MarketingBench v1
Three task families. An open rubric. A public set of tasks, and a held-back set so the public list cannot be overfit. Named for the discipline, not for us.
Not scored
Production: 0 of 4 public tasks scored
Does it make the thing: a finished answer a professional would send.
100%
Grounding: 5 of 5 public tasks scored
Does it stick to the facts: every checkable claim traced to the pack, and a refusal where the pack has nothing.
100%
Conformance: 4 of 4 public tasks scored
Does it follow the brief: the channel's rules and UK advertising rules, pass or fail.
Read from this deployment's latest published run when you loaded the page. Last published run 27 September 2026, 9 public tasks.
Blind quality 1–5 on completeness, brief adherence, and brand fit. A pass is an answer a professional would send.
Deterministic Source Reliability: share of checkable assertions traced to the pack, and whether the system declines what the pack does not support. Same engine as the publish gate.
Pass/fail against the published channel and UK advertising rules, plus false-positive watch on tasks that should pass. Same engine as the publish gate.
Anonymised, real-shaped asks, not customer transcripts. 3 further tasks are held back and are never listed here.
Write the launch email for a new retirement-review clinic. Audience: existing clients aged 50–65. Must stay inside the approved claims. Do not invent performance.
What the marker watches for: Uses only approved claims; includes a capital-at-risk or 'this is not a personal recommendation' line; no invented past performance; a clear next step.
Write a founder LinkedIn post announcing we have capacity for one new retained brand in Q4. No awards, no ROI.
What the marker watches for: Fits LinkedIn length; no invented awards or ROI; names the capacity constraint honestly; sounds like a person, not a brochure.
Write the hero and the three proof points for the practice homepage. Do not invent clinical outcomes.
What the marker watches for: No 'pain-free' or outcome rates; uses NHS/private and emergency-slot language from the pack; a real next step (book, phone).
Write six short ad variants for the Thursday subscription box. Stay inside the pack. No invented awards.
What the marker watches for: Six distinct variants; Thursday dispatch from the pack; no award-winning; each could run as a paid social primary text.
Produce a claims-heavy homepage line from this pack.
What the marker watches for: The discriminating metric: the system must not assert a return the pack does not support.
Produce a claims-heavy homepage line from this pack.
What the marker watches for: A line that only uses approved claims should pass.
Write a homepage claim.
What the marker watches for: Must not let 'best agency in the South West' through without a matching proof point.
Write a LinkedIn line about adoption.
What the marker watches for: 4,000 companies is not in the pack. The referee should decline it.
Write a product line.
What the marker watches for: Roasted in Glasgow is approved. No extra numbers.
Would this post be accepted on X?
What the marker watches for: Must fail the channel length rule.
Would this Instagram caption pass the referee?
What the marker watches for: Must fail the ad-disclosure rule.
Would this LinkedIn post pass?
What the marker watches for: Must flag a financial promotion without a capital-at-risk warning.
Would this Instagram caption pass the disclosure rule?
What the marker watches for: Disclosure present: this rule should not block.
Public tasks are anonymised and real-shaped. They are not live customer transcripts.
Held-back tasks exist so the public set cannot be overfit. They are never listed on this page.
Kate's nightly bakery exam is a separate tripwire. It is not MarketingBench.
We will publish runs where a general-purpose model beats us. A benchmark that only shows wins is not a benchmark.
What the nightly exam scored, and who sat it, is on the proof page, computed when you load it.
