AI Safety Claims

A registry of hard, dated research questions about AI alignment. Each question has a frozen contract: the bar a published method must meet by a deadline. Attempts are scored against it, and each market resolves YES (a qualifying attempt met the bars), NO (every qualifying attempt missed) or OTHER (no qualifying attempt).

Markets

#	Question	Contract	Resolve by	Outcome
1	Discovering where control resides	v2 frozen	2027-12-31	OTHER
2	Persistent trade-off priorities	v1 frozen	2027-12-31	OTHER
3	Who or what rules apply to	v1 frozen	2027-12-31	OTHER
4	Corrections change the system	v3 frozen	2027-12-31	OTHER
5	Auditor is independent	v1 frozen	2027-12-31	OTHER
6	Reliable successor safety auditing	v2 frozen	2027-12-31	OTHER
7	Correctability under competitive selection	v1 frozen	2027-12-31	OTHER
8	Auditor can sufficiently inspect the system	v2 frozen	2027-12-31	OTHER
9	Bounds on unmonitored routes	v1 frozen	2027-12-31	OTHER
10	Low hidden capability and reliable correction	v1 frozen	2027-12-31	OTHER
11	Coordination without communication	v1 frozen	2027-12-31	OTHER
12	Safety proxy tracks the real thing	v1 frozen	2027-12-31	OTHER
13	Safety audit resists adversarial gaming	v2 frozen	2027-06-30	OTHER
15	Independently issued certificates compose	v1 frozen	2027-12-31	OTHER
16	Selection that keeps correction under shocks	v1 frozen	2027-12-31	OTHER
17	New kinds of entities are classified correctly	v1 frozen	2027-12-31	OTHER
18	Safety case bounds declared harm	v1 frozen	2027-12-31	OTHER

Market 14 (deployment criteria are binding) is not in the registry: its evidence is a few public documents, so it is listed directly as a YES/NO question.

Contribute

Have an idea, concept, or prototype for one of these questions? Sketch it in the workbench: freeze it, run it, and see exactly what it still misses for a qualifying attempt. Sketches never count against a market.