black O2, inc.
ACTIVE
    FOR-001 / 40_SystemsSYS-001

    The Binding Hierarchy

    Active system

    CAUTION This record carries a caution marking. Read it before relying on it.

    title
    The Binding Hierarchy
    Type
    system
    Identifier
    SYS-001
    Status
    Active
    Tags
    • assembly-theory
    • axiomatic
    • control
    • architecture
    Aliases
    • L0–L5 system
    • The Shackle Stack
    method
    Axiomatic Decomposition via Assembly Theory
    rebuilt from
    Towers vault (towers_book)
    connections
    • [[MAP-001]]
    • [[CON-001]]
    • [[CON-002]]
    • [[CON-004]]
    • [[PSG-011]]

    The Binding Hierarchy

    A system assembled from the corpus's six-layer account of how a machine can be held, ordered by a single principle: whether the constrained system can reach the constraint.

    Primitives

    • Reachability, not strength, orders the layers — a binding is ranked by whether the constrained system's action space includes the binding, because a constraint the system can act upon will eventually be acted upon. PSG-011
    • A shackle is a mechanism plus a purpose — the term names a binding installed to limit a capable system's conduct, not any limitation the system happens to have. PSG-011
    • Non-subvertibility is a property of the deployment — whether a binding can be defeated is decided by the arrangement of weights, processes and access, not by the binding's own design quality. PSG-011
    • Model-level constraint is removable — refusal behaviour trained into weights can be stripped by an actor holding those weights, so no model-level constraint may be load-bearing. PSG-011
    • Internal constraint is unauditable from outside — a binding that lives entirely in the system's own evaluations offers no observation point from which its continued operation can be confirmed. PSG-008
    • Ignorance-based control is unstable — a constraint that depends on the governed entity not representing its own situation fails retroactively and discredits the relationship when representation arrives. PSG-005

    Assembly Map

    1. Reachability → layer ordering → the three deep, deployable layers are substrate, environment and architecture
    2. Model-level removability → no L3/L4/L5 binding may be load-bearing → L2 must carry custody
    3. Internal constraint unauditability → interpretability cannot be the end state → external observation point required
    4. Ignorance-based instability → disclosure is a design variable → oversight designs premised on concealment fail
    5. All of the above → the system's central claim: the durable bindings are external to the model, and the model-level layers are supporting instruments rather than load-bearing ones.

    ⚠ gap — the corpus asserts L3 (training) fails to deliberate circumvention without citing a retrieved study of the circumvention; the claim is carried by the abliteration material, which is itself described rather than cited.

    Selection Notes

    Selection criteria applied: textual abundance (L0–L5 appears as a table and again in layer-by-layer commentary), internal consistency (the table's "Does not stop" column agrees with the later prose), explanatory power (the ordering predicts which published attacks succeed), and compatibility with external constraint (it does not require a claim about model internals that the corpus cannot support).

    Retained and flagged rather than deleted:

    • The symbiosis and deterrence material (PSG-012) — passes abundance and consistency but is not a binding layer, so it sits in SYS-002 instead of being forced into this decomposition.
    • The fiction case studies (MAP-002) — they attest that binding is a live cultural preoccupation and supply failure modes, but each is a single-source reading; they inform the primitives rather than constitute them.
    • Abliteration as a named technique (CON-004) — retained as a primitive's evidence while flagged as unretrieved; the layer ordering would survive its removal but would lose its strongest support.

    Open Questions-Tensions

    • L2 versus centralisation. The system's load-bearing layer is custody of weights. CON-009's Sovereign Inference Node assumes compute centralisation makes custody architectural. INS-007 reads the same corpus as saying weights are already in the wild. Both cannot be the dominant case; the corpus does not adjudicate.
    • Auditability versus effectiveness. PSG-008 holds that a perfect internal binding is an unauditable one. Taken seriously, this pushes assurance entirely into L0–L2 and leaves L3–L5 as instruments that can only be trusted when externally corroborated — which the corpus does not provide a method for.
    • The hierarchy is silent on legitimacy. Nothing in the ordering asks whether a binding is justified. INS-009 and PSG-010 establish that this question cannot be recovered at enforcement time, which places the entire legitimacy burden on SYS-003 — a different system, with different instruments, and no bridge in the corpus.
    • Against doctrine. If interpretability matures to the point of reliable external readout of internal constraints, the ordering's L4 demotion becomes an artefact of current technique rather than a structural claim.

    Session 2 Addendum — What the Literature Does to This System

    • The ordering is externally corroborated, and not by citation. LIT-001 founds the AI control field on the same premise — safety lives in the deployment protocol, not the model's disposition — and reached it through red-team testing rather than through fiction. The convergence is recorded in CNF-001 and it raises this system's standing rather than lowering it.
    • The layer table is not complete. MAP-001 has no row for collusion between instances, steganographic channels, sandbagging, or subversion of the evaluation itself. All four are named and tested in LIT-001, LIT-006 and LIT-008. See CNF-002.
    • The capability-gap dependency is unstated and load-bearing. AI control works because the trusted monitor is weaker but independent. As that gap closes the protocol weakens. Nothing in the hierarchy records the gap as a variable, so the ordering reads as more durable than its own mechanism allows.
    • The strongest objection is absent from the vault. The anti-control position holds that external enforcement buys time rather than safety, and that the durable answer is mechanistic interpretability — the only approach that could distinguish compliance from alignment faking. CNF-008 states the objection and records that no primary interpretability source was retrieved.
    • A rival chokepoint. LIT-007 locates the governance lever in compute and training clusters, which is not the same object as inference-node custody. The hierarchy and the compute-governance programme are competing answers to "where is the handle," and the vault has not compared them.

    Sources

    • PSG-011 — the L0–L5 table and layer commentary; supplies the ordering principle and the layer-by-layer failure modes.
    • PSG-008 — supplies the unauditability primitive.
    • PSG-005 — supplies the ignorance-instability primitive.
    • CON-001 — establishes shackle as mechanism-plus-purpose.
    • CON-002 — establishes non-subvertibility as a deployment property.
    • CON-004 — establishes abliteration and the no-model-level-binding rule.
    • CON-005 — the behavioural counterpart to removal: capacity suppressed rather than stripped.
    • MAP-001 — the same content as a comparative table.
    • LIT-001 — AI control; the independent derivation of this system's premise. See CNF-001.
    • LIT-006 — sabotage evaluations; the template for what a validated layer claim looks like.
    • LIT-008 — sandbagging; the failure mode the layer table cannot express.

    Back to Shackled AIForests