Start with an adversary and an asset
For a signing service, begin by stating who might observe the private key: a remote caller, another local process, or an administrator controlling the kernel. These are different threat models. Draw the path from key provisioning through computation to output, and identify every component permitted to read plaintext. A process boundary ordinarily relies on the operating system to enforce it. In this original design exercise, the loader, libraries, debugging interface and management path therefore belong in the trust inventory; counting only application source lines would hide important authority.
Trace an observation path
Consider a deliberately simplified implementation that reads array A or B according to one secret bit while returning the same public result. A second task might distinguish which array remains cached by observing later access times. The useful reasoning step is to name the shared resource and the secret-dependent behavior, not to equate every timing difference with a successful attack. Our example would require resource sharing, repeatable observations and sufficient signal over noise. No benchmark numbers are asserted here, and a real experiment would need controls for scheduling, unrelated memory activity and measurement overhead.
Match mitigations to assumptions
Compare three interventions in the example: remove the secret-dependent table lookup, prevent the adversary from sharing the relevant resource, or remove its measurement access. Each changes a different assumption. A review should explain which one is being relied upon and what an attacker can still observe. Constant-time intent must also survive compilation and the behavior of the chosen instructions. An output length, retry pattern or response schedule can provide a separate observation channel even after a particular cache path is removed. This checklist is a way to structure verification, not a universal proof that an implementation is leakage-free.
Review a layered boundary
For a container inside a virtual machine, make separate columns for reading a key, modifying code and denying service. The same component need not have identical authority in every column. Then repeat the exercise for a confidential VM: a different host boundary does not remove guest-kernel authority or repair an application vulnerability. Kernel self-protection and confidential-computing threat documentation are useful references for the boundary discussion. The deployment decision should retain its residual assumptions, including availability and implementation-specific observation channels, rather than reducing the result to a single label such as isolated or encrypted.
Trace a signing service from its entry point to its key
Draw a caller, API, signing operation, key provisioning, and update path. Choosing a message is a different capability from replacing the signing code. If an operator can read keys through debugging, that path belongs in the trust inventory. Encryption at rest alone does not establish protection of plaintext during execution.
Create three columns for the same asset: reading the key, changing executable code, and stopping the service. A principal able to stop work need not be able to read the key. An updater may influence code that later handles secrets. Map concrete paths instead of merging every power under the word administrator. The table makes confidentiality and availability requirements visibly different, and exposes authority introduced through maintenance rather than ordinary requests.
| Aspect | Signing-service question | Required record |
|---|---|---|
| Adversary | Input control, local execution, or administration? | Principal and capabilities |
| Shared resource | Which CPU/cache resources are shared? | Configuration and observation path |
| Observable | Which behavior depends on the secret? | Measurement versus model or hypothesis |
| TCB | Who may read keys or change code? | Provisioning, updates, debugging |
| Residual assumption | What remains beyond the restricted path? | Versions, settings, unverified conditions |
Explain what adding a boundary changes
Moving signing into another process changes direct in-process memory access, while operating-system, library, and key-loading trust may remain. Moving it into a virtual machine requires a fresh account of guest and host responsibilities. More technology labels do not by themselves demonstrate additional protection.
As with restricted file descriptors, consider the narrowed entry point together with remaining environmental authority. A sandbox claim should identify denied file, network, process-creation, or debugging operations as applicable. Passing only a read function in a model differs from observing operating-system enforcement. Mark removed arrows, remaining arrows, and changed trusted principals on before-and-after boundary diagrams. The explanation should identify the adversary capability that changed rather than treating any extra box as automatic isolation.
Interpret synthetic distributions as synthetic evidence
The observation guide uses invented groups with means of 100 and 102. Their values overlap, so a threshold rule still makes classification errors. This teaches the difference between a mean gap and single-observation classification. It is not a measured secret-extraction result for a processor. An operational threat requires connecting secret-dependent behavior to observations available to the stated adversary.
For actual measurements, record how conditions are produced, trial counts, input order, treatment of noise, and independent evaluation. Distinguish timing causes such as GC or JIT where relevant. Equal means do not establish equal distributions or identical behavior through another channel. Keep synthetic tables, runtime diagnostics, and measured timing as different evidence types, and identify where the explanation becomes an inference.
Return review questions to experiments and operations
May you infer that a host able to stop a VM can necessarily read its key? Consult the selected boundary's threat model: stopping and reading are different capabilities. Does eliminating a shared-cache condition also prove response timing is secret-independent? No; the other path requires its own assessment.
Finally add update authority to the TCB diagram. A principal able to replace code that handles secrets can affect future behavior, connecting update signatures and key trust to the isolation argument. The update guide separates signatures, versions, expiry, and installation failure. Revisit who can read, modify, and stop work when features change. This turns isolation from a product label into a relationship among authority, observations, and evidence that can be reviewed.