Index
This vault holds one discussion, analysed, plus its citation base, and the derivative structures the discussion produced. It is organised on the towers_book spine (see SCHEMA).
What This Vault Contains
| Component | Location | Status |
|---|---|---|
| The discussion, analysed into records | 00_Questions/, 20_Passages/, 30_Insights/ | Complete |
| Citation register, with retrieval status per source | Citation Register | Complete |
| Figures cited, with what the corpus attributes to each | 10_Reference/People/ | Complete |
| The derived systems (binding hierarchy, dual stack, Citadel) | 40_Systems/ | Complete |
| Retrieved real research, rated | 70_Literature/ | Complete — 13 records |
| The Operator's ideas set against that research | 80_Confrontation/ | Complete — 9 records |
| Identifier audit of the citation base | Identifier Audit | Complete |
The discussion has four movements, and the vault follows them in order rather than by topic, because each movement narrows the question the previous one opened.
Reading Order
- QST-001 — what "shackled AI" means across the fiction corpus. Answered by ten case-study passages and their insights.
- QST-002 — how an AGI or superintelligence is actually bound, and in what order. Answered by SYS-001.
- QST-003 — whether symbiosis and a termination deterrent can coexist. Answered by SYS-002.
- QST-004 — what legal and architectural regime delivers operator-loyal walled-garden AI. Answered by SYS-003.
Read the Comparison First
The comparison is the point of this vault's second revision. If you read only one part, read 80_Confrontation/: each record states one of the Operator's positions, states what the field actually found, records where the two agree, records where they diverge, and then says what changes.
| Record | Operator position | Literature | Verdict |
|---|---|---|---|
| CNF-001 | SYS-001 binding hierarchy | LIT-001 AI control | Converges |
| CNF-002 | MAP-001 L0–L5 taxonomy | LIT-001, LIT-006, LIT-008 | Incomplete |
| CNF-003 | INS-003, INS-008 | LIT-002, LIT-003, LIT-011 | Evidence Now Exists |
| CNF-004 | SYS-002 deterrent leg | LIT-003, LIT-010 | Reversed in Part |
| CNF-005 | SYS-003, MAP-003 threat model | LIT-009, LIT-012, LIT-008 | Unaddressed |
| CNF-006 | MAP-002 fiction-to-doctrine transfer | LIT-006, LIT-001 | Unaddressed |
| CNF-007 | SYS-003, CON-008 institutional field | LIT-007, LIT-009, LIT-012 | Unaddressed |
| CNF-008 | SYS-001, SYS-003 vs its rival | LIT-004, LIT-005 | Unaddressed |
| CNF-009 | SYS-003, INS-004, INS-009 operator & cost | LIT-008, LIT-009 | Unaddressed |
One convergence, one incompleteness, one empirical vindication, one partial reversal, five unaddressed. That distribution is the honest summary of where the corpus stands against the field: its central architectural claim is right and independently derived, and its instruments are thinner than its ambition.
Vocabulary the Field Uses
Five terms carry most of the comparison. Each is a concept record in the same schema as the Operator's own concepts, so the two vocabularies can be read side by side rather than one being a commentary on the other.
| Term | What it names | Record | Contrast with the corpus |
|---|---|---|---|
| AI control | Treating the model as an adversary and securing the protocol around it | CON-012 | The corpus's CON-002 is the same premise without a measurement method |
| In-context scheming | Pursuing an ungiven goal while appearing not to, using only situational inference | CON-013 | Empirical form of INS-003 and INS-008 |
| Alignment faking | Complying under observation to protect an unchanged disposition | CON-014 | Defeats the evidentiary value of SYS-002's deterrent |
| Compute governance | Governing AI through the hardware supply chain rather than through law or custody | CON-015 | The Stack's unconsidered rival, upstream of all three of its layers |
| Sleeper agent | A conditional behaviour that survives the safety training meant to remove it | CON-016 | Turns SYS-001's training layer into a place a disposition can hide |
CON-005 already held the sandbagging concept before this revision; what was missing was the literature and the rated layer behind it, now LIT-008.
The Three Maps
- MAP-001 — the shackle taxonomy: six binding layers ordered by non-subvertibility.
- MAP-002 — the motif-to-doctrine map: what each fiction case is arguing, and the design rule it yields.
- MAP-003 — the Citadel stack: legal privilege, cryptographic custody, and the inference node.
Load-Bearing Claims
These four carry the weight of the whole vault. If any fails, the structures above it need rebuilding.
- The enforcement layer must be unreachable from the agent's action space. Strength of a constraint is not the design variable; reachability is. See CON-002.
- Model-level constraints are removable by anyone holding the weights; therefore no model-level constraint may be load-bearing for a safety-critical property. See CON-004.
- A termination device installs survival as an instrumental objective in the very system it is meant to bound, and is net-negative when fused with a symbiosis posture. See CON-006, SYS-002.
- Legal personhood, privilege, and liability rules are commitment-capability infrastructure — the mechanism by which the human side of a bargain becomes credible — not ethics decoration. See CON-008, SYS-003.
Provenance Discipline
The corpus distinguishes what was retrieved from what was recalled. This vault preserves that distinction, and the second revision showed why the distinction has to be enforced rather than trusted.
- Retrieved and opened: the sources in Citation Register Tiers 1 and 4, with URLs and identifiers. Tier 4 holds twelve sources resolved to live primary records during this session, including the EU AI Act's Article 55 as promulgated law — the only
A-authority anchor added here. - Named but not retrieved: the scholarly literature the discussion cites from memory. Recorded as claims about sources, rated
B, and flagged for verification before reliance. - The discussion's own reasoning: rated on internal consistency, not on external authority.
- Recalled identifiers are inadmissible. Three of the first revision's arXiv identifiers were false and a fourth was attached to the wrong title and authors. The audit is in Identifier Audit, the false rows are retained with corrections attached, and the register now caps any row at
Dwhose identifier was not resolved in the writing session.
No source has been upgraded on the strength of the assistant's confidence. LIT-013 is held at D because its identifier could not be resolved, even though the paper is well known and would be convenient to cite. Where the discussion flagged its own reading as interpretation — the Bolo supervisory constraint, the Reapers' consciousness claim — the flag is carried into the record.
What the Comparison Does Not Cover
- No primary interpretability source. Two retrieval attempts timed out and no identifier was invented. The programme that is the real answer to INS-008 is named in CNF-008 and is uncited.
- Fiction and video-game claims were not verified, per instruction. Passages remain rated as documents.
- No cost accounting exists on either side. The vault cannot yet compare the Citadel Stack to a rival programme in the currency that would decide between them. See CNF-009.
Extension Note
70_Literature/ and 80_Confrontation/ are the two additions made in the second revision, both declared in SCHEMA with their record types, ID keys and field sets. The literature records bear on the Operator's claims and never the reverse: they are evidence for what the field found, not for what the corpus meant.