Should Your EM Stay on a Free Coding Lane After the Pilot? Use This Term-Drift Sheet

Picture a Thursday standup. Six engineers. One engineering manager. Three sticky notes on the board: free lane, paid seats, self-host.

Nobody is fighting about which demo felt smart. They are fighting about a clock. The pilot felt cheap. The next planning window is ninety days. If the free terms move, who pays the exit?

I use that scene as a composite, not a named customer. I keep the same argument when I write the stay-or-leave call down. So which sticky note still matters if the sticker you liked last month is gone?

The call that can reverse

I am not asking whether AI coding "works." That question is too wide, and it lets everyone win the meeting.

I am asking something an engineering manager can actually own. Should this cohort stay on a free coding lane, buy a paid hosted lane, or self-host, once you price term drift?

Term drift is just this: quota, model set, server option, or data terms can change inside your planning window. You do not need a rumor to model it. You need a weight you will defend, a person who owns the row, and a date when the exception dies.

A scorecard is a conversation tool. It is not objective truth. If finance hates one input, change it and rerun. Why else keep a sheet?

What I will treat as fact

Disclosure: This article was prepared as part of MonkeyCode's product outreach.

The operator describes MonkeyCode as an open-source project with free model access and a free server option. Those two availability claims are the only product facts I will put in column one.

I am not naming models. I am not stating a token quota, a box size, an uptime number, or a duration. A round figure on a slide, including any "millions of tokens" sticker, stays out until someone opens the live terms and writes today's date beside it.

I drafted this on 2026-10-12. If you read it later, re-check the repo and the terms before you reuse my blanks. Would you want a planning sheet that froze a quota I cannot cite from this desk?

Name the inputs before the debate

The numbers that enter the fight

Fill these from your logs and your own quotes. Do not borrow mine past the example, and delete the example when you are done.

• N — engineers in the cohort. Useful context. Not the price.
• Rw — hours per month on a second human pass you can tie to lane output. Not all review. Only rework you would skip if the lane were absent.
• C — fully loaded hourly cost your finance partner already uses. A local number. Not a market average.
• I — monthly invoice for the paid hosted alternative, same cohort, quote dated this week.
• M — monthly self-host cost: people slice plus infra you would actually pay. Include on-call. Leave sunk laptops out.
• E — engineer-hours to leave the free lane. IDE config, CI remote, prompt files, and the habit of pasting into one box.
• W — planning window in months. Start at 3 unless your cycle is shorter.
• p — a planning weight from 0 to 1 for terms, or the free server option, moving inside W. Your risk appetite. Not a vendor forecast.

N does not enter the formula. I keep it so nobody proves the cost with headcount and walks out early. Seen that move before? I bet you have.

The three monthly costs

• Free: F = Rw * C + (p * E * C) / W
• Paid: P = I + (p_paid * E_paid * C) / W
• Host: H = M

Stay on the free lane only if F is lowest and every hard gate is still true. Otherwise buy or host. A low F never overrides a failed gate. Got that?

Paid can win on a higher sticker if exit load and rework make free worse. Self-host can win while looking expensive, if custody fails for both hosted paths. The unit is planning dollars per month, not a token sticker and not a seat count.

A filled example you should throw away

These numbers are a worksheet. I did not run them on a production org. They are not a product benchmark. Replace each one.

Suppose N = 6. Over 14 days the team labeled 9 pull requests lane-rework. Each second pass took about 25 minutes. That is 9 * 25 / 60 = 3.75 hours in two weeks, about 8 hours in a month. Set Rw = 8.

Finance says C = 110. Friction is 8 * 110 = 880 dollars a month. Small. Maybe too small. We will poke it in a minute.

A paid quote pasted into the doc this week says I = 1800. You think a contracted lane needs less exit work, so you set E_paid = 6 and p_paid = 0.2. Self-host is a quarter of a platform engineer plus a small box. You wrote M = 4600. Leaving the free lane looks like E = 16 hours. You set W = 3 and p = 0.5, because you will not bet a release train on an uncontracted server option.

Then the arithmetic:

• Free exit load: (0.5 * 16 * 110) / 3 ≈ 293
• F ≈ 880 + 293 = 1173
• Paid exit load: (0.2 * 6 * 110) / 3 ≈ 44
• P ≈ 1844
• H = 4600

On this sheet, free wins the cost line. Does the team stay? Only if the gates pass. A cheap lane that cannot hold secrets, or that nobody can leave, is not cheap. It is an unowned outage with a friendly price.

Three flips I want said out loud

Which single input, moved by a believable amount, reverses the call? If you cannot answer, you do not understand your own sheet.

1. Rework. If Rw rises from 8 to 22, friction becomes 2420. Free planning cost becomes about 2713, above paid. Buy wins on cost. I would not wait for a prettier demo to notice that. The label log would have screamed first.
2. Term weight. If p goes from 0.5 to 1.0, free exit load doubles to about 587. F becomes about 1467. Still under paid in this example. So term fear alone does not flip this particular sheet. Rework does. That tells the EM to instrument labels, not to collect hallway rumors.
3. Custody. If any repo in the cohort is secret-adjacent, and you have not read this week's data terms for the free lane and the free server option, both hosted paths fail the gate. Self-host wins even at 4600. Cost is no longer the argument. Stop quoting it.

The sheet is allowed to say stay. It is also allowed to embarrass the stay. Which outcome are you hoping gets buried in the appendix?

Hard gates, one owner, one expiry

I will not let a low F override these.

1. Data terms for the free model lane and the free server option were read by a named person in the last 7 days. Link the page. Write the date. A memory from onboarding does not count.
2. No production secret, customer payload, or unreleased vulnerability note goes through a lane whose retention you cannot describe in one sentence.
3. The free server option is not the only remote for a release-train job. A second path exists, even if it is slower and annoying.
4. One owner. Not a chat channel. A person who can pause the lane on a weekday without a committee.
5. Expiry. The exception dies on a calendar date inside W. Use 30 days, then a rerun. No silent renew.
6. Exit drill. The owner has run the leave-steps once on a scratch repo. If they have not, E is a fantasy. You may not use it.

Who owns the row? The engineering manager owns the decision. A platform lead owns the exit drill. Finance owns C and I. If those three names are blank, you do not have a decision. You have a vibe with arithmetic on top.

Log the rework, then run the math

The label, before the feelings

Agree on lane-rework before the two-week log. A pull request gets the label only when someone redid work because the lane's output was wrong for this repo. Spec changes do not count. Taste arguments without a redo do not count. Be strict, or Rw becomes a junk drawer.

# Proposal commands. They do nothing until the label exists in your repo.
# Window start is 14 days before the draft date, 2026-10-12.
gh pr list \
  --search "label:lane-rework updated:>=2026-09-28" \
  --state all \
  --limit 200 \
  --json number,title,updatedAt \
  > /tmp/lane-rework.json

jq 'length' /tmp/lane-rework.json

Put minutes in the PR body as a trailer line: rework-min: 25. Crude. Still better than a memory from standup.

gh pr list \
  --search "label:lane-rework updated:>=2026-09-28" \
  --state all \
  --limit 200 \
  --json number,body \
  > /tmp/lane-rework.json

python3 - <<'PY'
import json, re
rows = json.load(open('/tmp/lane-rework.json'))
mins = []
for row in rows:
    found = re.search(r'rework-min:\s*(\d+)', row.get('body') or '')
    if found:
        mins.append(int(found.group(1)))
print('prs', len(rows))
print('with_minutes', len(mins))
print('hours', round(sum(mins) / 60, 2))
PY

Double the two-week hours if you want a rough month. Do not pretend that extrapolation is a forecast. It is a stand-in until you have four weeks. If the two weeks included a holiday, say so in the notes. Why hide the weird week?

The calculator

This script is a proposal. I have not executed it against a real organization, and it does not call a vendor API. It will not tell you that a product is safe.

#!/usr/bin/env python3
'''Term-drift sheet. Replace every number. Not a benchmark.'''

def monthly(rw, c, p, e, w, invoice=0.0, host=0.0, mode='free'):
    if w <= 0:
        raise SystemExit('W must be > 0')
    exit_load = (p * e * c) / w
    if mode == 'free':
        return rw * c + exit_load
    if mode == 'paid':
        return invoice + exit_load
    if mode == 'host':
        return host
    raise SystemExit('unknown mode')

def main():
    c = 110.0
    free = monthly(8.0, c, p=0.5, e=16.0, w=3.0, mode='free')
    paid = monthly(0.0, c, p=0.2, e=6.0, w=3.0, invoice=1800.0, mode='paid')
    host = monthly(0.0, c, p=0.0, e=0.0, w=3.0, host=4600.0, mode='host')
    costs = {
        'free': round(free, 2),
        'paid': round(paid, 2),
        'host': round(host, 2),
    }
    winner = min(costs, key=costs.get)
    print(costs)
    print('cost_winner', winner)
    print('gates_still_required', True)

if __name__ == '__main__':
    main()

python3 term_drift.py
# Edit rw from 8 to 22 and run again. Did the winner move?

The flip point is the only number I want on the whiteboard. Everything else is scaffolding.

Who should not use this

Skip the sheet if you cannot label rework for two weeks. A blank Rw makes free look free. That is how a zero becomes a strategy.

Skip it if the cohort handles regulated data and nobody has read the current terms. Do not pilot first and ask legal later. Later is how an unreviewed server option becomes a finding.

Skip it if you wanted a model bake-off. This does not score answer quality. It scores whether a cheap lane still fits after drift, rework, and exit. A green toy prompt will not save the meeting.

Skip it if you need a procurement signature tonight. The sheet is not a contract review, a security assessment, or a capacity test of anyone's free server.

Skip it if the owner will not put an expiry on the calendar. An exception without a death date becomes the architecture. You do not need a brand name on that pattern to recognize it.

Limits I want in the notes

p is not a probability I measured. It is a weight. Write that next to the cell, or a later reader will quote it as science.

E is usually too low the first time. People forget prompt files, CI secrets, and the junior who only knows one remote. Add a second pass after the scratch-repo drill, then trust the number.

The paid quote goes stale. Date it. A quote older than W does not count, even if the PDF still looks official.

Self-host M hides on-call until the first weekend page. If your platform team is already saturated, add a delay cost or stop pretending M is complete.

I also did not compare features, latency, or model quality. I cannot, not from this desk, and I will not invent a leaderboard to decorate the post. If quality is the real fight, review your own repos. Then come back and put the rework hours here.

Open-source does not, by itself, answer custody or exit. A repo you can read is not the same as a server option you can leave on a Tuesday. Treat those as separate rows.

Where column one can come from

If a free model lane and a free server option are the hosted pair you want in the free column, read MonkeyCode's current repository and terms, write the date on the sheet, and only then drop in numbers you can defend. I would rather you delete my example than inherit it.

So which input flips your call: rework hours, the term weight, or a custody gate you have been skipping?