World◦Model: the world is your oyster

World◦Model: the world is your oyster

The AHT Game: How Slow Can You Be Before Mercor Cuts You

I pulled the billing and timing rules out of four Mercor project guidelines I've worked under. Two of them contradict each other outright. Inside the Data Factory, Episode 2 of 4.

Cong's avatar
Cong
Aug 03, 2026
∙ Paid

Last episode I wrote that every Mercor project has a corridor of acceptable speed, that nobody tells you where its walls are, and that people get terminated for hitting either one. Dozens of you wrote back with the same question.

The Dark Side of Mercor You Should Know Before You Start

The Dark Side of Mercor You Should Know Before You Start

Cong
·
Jul 28
Read full story

So what’s the number?

I want to answer that properly, because it’s the wrong question in a way that costs people their contracts. And the fastest route to why is a single line I found buried in a coding project’s guideline.

Time tracking was mandatory on every task. And the same document stated, plainly, that pay on that work was not calculated from the tracking at all.

Sit with that for a second. Logging was compulsory on a task where the money didn’t depend on it. Further down, the reason: the project screened for fraud by reading output quality, volume, and time records against each other.

Your time log is not a timesheet

That reframe explains almost everything confusing about speed on this platform.

On an hourly project, working faster does not earn you more. It earns you less. Twenty tasks in ten hours pays worse than twenty tasks in fifteen. So when an ops lead looks at a low average handling time, they are not admiring your efficiency. They’re asking whether the work actually happened, or whether it was skimmed, pattern-matched, generated by a model, or handed to a friend. And when they see a high one, they ask the mirror question: is this person padding hours, or just not good at this?

The upper wall has arithmetic behind it, and the arithmetic belongs to the client lab rather than to Mercor. The lab budgets per datapoint and pays per hour, so cost per task equals your handling time times your rate. That same coding project published both halves: $85 an hour, against a task the guideline sized at roughly one hour. That isn’t pacing advice. That’s the unit price of the dataset, stated in advance. Work at ninety minutes a task and every datapoint you produce costs the lab fifty percent more than the budget it was approved against. Nobody emails you about that. It surfaces as a capacity review you’re not in.

So the corridor is real, and both walls have owners. The upper wall is the client’s unit cost. The lower wall is the fraud model.

There's one more thing the instrument does, and it's the only good news in this episode, so I'll come back to it at the end: handling time is also an input when writers get promoted to reviewer. It's watching for fraud, and it's quietly building a shortlist.

The rules contradict each other across projects

Here’s why “what’s the number” has no answer, and it’s worse than most people assume.

I went back through the billing language on four projects I’ve worked. On one, the rule was blunt: charging time outside onboarding and active task work would end the contract and follow you to future work with Mercor. Sitting and waiting for a task to appear was explicitly not loggable.

On another project, running on the same platform in the same period, reading the documentation and keeping up with Slack were put on the clock as work you were expected to log.

Those are opposite rules. Log your Slack reading on the first project and you’ve committed a terminable offense. Fail to log it on the second and you worked for free, and you also depressed your own handling time in a way that reads as suspiciously fast.

There is no platform-wide billing policy at Mercor. There is a per-project policy, written by whoever stood up that engagement, and the only safe assumption is that your last project’s rules do not transfer. Which means the instinct you built on your first project is not an asset. It’s a liability you carry into your second.

And it gets one layer worse, because the written rule and the enforced rule are not the same rule either. Every project has a special project lead, the SPL, and some are markedly more tolerant than others. The same logged hour that draws a warning under one lead passes without comment under the next. Neither of them changed the document. They’re just different people reading it.

The rates that are arithmetically impossible

Then there’s the case where no amount of calibration helps, because the number was never reachable.

The clearest instance I’ve seen described recently: a project asking contractors to read roughly eighty pages of dense technical material they had never seen, produce a careful analysis, use no AI whatsoever, and finish inside forty-five minutes. The role required a professional doctorate. Contractors ran the obvious numbers. A page a minute is already brisk for technical reading, which leaves zero minutes for the analysis they were actually hired to write.

And the enforcement is not theoretical. One contractor reports being removed from two separate projects for exceeding handling time. The ramp problem compounds it: several describe project managers who allow no settling-in period at all, so the week when you are slowest by definition, your first, is the week you get measured in.

There is a much less frightening version of how those thresholds actually bite, and I’ll give it to you below the paywall, because it’s the single most useful thing I know about this and it changes what you should worry about. But I’m not going to soften the front half. People do get cut for this.

The sieve

There’s a darker reading of all this that circulates constantly among contractors, and it deserves stating plainly even though I can’t prove anyone intended it.

Put the pieces side by side. A rate that cannot be hit honestly. A rule against the one tool that would close the gap. And a contractor pool deep enough that anybody who objects is replaceable by Friday.

Whatever was intended, that combination selects. It filters toward people willing to make up the difference with unpaid time, or with tools they’ve been told not to use. It filters out the people who log honestly and work at human speed. Nobody at Mercor has to have designed that for it to be what the machine does. And the people it selects for are the ones who look, on the dashboard, like the good workers.

If that’s right, the number isn’t a target. It’s a sieve.

So the real question was never what the number is. It’s how you find the number, on a project you’ve never worked, under a lead whose tolerance you haven’t measured, at a rate that may not be reachable, before you’ve logged an hour you can’t take back.

I’ve been paid on time, every week, across ten-plus projects, and I’m still here. So the corridor is survivable. The rest of this episode is how, and it’s the part I charge for, because it took me ten-plus projects and a few offboardings to assemble.

What’s below the paywall:

  • What the four numbers on your dashboard actually mean, and which one holds veto power over the other three

  • Why the offboarding threat in your project Slack lands far less often than it’s stated, and the single factor that decides whether it lands on you

  • How to price a project’s expected pace from documents already in your hands, before you accept one task

  • The three moves that work when the rate is impossible, including the question to a reviewer that usually reveals the real bar is lower than what you’ve been producing

  • Which parts of a task you can safely automate, and which will get your logged hours stripped

  • Which of your two timers your handling time is computed from, and the second time record you’re creating without knowing it

  • One thing I do that I don’t recommend, and why I’m telling you anyway

  • Whether an AHT offboarding follows you to the next project. A subscriber asked me directly; this is the real answer

  • The reading of your handling time that works in your favour, because it’s also an input when reviewer seats get filled

Episodes 3 and 4 build directly on this one, and both are paid.

And paid gets you the thing this newsletter is actually built on. I read job listings the way I just read those four guidelines. Tell me your background, and each week I send the roles I’d apply to if I were you, by DM and in your own weekly email, with what I found when I checked them.

One placement pays for years of this.

User's avatar

Continue reading this post for free, courtesy of Cong.

Or purchase a paid subscription.
© 2026 Wayne · Privacy ∙ Terms ∙ Collection notice
Start your SubstackGet the app
Substack is the home for great culture