Requirements
1–3 daysQuestion
What is failing, and what would “better” mean?
Evidence
Three real failures, data rights, acceptance metric, budget
Exit rule
No quote without a measurable target.
PROCESS
Six gates.
No skipping the baseline.
Every project moves through the same gates. If a gate fails, we stop, change the approach, or shrink the scope — we do not train through it.
Baseline
2–5 daysQuestion
How does the current model actually perform?
Evidence
Fixed test set, baseline outputs, failure taxonomy
Exit rule
No training without a recorded baseline.
Small experiment
1–2 weeksQuestion
Is there a measurable signal in the right direction?
Evidence
Small high-quality dataset, before/after comparison
Exit rule
No scale-up without a signal.
Full training
1–4 weeksQuestion
Can the signal hold at production scale?
Evidence
Frozen config, seed, data manifest, training log
Exit rule
No launch without client budget approval.
Acceptance
2–5 daysQuestion
Does it hit the metric without unacceptable regression?
Evidence
Held-out evaluation, regression suite, failure cases
Exit rule
No delivery without a signed acceptance.
Productize
OngoingQuestion
What can be reused on the next project?
Evidence
Reusable eval, data recipe, report template, or adapter recipe
Exit rule
At least one reusable asset per project.
→Deliverables
What “done”
looks like.
The exact package depends on the service tier. A full delivery includes the artifacts below; a diagnostic or experiment delivers the relevant subset.
- 01-requirements/ — scope, acceptance metric, data rights
- 02-data/ — data manifest, cleaning log, held-out test set
- 03-training-config/ — reproducible config, versions, seed
- 04-model-or-adapter/ — adapter or merged model (pointer)
- 05-evaluation/ — baseline/post-training comparison, regression, failures
- 06-inference-example/ — minimal runnable example
- 07-deployment-guide/ — hardware, latency, cost, deployment notes
- 08-limitations/ — known limits, failure modes, risks
→Roles and change control
Who does what.
- Owner — positioning, pricing, contract, client relationship, final acceptance.
- AI operator — diagnostic, data, training, evaluation, report, project record.
- Client — data, business acceptance, deployment environment.
What does not move.
- Acceptance metrics are frozen before training.
- New scope is quoted separately; it does not enter the original project.
- Cloud spend above the cap requires new approval.
- Evaluation data never enters the training set.
ACCEPTANCE
Measured,
not asserted.
A project is accepted on the held-out test set agreed in the SOW: the target metric, the regression suite, and a review of failure cases.
- Primary metricWritten before training
- Test setHeld out and frozen
- RegressionNo silent capability loss
- Failure casesReviewed and documented
- Sign-offClient acceptance
START HERE
Bring a real
failure case.
The fastest way to start is three examples where the current model failed, plus the result you expected. We will tell you which gate applies.