dynamicsystemsarchitecture.org

Isolation Under Adversarial Pressure

Not "does NCFCA's isolation work in principle" — does it hold up better than realistic alternatives under real adversarial pressure. Two escalating comparisons, both run directly, not asserted.

Round 1 — against the standard, honest baseline

Most real systems keeping data channels separate use a naive policy check: a conditional guard at each call site (if source == allowed_source). Not a strawman — it's the actual standard pattern in code without a dedicated schema-enforcement layer. Tested both approaches against 500 simulated callers: legitimate, careless, hostile, and legacy-refactored:

Naive policy-check isolation (typical existing approach):
  Contamination events that got through: 277
  Correctly blocked:                     99

NCFCA structural isolation:
  Contamination events that got through: 0
  Correctly blocked:                     376

The naive approach isn't badly written — it's the standard pattern, and it has a bypass parameter that can be (and here, was) forgotten or skipped. NCFCA's isolation has no bypass parameter to forget in the first place — the only way to write is through a path that always checks.

Round 2 — against a genuinely hard competitor

Capability-token isolation is how real, security-conscious systems handle this when they take it seriously: a write requires a cryptographically signed, scoped token from a trusted authority, not a string comparison. No skip parameter on the write path — same design rigor as NCFCA's own approach. This is a real competitor, not a second strawman. Tested against 500 calls: legitimate, token replay, hostile forgery, wrong-scope reuse:

Capability-token isolation (strong alternative):
  Breaches: 20
  Blocked:  352

NCFCA structural isolation:
  Breaches: 0
  Blocked:  240

The token system's 20 breaches came specifically from replay — a legitimately-issued token reused outside its intended scope. NCFCA has no concept of a reusable token at all; it checks source identity per call, which makes an entire failure class (replay) structurally inapplicable rather than something that has to be separately guarded against.

Summary

ApproachDesign rigorBreaches / 500
Naive policy checkStandard, honest, real-world common277
Capability tokensStrong — cryptographic, scoped, no skip parameter20
NCFCA structural isolationNo bypass exists to forget or replay0

Source

isolation_baseline.py · strong_alternative.py