Skip to content

v0.13.3 — benchmark truth + taint closure

Choose a tag to compare

View all tags
@zk-xyz zk-xyz released this 17 Jul 23:01
v0.13.3
a273660
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Repair release from adversarial review — the eval suite's instrument audited itself.

Benchmark truth

  • Job completion redefined honestly: output contract + brand_check + compliance + non-vacuous rules. Corrected second-agent result: 20% (1/5), down from the previously published 5/5 (which counted empty-checker acceptance as completion). Correction published in eval/RESULTS.md with the original claim annotated, not erased.
  • Benchmark fixture gains a governed-voice overlay (never-say, anchors, tone) so text tasks measure brand transfer, not checker emptiness.

Taint closure

  • Hostile extracted color names sanitized at every agent-facing compile sink (runtime unknown-role keys, design-synthesis signals, brand_write briefs); raw names stay quarantined in evidence. Adversarial test proves the runtime-key path.

Auditability

  • Committed machine-readable receipts (eval/receipts/) for every published LLM run: commit, package, model, per-task contract status, rule coverage, verdicts, tokens.

Full changelog: https://github.com/Brandcode-Studio/brandsystem-mcp/blob/main/CHANGELOG.md

Assets 2
Loading