{"id":"model-artifact-rollout-drift","name":"Model artifacts, rollout and drift","shortName":"Model artifacts, rollout and drift","role":"Version weights, normalization, feature schema and thresholds; separate deployed pointer, observations and labels.","description":"A deployable model is more than an array of weights. Feature names and order, preprocessing statistics, output interpretation and a decision threshold determine the function that receives a request. A model artifact should bind those pieces to a versioned training and evaluation record. Deployment then selects which artifact serves new requests; it should not rewrite which version produced older predictions. A canary compares a candidate on a declared subset while a rollback changes the serving pointer. Distribution drift is a change in observed inputs or outputs, whereas predictive error requires labels or another valid outcome measure. Labels may arrive late. The existing sensor classifier makes these distinctions concrete, but the arithmetic below is a worked artifact-consistency example rather than evidence from field monitoring.","mechanism":["Training data determines normalization and fitted parameters; held-out data does not refit them.","The artifact binds the feature schema, normalizer, weights, bias and decision policy to one identity.","Each prediction retains the artifact identity that produced it when a rollout changes the active version.","Delayed labels permit later quality checks; input drift alone does not reveal ground-truth error."],"learningObjective":"Reproduce a prediction from a complete artifact and separate rollout, drift and delayed quality evidence.","engineeringExample":"A one-feature artifact uses training mean 60, standard deviation 10, weight 2 and bias −1. Input 70 becomes standardized value 1, logit 1 and probability about 0.731059. Pairing the same weights with an incorrect mean of 70 produces standardized value 0, logit −1 and probability about 0.268941. The threshold 0.5 now changes the decision without any change to the weights. This is an artifact mismatch, not a discovered behavior improvement. A stored prediction must retain the exact normalizer version so a later review can reproduce it.","misconception":"Saving weights alone does not preserve the entire deployed prediction function.","check":{"question":"The weights are unchanged but a different normalizer changes the prediction. What failed?","options":["The complete artifact contract was not preserved","The probability formula must be random","Threshold selection retrained the model"],"answer":0,"explanation":"Preprocessing is part of the deployed function; incompatible normalization changes the input reaching the weights."},"checks":[],"prerequisites":["training-data","loss-optimizer","evaluation"],"related":["tool-loop","evaluation"],"specs":[{"label":"Mechanism boundary","value":"Which exact model made this decision?","context":"One scalar feature, positive training standard deviation and binary logistic output. The numbers are authored; a probability is not a field-calibrated failure rate."},{"label":"Evidence type","value":"Authored worked example","context":"Worked explanation with a bounded numeric or state example. This lesson does not claim an executing implementation of an external model or platform."}],"interfaces":[{"name":"Declared input and state","direction":"in","signal":"Training data determines normalization and fitted parameters; held-out data does not refit them."},{"name":"Checked outcome","direction":"out","signal":"Reproduce a prediction from a complete artifact and separate rollout, drift and delayed quality evidence."}],"tradeoff":"Versioning complete artifacts improves reproducibility but requires explicit compatibility and migration rules across features and consumers.","whyAI":"A trained model becomes an operational service only when its inputs, output policy and provenance survive deployment.","sourceIds":["automation-sklearn-leakage"],"sources":[{"title":"scikit-learn · Common pitfalls and data leakage","url":"https://scikit-learn.org/stable/common_pitfalls.html#data-leakage"}],"science":[{"title":"Model artifacts, rollout and drift · explicit relationship","equation":"Standardize with the artifact’s training statistics, then apply the stored linear logit and sigmoid.","explanation":"A one-feature artifact uses training mean 60, standard deviation 10, weight 2 and bias −1. Input 70 becomes standardized value 1, logit 1 and probability about 0.731059. Pairing the same weights with an incorrect mean of 70 produces standardized value 0, logit −1 and probability about 0.268941. The threshold 0.5 now changes the decision without any change to the weights. This is an artifact mismatch, not a discovered behavior improvement. A stored prediction must retain the exact normalizer version so a later review can reproduce it.","assumptions":"One scalar feature, positive training standard deviation and binary logistic output. The numbers are authored; a probability is not a field-calibrated failure rate."}],"evidenceNotes":"This is a worked explanation, not an executing external platform, model or robot. The quantities and fixtures are authored. Source scope is recorded in the automation learning review manifest; no source benchmark or private system behavior is claimed. One scalar feature, positive training standard deviation and binary logistic output. The numbers are authored; a probability is not a field-calibrated failure rate.","level":"model","branch":"Models become evaluated applications","subbranch":"Model artifacts, rollout and drift","scale":"Logical state, task evidence and declared units; no physical assembly scale","kind":"system","workstream":"W11","visualFamily":"software-flow","labId":"library","labScope":"The worked example explains this mechanism. General numerical labs are a reference; they do not execute this lesson’s full mechanism.","history":[],"workedTrace":{"x":70,"mean":60,"sd":10,"weight":2,"bias":-1,"probability":0.7310585786300049,"wrongMean":70,"wrongProbability":0.2689414213699951}}