Here is my review process: [describe - what arrives, who reviews it, how long each takes, what the decision is].<br/><br/>Tell me what this process assumed about the cost of producing one submission. Be specific: how much effort did a submitter previously have to spend, and what did that effort filter out for free?<br/><br/>Then tell me what happens to each stage now that producing a plausible submission is close to free.<br/><br/>Finish with the one assumption whose failure explains most of my current problem.
Below are examples of submissions I rate as good and as poor: [paste both sets].<br/><br/>Identify the signals that still separate them now that generated text is fluent. Explicitly discard anything that only measured effort or polish - length, formatting, grammar, structure, confident tone.<br/><br/>What I need are signals that are expensive to fake: specific verifiable claims, engagement with the particulars of my context, admissions of uncertainty, things only someone who did the work would know.<br/><br/>Rank them by how hard each is to fake, and tell me which I could check automatically.
My queue is overwhelmed. I can rate limit submitters, filter submissions, or both.<br/><br/>Work through the trade for my situation: [describe].<br/><br/>Rate limiting is cheap to ship and punishes volume regardless of quality - it hits the prolific and legitimate alongside the abusive.<br/>Filtering is slower to build, needs a signal that holds, and fails in a different direction: it rejects good work that looks wrong.<br/><br/>Tell me which my situation actually calls for, what the false-positive cost looks like for each, and whether a rate limit as a holding measure buys enough time to build the filter.
I want a first-pass triage that removes obvious non-starters before a human looks.<br/><br/>Here is what arrives and what makes something a non-starter: [describe].<br/><br/>Design the triage so it fails safe - when uncertain, it passes things through to a human rather than rejecting them. Tell me the specific checks, in order of cost, and what each one's false-rejection risk is.<br/><br/>Then tell me what proportion I could realistically remove this way, and what will still reach a human regardless.
Here is data on how my reviewers spend time: [paste, or describe].<br/><br/>Split the time between: deciding on genuinely borderline cases, confirming obvious accepts, confirming obvious rejects, and overhead - chasing information, handling complaints, administration.<br/><br/>If the overhead or the obvious cases dominate, the problem is not volume, it is that skilled people are doing unskilled work.<br/><br/>Tell me which category to attack first for the largest return on reviewer time.
I am about to introduce [limit or filter] on my process.<br/><br/>Write the honest public explanation: what the limit is, why it exists, who it will unfairly affect, and what has to be true for it to be removed.<br/><br/>Do not write it as a policy announcement. Write it as something a person who is about to be inconvenienced by it would find reasonable.<br/><br/>Then tell me what question I will be asked most often in reply, and whether I have a good answer.