Prepared for Example Enterprise · Take-home submission draft, not an official Applied Compute document.
PRELIMINARY: DO NOT SEND CUSTOMER: HOLD - MODEL
In short
You are buying 3 managed reinforcement-learning (RL) cycles to improve Held-out pass rate for Internal workflow automation, from a measured baseline of 60% toward a target of 80%, over 6 weeks. The model is Qwen3.6-35B-A3B, an open-weight model with 35 billion parameters, of which 3 billion are active for each token. The total is $247,247. The 80% target is a review point, not a guarantee: when it is reached, or when results stop improving, you decide whether to continue. You approve every cycle, and each invoice is issued when its milestone completes. The total below covers everything in the three phases; only usage above the included allowances and the items under Not included are extra.
Where your money goes
Setup package: evaluation and environment $49,449(20%)
Managed RL cycles (3) $173,073(70%)
Handoff and independent verification $24,725(10%)
What you pay for, phase by phase
Phase 1 · Weeks 1–2
Setup package: evaluation and environment
$49,449
What you get
Workflow and data review, with your subject-matter experts
Frozen held-out evaluation set (200 baseline cases) and baseline score
Grader and reward specification, versioned and signed off by you
Production-realistic environment with a smoke-tested end-to-end RL dry run
Go / no-go recommendation with the expected improvement range
Why it costs this
The foundation every RL cycle depends on: an evaluation you can trust, a grader that measures what matters, and a baseline to measure improvement against. You keep these assets and can reuse them with any future model.
Billed: Week 2, when the go/no-go gate is signed
Phase 2 · One cycle every 1 week
Managed RL cycles (3)
$173,073
What you get
Rollouts: 4,000 attempts per cycle
Held-out evaluation: 200 cases per cycle, run on untouched data
Managed compute: 3,870 GPU-hours included across the engagement
Agent platform usage: 172.5 million tokens included
A written score report and your stop or continue decision after each cycle
Why it costs this
Each cycle runs the full loop for a standard-size model: train, generate rollouts, evaluate on held-out cases, and report. This phase is all-inclusive: compute, the agent platform, the engineers who run each cycle, and evaluation. You receive no separate compute or engineering invoices. Usage above the included allowances is billed as set out in the usage section.
Billed: Invoiced as each cycle completes
Phase 3 · Week 6
Handoff and independent verification
$24,725
What you get
Final evaluation on data never used in training
Checkpoint and weights package, with configuration and traces
Reproducibility report and an operating runbook
Signed handoff acceptance against the agreed evidence
Why it costs this
An independent check on data the model never trained on, plus everything your team needs to run the model without us.
Billed: Week 6, on signed handoff
When you are invoiced
Invoice
When
Amount
Running total
Setup gate
Week 2
$49,449
$49,449
Cycle 1 of 3
Managed cycle, usage included
Week 3
$57,691
$107,140
Cycle 2 of 3
Managed cycle, usage included
Week 4
$57,691
$164,831
Cycle 3 of 3
Managed cycle, usage included
Week 5
$57,691
$222,522
Handoff and verification
Week 6
$24,725
$247,247
Total training fee
$247,247
$247,247
What changes the price
If you…
Price effect
Why
One more cycle
+$42,403
One more managed cycle beyond the approved count, with its compute, platform usage and engineering included.
One fewer cycle
-$51,634
One managed cycle less. The setup and handoff packages do not change.
Stop after cycle 1
$107,140
The setup invoice and the first cycle invoice. No later invoices are issued. The handoff package is not billed unless the handoff is signed.
Another workflow, data source or region
Change order
Scope changes are quoted separately, through a change order, before any work starts.
What you can be charged
Maximum committed fee
$247,247. The three invoices add up to this and no more, whatever the cycle timing.
Included usage
3,870 GPU-hours of managed compute and 172.5 million tokens on the agent platform, which includes a 15% buffer over planned use.
Usage above those allowances
Only with your written approval, billed at the rates in the usage section, and up to $25,000 in total without a further change order.
Committed-cycles price
Available when you commit to 4 or more cycles (5% off) or 6 or more (10% off).
Rates
Fixed for the term of this statement of work.
Price validity
30 days from issue.
What one cycle includes, and what you pay if you stop
Each managed cycle includes the engineers who run it, the agent platform, the compute to train and roll out the model, and the held-out evaluation. Cycles are invoiced in equal parts of the cycle total, because each one carries the same work.
If you stop at
Total invoiced to that point
Stop after setup
$49,449
Stop after cycle 1
$107,140
Stop after cycle 2
$164,831
Full engagement (with acceptance)
$247,247
Fees are not contingent on the score. Reaching the target, or a plateau, triggers a review; you then decide whether to continue, and no later invoice is issued unless you approve it.
Adding a cycle costs $42,403: only that cycle's compute, platform use and engineering are added, while the setup and handoff do not change. The per-cycle invoice is a share of the cycle total, so it is not the price of one more cycle.
Ownership of deliverables
[The trained weights, grader, held-out evaluation set, environment and traces delivered to you: ownership and licence terms, including the base-model licence, to be confirmed by counsel.]
Not included
Production model serving (a separate order, priced after a load test)
Preparing your data beyond the agreed sample (quoted as a change order)
Additional workflows, data sources or regions
Licences for the base model or any third-party tool you choose to add
Cycles beyond the approved count, and usage beyond the stated allowances (billed as set out in the usage section)
Words used in this proposal
RL cycle
One round of training on the model's own attempts, scoring them, and updating the model.
Rollout
One attempt by the model at a task, recorded with its result.
Held-out evaluation
A test set the model never trains on, used to measure real improvement.
Grader
The rules that score an attempt, for example pass or fail, or a rubric score.
Allowance
The compute and agent platform usage included in the price for the whole engagement. Usage above it is billed at the rates in the usage section.
Checkpoint
A saved version of the trained model, with its settings, so results can be reproduced.
Go / no-go gate
A decision point at the end of the foundation phase. You approve it before any cycle is paid for.
Change order
A short written amendment that adds or changes scope and its price before the work starts.
Statement of work (SOW)
The contract schedule that defines scope, phases, evidence, invoices and amendment rules.
Common questions
Why do I pay for the foundation before training starts?
Training is only as good as the test that measures it. The foundation builds that test, the grader and a baseline, and you keep them. It is the one part of the work that makes every later cycle measurable.
Is engineering billed separately?
No. Engineering is included in the setup and cycle prices. You are not invoiced for hours; each phase is priced for the work it delivers.
What is included in the total?
Everything in the three phases: setup, managed cycles with compute and platform usage up to the allowances, and handoff. Only usage above the allowances and the items under Not included are extra.
Why does the price depend on the model?
A larger model needs more compute for each cycle, so the managed cycle price is higher. The price shown already reflects the model you have chosen.
What if a cycle does not improve the score?
You decide after each cycle whether to continue, change direction or stop. You pay for the cycles and milestones completed. We do not stop contracted work silently, and we do not charge for work you have stopped before it starts.
What happens if I stop early?
You pay the invoices up to the stop point, and nothing more. Any third-party compute we committed for you on your approval is billed at cost. The acceptance payment is billed only on a signed handoff.
Acceptance
Each phase is accepted against the objective evidence listed in the statement of work, and each invoice follows the same milestone.