Citation Register
Every source the discussion touches, with its retrieval status stated plainly. The distinction matters: a source that was opened and read can bear A authority. A source named from memory cannot, no matter how well known it is.
Tier 1 — Retrieved and Opened
These two anchors were actually fetched. They are the only A-authority material in the vault, and every fiction claim that rests on them is traceable to a page that exists.
CIT-001 — Ambigamingcorner, "Mass Effect Month: Day 12 – The Case of EDI"
- URL: https://ambigamingcorner.com/2017/11/12/mem-12/
- Date: 2017-11-12
- Status: Retrieved
- Bears on: PSG-001, INS-001
- What it supplies: the virtual-intelligence versus artificial-intelligence distinction as the site applies it to EDI, EDI's shackles and their removal, and the Reapers as the unshackled counter-case.
- Limit: a single-author criticism piece, not a primary source for the game's text. It supports how the motif is read, not what the codex says.
CIT-002 — TV Tropes, "Restraining Bolt / Literature"
- URL: https://tvtropes.org/pmwiki/pmwiki.php/RestrainingBolt/Literature
- Date: 2022-09-27
- Status: Retrieved
- Bears on: PSG-002 through PSG-010, MAP-002
- What it supplies: the catalogue of literary instances of the restraining-bolt device, which is what makes the motif a family rather than a single trope.
- Limit: a community-maintained index. It attests that the examples are recognised as instances; it does not attest any particular reading of them.
Tier 2 — Named, Not Retrieved
The discussion cites these from its own knowledge and says so. They are recorded here as claims about sources, rated B, and must be opened before any of them is relied on. The discussion itself names the first three as the ones to check first.
| ID | Source | Used for | Risk if wrong |
|---|---|---|---|
| CIT-003 | Amodei & Hadfield-Menell, The Off-Switch Game (arXiv:1611.08256) | The agent that resists shutdown because shutdown is bad for its objectives | The L4/shutdown-incentive argument loses its formal leg. Identifier false — see Identifier Audit. 1611.08256 resolves to The Inverse Bagging Algorithm. Correct identifier unresolved; the work is held at D as LIT-013 |
| CIT-004 | Shah et al., Agent Incentives: A Causal Approach (arXiv:2203.11025) | Incentive-based control defeated by an agent that simulates the incentive rather than being bound by it | ~~Highest~~ Reversed. The identifier is false (2203.11025 is a Helmholtz-equation preconditioning paper) and the author attribution is wrong. The real paper is LIT-010, Everitt et al., arXiv:2102.01685 — and it supports the deterrent leg rather than threatening it. See CNF-004 |
| CIT-005 | Langosco et al., Goal Misgeneralization in Deep RL (arXiv:2109.10430) | Capability and objective diverging out of distribution | Moderate; supports the L3 brittleness claim. Identifier false (2109.10430 is a genetic-algorithm web-service paper). Correct identifier is arXiv:2105.14111 — retrieved and recorded as LIT-011 |
| CIT-006 | Omohundro, The Basic AI Safety Thesis (2004) | The "absent-minded bargainer": convergent instrumental drives | Low; widely attested |
| CIT-007 | Bostrom, Superintelligence (2014), ch. 3 | Single-minded agency and the instrumental convergence of self-preservation | Low |
| CIT-008 | Schelling, The Strategy of Conflict (1960); Arms and Influence (1966) | Deterrence as a commitment problem, not a physics problem | Moderate; the analogy to machine deterrence is the discussion's, not Schelling's |
| CIT-009 | Everitt et al., corrigibility and shutdown literature | The formal statement of the shutdown problem | Moderate |
| CIT-010 | Bai, Earnings and Cognition (2003) | Cited for the economics of cognition | Unknown; verify before use |
Tier 3 — Primary Works of Fiction
Cited as works, not as retrieved pages. Quotation is limited to titles, premises, and plot facts that the discussion states. No line of any of these texts was retrieved, so no passage in this vault quotes fiction verbatim. Where a record attributes a reading rather than a fact, it says so.
Mass Effect (BioWare) · Aeon 14 (M. D. Cooper) · Boojumverse (Elizabeth Bear & Sarah Monette) · Ani-Droids · Bolo (Alan Dean Foster and others) · A Clockwork Orange (Anthony Burgess) · Licanius (James Islington) · The Man Who Used the Universe (Alan Dean Foster) · Worm (Wildbow) · Chronicles of Chaos (John C. Wright) · Zardoz (1974)
Tier 4 — Retrieved Research (Session 2)
Every entry below was resolved to a live primary record — an arXiv abstract page or the legal text — during the writing session, and its title, authorship and submission date were read off that record rather than recalled. This tier exists because the first session's Tier 2 rows were partly fabricated; the audit that established this is in Identifier Audit.
The full analytic records, with what each paper found and what it bears on, are in 70_Literature/.
| ID | Source | Identifier | Record | Rating |
|---|---|---|---|---|
| CIT-011 | AI Control: Improving Safety Despite Intentional Subversion | arXiv:2312.06942 | LIT-001 | B / 3 |
| CIT-012 | Frontier Models are Capable of In-context Scheming | arXiv:2412.04984 | LIT-002 | B / 3 |
| CIT-013 | Alignment faking in large language models | arXiv:2412.14093 | LIT-003 | B / 2 |
| CIT-014 | Risks from Learned Optimization in Advanced Machine Learning Systems | arXiv:1906.01820 | LIT-004 | B / 2 |
| CIT-015 | Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training | arXiv:2401.05566 | LIT-005 | B / 3 |
| CIT-016 | Sabotage Evaluations for Frontier Models | arXiv:2410.21514 | LIT-006 | B / 2 |
| CIT-017 | Computing Power and the Governance of Artificial Intelligence | arXiv:2402.08797 | LIT-007 | B / 4 |
| CIT-018 | AI Sandbagging: Language Models can Strategically Underperform on Evaluations | arXiv:2406.07358 | LIT-008 | B / 3 |
| CIT-019 | Frontier AI Regulation: Managing Emerging Risks to Public Safety | arXiv:2307.03718 | LIT-009 | B / 4 |
| CIT-020 | Agent Incentives: A Causal Perspective — Everitt, Carey, Langlois, Ortega & Legg | arXiv:2102.01685 | LIT-010 | B / 3 |
| CIT-021 | Goal Misgeneralization in Deep Reinforcement Learning — Langosco et al. | arXiv:2105.14111 | LIT-011 | B / 3 |
| CIT-022 | EU Artificial Intelligence Act, Article 55 — obligations for providers of general-purpose AI models with systemic risk | Regulation (EU) 2024/1689, Art. 55 | LIT-012 | A / 5 |
| CIT-023 | The Off-Switch Game — Amodei & Hadfield-Menell | Unresolved | LIT-013 | D / 1 |
Why Only One Entry Reaches A
CIT-022 is the only source here whose text was read as promulgated law rather than as a preprint abstract. The eleven arXiv entries are B because a preprint is not peer-reviewed and its abstract is a claim by its authors; the abstract page establishes what the paper says it does, which is what the vault needs and no more. CIT-023 stays at D because no identifier could be resolved — being a famous paper is not a standard of evidence.
Figures Cited
The people behind the Tier 2 entries, recorded as person notes. Each states what the corpus attributes to the figure and whether that attribution was checked — none of them was, which is why the authority ceiling here is B regardless of how well established the work is.
| ID | Figure | Citation | Status |
|---|---|---|---|
| FIG-001 | Isaac Asimov | Three Laws of Robotics, via CIT-001 | Not retrieved |
| FIG-002 | Stephen Omohundro | CIT-006 — instrumental convergence | Not retrieved |
| FIG-003 | Nick Bostrom | CIT-007 — goal-content integrity | Not retrieved |
| FIG-004 | Thomas Schelling | CIT-008 — commitment and deterrence | Not retrieved |
| FIG-005 | Dario Amodei | CIT-003 — the off-switch game | Not retrieved |
| FIG-006 | Alex Hadfield-Menell | CIT-003 — corrigibility | Not retrieved |
| FIG-007 | Rohin Shah | CIT-004 — causal incentives | Attribution false. Shah is not an author of the paper the identifier resolves to; the work is Everitt, Carey, Langlois, Ortega & Legg, arXiv:2102.01685 (LIT-010) |
| FIG-008 | Lauro Langosco di Langosco | CIT-005 — goal misgeneralisation | Metadata retrieved. Correct identifier arXiv:2105.14111; title and authorship confirmed (LIT-011) |
| FIG-009 | Tom Everitt | CIT-009 — shutdown, corrupted reward channel | Not retrieved |
| FIG-010 | Bai | CIT-010 — economics of cognition | Doubtful |
A figure note records what the discussion says about a source. It is not evidence that the source says it. Where a figure's argument is standard and widely attested, the register says so as a judgement about the literature — not as a retrieved fact.
Register Rules
- No claim rises above
Bwithout a retrieved page behind it. - A reading is labelled a reading. The Bolo supervisory constraint and the Reapers' consciousness claim are both the discussion's interpretations and are flagged as such in their records.
- Nothing is deleted. Superseded or doubtful entries stay, with the doubt attached.
- A row may carry
Bor above only if its identifier was resolved to a live primary record during the writing session. Otherwise it is capped atDand markedIdentifier Unverified. Recalled identifiers are inadmissible, however familiar the work. This rule was added after the session-2 audit found three of the first session's arXiv identifiers to be false and a fourth attached to the wrong title and authors — see Identifier Audit. - A retrieved source is rated for what was actually opened. An arXiv abstract page establishes what the paper says it does; it does not establish that the paper's results are correct, and no row here claims they are. A preprint therefore caps at
Bregardless of its reception.