The L0–L5 binding hierarchy
The passage, verbatim. The discussion's central technical table, reproduced without alteration:
| Layer | Mechanism | Stops | Does not stop |
|---|---|---|---|
| L0 Substrate | Energy, fabrication, materials, compute, physical placement | Self-improvement, self-replication, scaling | Persuasion, lateral influence |
| L1 Environment | Default-deny egress, tool broker, credential isolation, sandbox | Real-world action, exfiltration, lateral movement | In-context reasoning about the constraint |
| L2 Architecture | Read-only weights, no self-modification path, separate control process | Weight theft, self-edit, recursive improvement | Deception, output-channel attacks |
| L3 Training | Corpus filtering, curriculum, reward shaping, unlearning | Accidental capability acquisition | Deliberate circumvention, in-context override |
| L4 Interpretability | SAE feature clamps, circuit breakers, refusal-direction ablation | Specific behaviours, specific leakage channels | General capability, novel exploits of the same channel |
| L5 Oversight | Human approval, monitoring, deception probes, audits | Silent divergence, some deceptions | Adversarially aware agents |
Why it is load-bearing. This is the vault's spine. Every other architectural claim is either a defence of this ordering or a consequence drawn from it. The ordering is not by technical sophistication and not by strength — it is by whether the constrained system can reach the enforcement layer, which is why substrate and environment outrank interpretability.
Provenance note. Verbatim from the retrieved discussion file, authority A — this is the corpus speaking, not a secondary reading. The table is the discussion's synthesis; the individual mechanisms are standard practice and are not attributed to any cited paper.