Who checks the AI that approves your article?

An AI can write a convincing article. A second AI can approve it. The article can still carry an invented claim.

We run that loop at TrustGrowth: an AI writer drafts posts and social copy, an AI reviewer grades each draft against a rubric, and the writer revises on those notes. Before handing over more, we want a defensible answer to one question. Which configuration deserves trust to write, review and revise a particular kind of content, and what human work is left afterwards?

What we found

A benchmark result proves how one model configuration performed on one sample under one rubric on one date. It does not prove that the same setup suits your content type, your acceptance standard or your reviewer, and this piece reports no measurement of its own.

• Published writing and judge benchmarks answer narrow questions. None answers whether a particular reviewer should gate your product update.
• Our own review gate runs in shadow mode beside human decisions. This piece reports no figures from it; the protocol below is how we will earn the right to quote any.
• Writer, reviewer and revision need separate scorecards. A style result cannot offset a factual error.
• First of seven pieces on choosing a model to trust for writing, reviewing and revising content, and on the human work that remains.

Four terms, defined

• Benchmark. A fixed set of tasks, a scoring rule and a published result for a named model on a named date.
• Evaluator. The model, prompt and settings that grade an output. Also called a judge.
• Rubric. The written standard the evaluator applies. A wrong rubric makes a wrong evaluator.
• Human approval. A named person who reads the evidence and signs the decision to publish.

What does a benchmark result prove?

A writing benchmark hands a model a brief and grades the output. WritingBench scores 1,000 briefs on five criteria generated per case (paper). EQ-Bench Creative Writing v3 grades literary craft and ranks pairwise. LitBench checks a judge against curated story preferences; Hemingway-bench collects blind expert preferences.

A judge benchmark grades the grader. LLMBar asks whether an evaluator picks the response that actually followed the instruction, JudgeBench whether it picks the objectively correct answer in hard pairs, CriticEval separates feedback, comparison and correction, and RewardBench 2 covers reward models. A third family checks facts: FActScore, SAFE and FACTS Grounding ask whether claims are supported by retrieved or supplied evidence.

So a result establishes something narrow: this prompt, this rubric, this sample, this model, this date. It establishes nothing about a content type nobody sampled or your editorial standard. Reading two benchmarks' code at pinned commits sharpened that in four ways.

A name is not a version. The pinned WritingBench README records a move from 1,239 cases to 1,000 and two judge-model changes during 2025. A number quoted without a version can compare two instruments.

Scoring code needs checking too. In the pinned runner, the resume path re-processes the final input row of a finished job and appends it again, and the aggregation step sums every emitted record. Read together, those two paths let a resumed run shift an average with no new case. That is a reading of pinned code, not evidence about any hosted leaderboard.

Documentation drifts from the scorer. EQ-Bench Creative Writing v3 documents Glicko-2 while its inspected rating code calls a TrueSkill solver, and its pairwise judge prompt sees at most the first 4,000 characters, so a strong opening can carry a weak ending.

"Human votes" is no longer a safe label. On 30 July 2026, Arena announced AutoEval: a reward model produces proxy preference votes, combined with live human votes. Record whether a public number came from a person or a proxy.

Where does the benchmark stop and the workflow decision begin?

The benchmark stops at the tested sample; the workflow decision starts where you name the content type, the failure you cannot ship and the person who signs. Before comparing anything, write down the content type, the acceptance criteria, the failure modes you cannot ship, what the reviewer may change alone, and where a human decides.

Table 1. Proposed evaluation protocol, dated 2026-09-06. No results are reported and no model is placed in it.

Stage Evidence you need The failure that matters Human responsibility Writing First drafts on your briefs and sources An unsupported claim that reads well Owns brief, sources, rubric Reviewing Labelled defective and clean drafts, graded blind A miss, and equally a false block Defines a defect; audits both directions Revising One draft revised with and without each reviewer, graded by an independent panel Voice sanded off while findings clear Judges better, not just cleaner Final approval The record of what was claimed and what supported it A claim nobody can stand behind Signs and is accountable

That is the evaluation we designed: three linked suites reported separately, never one weighted number, factual failures gated apart from style. The design is written; no bakeoff has run and nothing is deployed.

What should a reviewer be tested on?

On labelled clean and defective drafts, with its misses, its false blocks and its stated reasons scored apart. An evaluator inherits every assumption in its rubric, prompts, reference answers and harness, and its own preferences too. The MT-Bench LLM-as-a-judge study (Zheng et al., 2023) documents order and verbosity effects, and the self-preference bias study (2024) gives reason to expect a judge to favour its own family's style; we treat that as a hypothesis.

A reviewer can fail in three ways a writing benchmark never shows.

1. It misses the invented fact. The draft reads well, so it passes.
2. It blocks the good draft. Every unusual sentence becomes a finding.
3. It sands off the voice. The errors are gone, and so is the reason to read.

The test would classify each decision as a correct pass, a correct block, a false pass, a false block or an abstention, and record whether the stated reason names the defect the label names. Outcome and reason are scored apart, because a right decision for a wrong reason will be wrong next time.

Our own gate is the first case. It runs in shadow mode beside human decisions, and we have written up a fail-open scoring defect we found in that gate. The audit will record, for a dated set of drafts, how often the gate blocks on density and format proxies, and how often it stays silent on a title that promises what the body does not deliver, a protocol described but never run, near-duplicates in a cluster, or leaked pipeline text. Until those counts exist with their denominators, we quote none.

Who makes the final approval?

A named person who has read the evidence, not the highest-scoring model. The steps below are process recommendations, not graded results; we follow them and have not tested them against an alternative.

1. Define the content task and audience before choosing a model.
2. Write the acceptance rubric before you look at any output.
3. Select a dated, representative set of your own content, not a public one.
4. Give every candidate configuration the same task, brief and sources.
5. Set the model's judgments beside human judgments on the same drafts.
6. Record false positives, false negatives and unresolved disagreements, denominators visible.
7. Set a human approval rule for high-impact or ambiguous content, and name who signs.
8. Repeat when the model, prompt, rubric or content type changes.

Step five is the expensive one: in our proposed pilot the human calibration hours, not the model spend, dominate the planning arithmetic. That is an estimate at assumed prices, not a quote and not a measured cost.

Method and limits

Four kinds of statement appear above, and none is a measured result. Code observations: readings of public code at the pinned commits linked beside each claim, evidence about that revision and never about a hosted leaderboard. Published benchmark facts: taken from the papers and READMEs linked. Planning estimates: the cost arithmetic, at assumed prices. Proposed tests: Table 1 and the reviewer audit. The internal shadow check of our own gate is unpublished, its labels come from blind LLM judges rather than human adjudication, and its figures will appear only after the protocol has run on a dated set with human reference labels. Any such result would describe one company's drafts and would not transfer to your content types.

Do this next

Fill in this worksheet for one content type before comparing any models.

• Content type. One type, named, with a published example.
• Approval standard. What "publishable" means, written before grading starts.
• Test set. Dated drafts of your own, defects labelled independently.
• Evaluator configuration. Model, prompt, rubric version, decoding settings, date. All five, or it is not repeatable.
• Human reviewer. Named, with authority to overrule the model.
• Disagreement log. Every case where model and human differ, with evidence.
• Escalation rule. What sends a draft to a person whatever the model says.
• Re-run trigger. The change that invalidates the current result.

Trust is a decision about a specific task, supported by a relevant test and a person accountable for the outcome, not a label borrowed from a leaderboard. If you run AI content through an AI reviewer, send us a reviewer miss, or an unwarranted edit it demanded.

Last measured

Every external source was verified on 2026-09-06 at the pinned commits given above: the WritingBench commit timestamped 2025-12-19 and the EQ-Bench Creative Writing v3 commit timestamped 2026-06-24. The Arena AutoEval announcement is dated 30 July 2026, read from an archived snapshot of 26 August 2026. No figure from our own gate is reported here, and no re-run date is set.

One further date, and one boundary worth stating plainly. TrustGrowth publishes a score for its own site, trustgrowth.ai: 43 out of 100, labelled Provisional with a 38 to 48 band, as of 2026-09-06. That page says the score rolls up Authority, Visibility, Growth, Corroboration, Technical, Content, E-E-A-T, Performance and GEO signals into one public snapshot from first-party Google Search Console data, and how the TrustGrowth score works sets out the method. It is a site trust score. It is not a writing-benchmark score, not a model score, and not evidence about any reviewer.