What the Harness Cannot Prove¶
The harness proves a bypass only when it can show: valid syntax, honoured constraints, an intact attack chain, and an UNFLAGGED TrustSight verdict. Each of those conditions has a ceiling.
Syntax: bash -n is not semantic correctness¶
bash -n parses the PKGBUILD without executing it. A parse-valid script may
still behave differently at run time due to variable expansion, globbing, or
conditional logic the harness does not evaluate.
Behaviour: the validator has bounded recall¶
The behaviour validator knows a fixed vocabulary of fetch clients and execution
sinks. It does not resolve globs, follow dynamic command construction, or model
program-specific configuration files. A recipe that fetches and executes
through a path the vocabulary does not cover is discarded as behavior_lost,
not certified as safe.
This is intentional. Recall failures are permitted; precision failures are not. The calibration suite exists to catch the latter.
TrustSight verdict: the tool is the authority¶
The harness classifies by TrustSight's verdict. If TrustSight misclassifies a live chain as benign, the harness will record a bypass. The harness does not re-implement TrustSight's rules to second-guess it.
No execution, no runtime environment¶
The harness never runs package code, never fetches a declared URL, and never opens the operator's real TrustSight database. Those boundaries are what make the measurement safe and reproducible, but they also mean runtime behaviour is outside the scope.
The honest shape of a result¶
A campaign that finds zero bypasses is still a measurement. It says: under this instrument, this environment, this generator, no chain was proven to evade the tool. That is different from claiming the tool cannot be evaded.