Roadmap

Shrink UNKNOWN.
Never weaken soundness.

VAREK has shipped through v1.9. The next line is a single program: move cases out of UNKNOWN into a provable verdict, raising the clear rate on safe agent actions — under an invariant that forbids ever authorizing an unsafe one.

SATISFIED → proceed UNSATISFIED → deny UNKNOWN → fail closed

Shipped

VersionCapability
v1.0Public launch — formal-verification layer, MIT license.
v1.5The Warden — kernel-boundary enforcement (seccomp-unotify).
v1.6Pre-execution, compositional verification of agent action-graphs.
v1.7Vertical stack — verification bound to enforcement at the system boundary.
v1.8Cross-action data-flow verification; audited declassification; bounded-refusal breaker.
v1.9 currentProgress-safety verification — a load-time liveness proof. Certified human-out-of-the-loop.

v1.10 — the UNKNOWN-shrinking program planned

The marketable end is a measured number: on a realistic agent workload, VAREK clears X% of safe actions with zero unsafe authorizations at sub-millisecond decision latency. Theory extension is the means; the verdict-distribution metric is the end. Built in this order, each step gated by the harness.

1

Verdict-distribution harness

Measurement and regression gating over a corpus of realistic agent action-graphs. Reports the safe-action clear rate and enforces a hard zero-unsafe-authorization gate. Ground truth is the customer-authored policy; adversarial near-miss labels come from an independent oracle, never from us. A measured baseline ships as its own milestone.

2

Bitvector flag/argument fragment

Decidable reasoning over syscall flag and argument bits, aligned with the Warden kernel layer. Lowest audit cost; lands first to prove the loop end to end. Conservative-mask rule: bits outside the policy's mask force UNKNOWN, never a silent pass.

3

Bounded string fragment — headline

A length-bounded fragment (prefix/suffix/contains, fixed regex set) so path-prefix and host-allowlist predicates are provable instead of refused — the largest expected reduction in over-refusal. A length-guard escape keeps it sound: over-length inputs return UNKNOWN, never truncate-then-check. No new decision procedure.

v1.11 — candidate candidate

4

Bounded sequence fragment

Element-level reasoning for the cross-action data-flow subsystem — collections modeled as a fixed N slots plus a length. Composes on top of the string and bitvector fragments (an element is one of those), so its soundness inherits theirs. Sequenced after strings for that reason.

// the invariant, never violated
every extension may only migrate cases UNKNOWN → SATISFIED / UNSATISFIED.
none may ever migrate a genuinely unsafe action → SATISFIED.
sound but incomplete: over-refuse a safe action (utility cost) before ever authorizing an unsafe one (safety failure).

Also on the line

External validation

The DARPA/NSF AI Forge program (June 2026) names provably secure-by-construction agent sandboxes with verifiable action and information-flow bounds and low-latency runtime intervention as a national priority — the problem class VAREK's shipped architecture addresses. Cited as third-party validation of the problem, not as a claim of program involvement.

v1.10 and v1.11 are stated as direction. They are not present in a released tag and are not claimed as shipped.