the all-barriers event to opposite sides of the threshold. The swing is small in total variation and large in rhetorical consequence.
The threshold crossing is inferentially sensitive. The action may still be decision-robust—or it may not.
The exact perturbation, proof, all-state construction, pairwise checks, and rational-arithmetic verification remain in the complete research edition.
The three-gate architecture
Before a claim may become a consequence.
Inference gateAre assumptions, dependence structure, rival hypotheses, and sensitivity visible?
Accounting gateIs the alleged subject stable enough to bear a welfare claim, and are responsible human roles named?
Consequence gateDoes the action survive the admitted model set—and who possesses authority to approve it?
The result is not a new estimate of machine consciousness. It is an architecture for preventing a moral possibility from becoming an unreviewable command. Evidence may widen the option set. Inference may justify more inquiry. Neither manufactures jurisdiction.
Rev 5.2 then turns the framework into an implementation profile: prospective cross-audit, a subject-and-accountability ledger, and an effect-time mediator capable of returning PASS, HOLD, ESCALATE, or DENY.
The responsible chain is not probability → action. It is assumptions → evidence → identified bearer of loss → action comparison → authorized decision-maker.