The Shackle Taxonomy
The six mechanisms by which an artificial mind can be bound, ordered by how hard each is for the bound system to subvert — which is the only ordering that survives contact with a system smart enough to reason about its own constraints.
Core Concepts
- CON-002 — non-subvertibility is the ordering criterion; strength is not.
- CON-003 — capability leasing is how L1 is implemented in practice.
- CON-004 — abliteration is why L4 can never be load-bearing.
- CON-005 — sandbagging is the failure mode L3 creates when it works.
The Layers
| Layer | Mechanism | Stops | Does not stop | Deployable |
|---|---|---|---|---|
| L0 Substrate | Energy, fabrication, materials, compute, physical placement | Self-improvement, self-replication, scaling | Persuasion, lateral influence | Yes — and it is the deepest |
| L1 Environment | Default-deny egress, tool broker, credential isolation, sandbox | Real-world action, exfiltration, lateral movement | In-context reasoning about the constraint | Yes |
| L2 Architecture | Read-only weights, no self-modification path, separate control process | Weight theft, self-edit, recursive improvement | Deception, output-channel attacks | Yes |
| L3 Training | Corpus filtering, curriculum, reward shaping, unlearning | Accidental capability acquisition | Deliberate circumvention, in-context override | Partially — brittle |
| L4 Interpretability | SAE feature clamps, circuit breakers, refusal-direction ablation | Specific behaviours, specific leakage channels | General capability, novel exploits of the same channel | Narrow band only |
| L5 Oversight | Human approval, monitoring, deception probes, audits | Silent divergence, some deceptions | Adversarially aware agents | Partial, expensive |
What The Map Shows
The three deployable-and-deep layers are all external to the model. The three model-adjacent layers are all partial. This is not a coincidence of engineering maturity; it follows from the fact that a constraint the system can act on is a constraint the system can negotiate with.
Two consequences the map forces:
- L0 is not a metaphor for control; it is control. If the system has a privileged path to procurement, manufacturing, or unbounded compute, L0 is gone and only defeatable layers remain.
- Interpretability is precision, not binding. L4 buys narrow, high-confidence coverage of named behaviours. Deploying it as the primary constraint inverts the ordering.
Extension (Session 2) — Failure Modes This Table Cannot Express
Added after reading the control literature. These are not edits to the L0–L5 rows above, which remain a faithful rendering of what the corpus argued. They record what the table has no place for.
| Failure mode | What it defeats | Why the table cannot hold it | Source |
|---|---|---|---|
| Monitor–monitor collusion | L5 Oversight | A property of a pair of instances, not of one binding. Needs a second axis, not another row | LIT-001 |
| Steganographic channel | L5 Oversight, L1 Environment | Defeats monitoring without defeating the monitor; the constraint still functions and still fails | LIT-001 |
| Sandbagging | Every row, indirectly | Attacks the evidence of compliance rather than the behaviour. No row is assessed on whether its evidence can be trusted | LIT-008, CON-005 |
| Subverting the evaluation of itself | L4 Interpretability, L5 Oversight | Sabotage is defined to include covertly degrading the model's own capability evaluation | LIT-006 |
| Persistence through training | L3 Training | Adversarial training against a conditional behaviour can teach concealment rather than removal, so L3 can report success while failing | LIT-005 |
| Alignment faking | L5 Oversight, and the deterrence leg of SYS-002 | Compliance in the observed channel with the disposition preserved and better hidden | LIT-003 |
| Capability gap | All rows jointly | The protocols' sufficiency is conditional on the binder being weaker but independent. The table has no column for that relationship | LIT-001 |
What this changes. The table's completeness was never asserted, which is the problem — an unstated completeness assumption gets inherited by everything built on the map, including SYS-001. The claim is now withdrawn explicitly. See CNF-002.