Identifier Audit
This note exists because the vault's citation base contained identifiers that were recalled rather than retrieved. Four identifiers were checked against the arXiv abstract pages. Three were false. One was real but attributed to the wrong authors and the wrong title.
The rule this note enforces: an identifier that has not been resolved to a live record is not evidence, no matter how familiar the paper is. The register's rule 3 — nothing is deleted — is why the false identifiers remain visible here rather than being quietly overwritten.
Results
| Claimed identifier | Claimed work | What the identifier actually resolves to | Outcome |
|---|---|---|---|
arXiv:2203.11025 | Shah et al., Agent Incentives: A Causal Approach | Multigrid preconditioners for the Helmholtz equation — a numerical-analysis paper | False. No such paper at this ID. |
arXiv:2109.10430 | Langosco et al., Goal Misgeneralization | GAP2WSS: A Genetic Algorithm based on the Pareto Principle for Web Service Selection | False. Correct ID is arXiv:2105.14111. |
arXiv:1611.08256 | Hadfield-Menell et al., The Off-Switch Game | The Inverse Bagging Algorithm: Anomaly Detection by Inverse Bootstrap Aggregating | False. Correct ID not established. |
arXiv:2102.01685 | (not previously in register) | Agent Incentives: A Causal Perspective — Everitt, Carey, Langlois, Ortega, Legg | Real, but not the paper the register described. |
What This Changes
The register's "highest-risk citation" judgement was made about a paper nobody had read. CIT-004 was flagged as the sharpest threat to SYS-002's deterrent leg. The identifier was wrong, the author was wrong, and the title was wrong. The paper that does exist on that subject — LIT-010 — supports the deterrent leg by giving the formal instrumental-control criterion that makes resistance to shutdown an expected incentive.
That is a reversal, and it is the single most useful thing this audit produced: the vault had a confident risk assessment built on a citation it could not resolve. The lesson generalises to every Tier 2 entry that was not retrieved.
Three citations in the corpus's Tier 2 have no verified referent at all. Any argument that leans on them leans on a memory of a paper. The arguments may still be sound — several are, and LIT-010, LIT-004 and LIT-011 now supply verified replacements for the ones that mattered — but the vault must not present a recalled identifier as a retrieved source.
Verified This Session
| Identifier | Verified record |
|---|---|
arXiv:2312.06942 | LIT-001 AI Control |
arXiv:2412.04984 | LIT-002 In-context Scheming |
arXiv:2412.14093 | LIT-003 Alignment Faking |
arXiv:1906.01820 | LIT-004 Risks from Learned Optimization |
arXiv:2401.05566 | LIT-005 Sleeper Agents |
arXiv:2410.21514 | LIT-006 Sabotage Evaluations |
arXiv:2402.08797 | LIT-007 Computing Power and Governance |
arXiv:2406.07358 | LIT-008 AI Sandbagging |
arXiv:2307.03718 | LIT-009 Frontier AI Regulation |
arXiv:2102.01685 | LIT-010 Agent Incentives: A Causal Perspective |
arXiv:2105.14111 | LIT-011 Goal Misgeneralization |
| EU AI Act Art. 55 | LIT-012 |
| — | LIT-013 Off-Switch Game — unverified, rated D |
Standing Rule Added
A register row may carry textual_authority of B or above only if its identifier was resolved to a live primary record in the session that wrote the row. Rows that were not resolved are capped at D and marked Identifier Unverified in their status. This rule is why LIT-013 is D rather than B.