Why Manual Quality Evaluations Are Essential for LLMs
The instinct to automate human review out of the loop is understandable. Review takes time, it doesn't scale linearly, and every approval gate feels like friction between generation and publish. But that instinct is wrong, and acting on it moves the cost of errors downstream to the worst possible moment: when a real reader notices the model got something wrong.
The Review Inbox Is the Product, Not a Feature
Kiln routes every generated draft to a review inbox before any publish action fires. That's not a UX decision. It's an architectural one. The control plane triggers generation; the job plane runs the worker; the draft lands in the inbox. Nothing touches LinkedIn, Instagram, or a blog export until a human approves it. The one-directional flow from generation to inbox to publish is explicit by design.
The reason this matters is that autonomous publishing removes the single checkpoint where accuracy, voice, and intent can be verified against the source material. The model received a prompt containing your brand profile and a set of retrieved source chunks. It produced something. But "something" is not the same as "what you'd actually publish." The gap between those two things is exactly what a reviewer catches, and no downstream system catches it afterward at lower cost.
The argument against this is that approval gates introduce latency and human bottlenecks. That's fair at scale. At the volume Kiln is designed for, sub-100 generations a week, the latency is trivial and the bottleneck is a feature. You're not running a wire service. You're building a durable body of published work, and every piece that goes out under your name carries your credibility. The inbox is where that credibility gets defended.
Append-Only Draft History Turns Human Judgment Into Durable Signal
Every draft version in Kiln is preserved. Edits don't overwrite the prior state; they append to a history. That means the delta between what the model produced and what a reviewer actually changed is recorded, permanently, as a structured artifact.
This is higher-signal data than any automated metric because it captures practitioner judgment at the moment of decision. When a reviewer changes the opening paragraph, rewrites a claim, or cuts a section entirely, that edit encodes something specific: the model's output was wrong in a way the reviewer could identify and correct. Aggregate enough of those deltas and you have a detailed record of where the model consistently falls short for your brand, your sources, and your audience.
Automated metrics can tell you a draft hit 900 words and included an H1. They can't tell you that the model consistently overstates a product's capabilities when the source material is cautious, or that it drops the hedging language your legal team cares about. The edit history tells you that, provided you preserve it. Append-only storage is the mechanism that makes the history trustworthy. There's no way to quietly fix a draft and erase the evidence of what changed.
Automated Evals Catch Format; Humans Catch Meaning
Kiln's eval layer can verify structural rules reliably. A baseline blog eval checks for H1 and H2 structure, a word count target of 800 to 1100 words, and an opening paragraph that functions as a hook rather than a summary. Those checks are worth running. They catch regressions in generation format before a reviewer ever sees the draft.
But format is not meaning. An eval can confirm that a draft has the right structure and hits the word count. It cannot confirm that the draft faithfully represents the source chunks the model was given. That failure mode, a draft that is structurally correct but factually misrepresents the source material, is the one that damages trust with an audience. It's also the one that's hardest to detect automatically, because the model produces fluent, confident prose regardless of whether it's accurate.
The mechanism here is retrieval. Generation in Kiln is source-grounded: prompts contain the brand profile and selected chunks from ingested sources. The model synthesizes from those chunks. But synthesis is not quotation. The model can misread emphasis, drop qualifications, or blend two separate claims into one. A human reviewer who knows the source material catches this. An eval that checks word count does not.
The practical split is: run automated evals to catch format regressions before human review, not instead of it. The eval pass is a triage filter. The human pass is the quality gate.
Skipping manual eval doesn't save time. It shifts the cost to the moment a reader finds the error, which is after publish, after the piece has been shared, and after your credibility has taken the hit. The review inbox, the edit history, and the distinction between format checks and meaning checks all exist to prevent that moment. Build the pipeline so a human sees every draft before it goes out, preserve every edit they make, and treat that record as the most valuable data your generation system produces.