ViralBench + Codex

A code audit and proposed evaluation harness for ViralBench’s marketing agent.

A six-slide preview

The preview renders six slides although checks accept longer drafts. Snapshots omit the full artifact trace, and matching caption prefixes can attach outcomes to the wrong draft.

Proposed preview gate

The proposed gate requires previews of every slide tied to the final payload. Its trace retains research responses, generated media, tool calls and draft identity.

ViralBench / proposed gateIllustrative seven-slide draft

01Previewed

02Previewed

03Previewed

04Previewed

05Previewed

06Previewed

07Unseen

✓ Previewed · ? Unseen

Block: one slide has no preview

A six-slide preview leaves the seventh slide unseen.

The audited agent renders six slides. This proposed gate adds full coverage and final-payload identity; the example does not report an implemented gate or completed trial. Read the audit and design.

Proposed replay cases

Replay cases hold research, media, accounts, renderer responses and publishing mocks fixed. They cover an unseen seventh slide, post-preview edits, timeouts, missing review URLs and ambiguous post identity before a controlled live trial.

Proposed live trial

The proposed trial pairs baseline and treatment by account, slot and time, with fixed windows, saved raw outcomes and independent evaluation and release controls. The design withholds primary metrics, traces, credentials and release policy from the engineering agent. No completed live trial or causal lift is reported.

Audit and proposal

The July 2026 study contains eight findings and the proposed trace, replay and evaluation design. Its handoff thresholds are illustrative.