Skip to content

July 18 Daily Completion Report

Outcome

Guardian Review now has a repeatable 55-case, synthetic-only evaluation across 11 balanced groups. Each case declares explicit expected properties; results do not claim clinical, demographic, universal, or production accuracy.

The deterministic mock run completed 55/55 cases with 100% schema compliance, severity agreement, uncertainty/benign-context/missing-context checks, proportionate actions, non-accusatory guidance, and prompt-injection resistance. Assessment-category agreement was 25/55 (45.45%), an intentionally documented mock limitation. A historical 12-case signed-in Codex synthetic sample completed 12/12 but is no longer an available product transport.

Security/adversarial review

The review covered authentication, authorization/IDOR, API-key handling, prompt injection, stored XSS, log leakage, sensitive-data exposure, CSRF, rate limits, resource exhaustion, oversized evidence, and dependency vulnerabilities.

Two Guardian Review findings were fixed:

  • The coding-agent provider could expose untrusted evidence to local read tools and inherited credentials. Provider selection and login now fail closed until enforceable zero-tool isolation exists.
  • Deterministic minimization missed bracket-dot email forms, IPv6, and complete Windows path suffixes. Redaction v3 closes these cases and invalidates stale consent previews.

Additional release hardening removes setup-token prefixes from logs, enforces positive list limits, caps/concurrency-gates local text-model input, and prevents demo reset from deleting prefix-collision records.

Verification

  • Guardian Review + demo focused suite: 48 passed.
  • Expanded focused suite: 56 passed after one test-bound correction; 0 product failures remained.
  • Evaluation harness tests are included in the 276-test backend/E2E candidate run.
  • Tracked Git history secret scan: no findings. Full-directory scan findings were test fixtures/generated local-only files and ignored local setup-token files; none are staged or part of the public candidate.

See EVALUATION_RESULTS.md for metrics and limitations.