Machine Learning

Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion

Evaluate whether an adaptive learning policy should ship by inspecting logged propensities, IPS/SNIPS/DR estimates, holdout evidence, guardrails, coverage, and promotion thresholds.

status: reviewimportance: importantdifficulty 4/5math: graduateread: 30mlive demo

Back to Machine Learning Next: Learner Outcome Experiment Design: Outcomes, Power, Retention, Guardrails (review)

Concept Structure

Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion

01Intuition

Start with the picture, metaphor, or geometric mechanism.

02Math

Make the objects explicit and connect them with notation.

03Code

Mirror the equations with runnable implementation details.

04Interactive Demo

Manipulate the mechanism and watch the idea respond.

3prerequisites

1next concepts

4related links

Learner Contract

What this page should let you do.

You are here becauseEvaluate whether an adaptive learning policy should ship by inspecting logged propensities, IPS/SNIPS/DR estimates, holdout evidence, guardrails, coverage, and promotion thresholds.

This Machine Learning concept is the current idea: keep the same invariant visible across Intuition, Math, Code, Interactive Demo.

Before thisAdaptive Curriculum Optimization: Policy, Utility, Exploration, Audit (review)

3 prerequisites listed; refresh them before leaning on the math or code.

Adaptive Curriculum Optimization: Policy, Utility, Exploration, Audit (review)Evaluation Pipelines Causal Inference Basics: Correlation versus Intervention (review)

By the end4/4 sections ready | runnable code expected | live demo

Explain the mechanism, trace the main notation, and test one prediction in the live demo.

Do this firstIntuition

Read the intuition before the notation; the math should name a mechanism you already felt.

Then go nextLearner Outcome Experiment Design: Outcomes, Power, Retention, Guardrails (review)

Follow this edge after making one prediction here; the next page should reuse the result, not restart the route.

Test the linkManipulate one control and predict the visible change.Then continue to Learner Outcome Experiment Design: Outcomes, Power, Retention, Guardrails (review)

Claim/source review status

Substantive review recorded

2/2 claims have bounded review metadata; still check caveats and source scope.Metadata-derived; review may be AI-assisted. Not a human certification.

Claims2/2 reviewed

Sources5 cited

Codeattached

Demolive

Reviewed2026-07-05

Updatedpage 2026-07-05

Learning item flow

4/4 sections readyAsk about this Research room

ConceptLearning System Policy Evaluation: Logs, Estimators, Guardrails, Prom...Machine Learning EquationLearning System Policy Evaluation: Logs, Estimators, Guardrails, Prom...Exact equation CodeLearning System Policy Evaluation: Logs, Estimators, Guardrails, Prom...Exact runnable code DemoLearning System Policy Evaluation: Logs, Estimators, Guardrails, Prom...Interactive demo ClaimLearning-system policy evaluation needs logged context, action, rewar...Exact claim check SourceUnbiased Offline Evaluation of Contextual-bandit-based News Article R...Exact source

ConceptLearning System Policy Evaluation: Logs, Estimators, Guardrails, PromotionMachine Learning

5 sources attachedLocal snapshot ready

concept:machine-learning/learning-system-policy-evaluation

Codenearby Predictbefore reveal Roomitem handoff

01IntuitionStart with the picture, metaphor, or geometric mechanism.02MathMake the objects explicit and connect them with notation.03CodeMirror the equations with runnable implementation details.04Interactive DemoManipulate the mechanism and watch the idea respond.

Intuition

Build the mental picture first so the rest of the page has something to attach to.

Section prompt

Adaptive curriculum optimization chooses a promising next policy. Learning-system policy evaluation asks whether that policy deserves to ship.

That is a different job. A policy can look good in a utility score and still fail because the logs do not cover the new actions, the reward model is biased, the confidence interval crosses zero, a learner guardrail regresses, or the improvement only appears in one subgroup.

A policy-evaluation workbench should keep separate evidence lanes:

logged evidence: context, action, reward, propensity, and timestamp,
overlap: whether the old logging policy tried the actions the new policy wants,
estimator choice: replay, IPS, self-normalized IPS, direct model, or doubly robust,
online holdout: randomized treatment/control evidence,
guardrails: frustration, assessment reliability, coverage, retention, and accessibility risk,
promotion threshold: what must be true before rollout,
audit trail: what was estimated, what was blocked, and why.

The core habit is to avoid treating "the policy scored higher" as proof. The system should say: this estimate comes from these logs, with this overlap, this uncertainty, these guardrails, and this promotion decision.

Math

Translate the story into symbols, assumptions, and a derivation you can inspect.

Section prompt

Equation 1 $q_i = q(a_i \mid x_i).$ Equation 2 $w_i = \frac{\pi(a_i \mid x_i)}{q(a_i \mid x_i)}.$

Each logged event has context $x_i$ , action $a_i$ , reward $r_i$ , and logging propensity

q_i = q(a_i \mid x_i).

For a target policy $\pi$ , the importance weight is

w_i = \frac{\pi(a_i \mid x_i)}{q(a_i \mid x_i)}.

The inverse-propensity estimate is

\widehat V_{\text{IPS}}(\pi) = \frac{1}{n}\sum_{i=1}^n w_i r_i.

The self-normalized variant trades some bias for stability:

\widehat V_{\text{SNIPS}}(\pi) = \frac{\sum_i w_i r_i}{\sum_i w_i}.

With a reward model $\hat r(x,a)$ , a compact doubly robust estimator has the shape

\widehat V_{\text{DR}}(\pi) = \frac{1}{n}\sum_i \left[ \sum_a \pi(a \mid x_i)\hat r(x_i,a) + w_i \left(r_i - \hat r(x_i,a_i)\right) \right].

Estimator health also depends on overlap. A simple effective-sample-size diagnostic is

\operatorname{ESS} = \frac{(\sum_i w_i)^2}{\sum_i w_i^2}.

Promotion is a separate decision:

\operatorname{promote} = \mathbf 1[ \Delta_{\text{holdout}} > \tau \land \operatorname{lowerCI}_{\text{DR}} > 0 \land \operatorname{ESS}/n > \rho \land \operatorname{guardrails}=\text{pass} ].

The formulas do not remove judgment. They make the judgment inspectable.

Code

Keep the implementation aligned with the notation so the algorithm is legible.

Section prompt

Runnable code 1events = [ {"reward": 0.62, "propensity": 0.50, "target_prob": 0.50, "target_model": 0.58, "l...python

events = [
    {"reward": 0.62, "propensity": 0.50, "target_prob": 0.50, "target_model": 0.58, "logged_model": 0.56},
    {"reward": 0.70, "propensity": 0.50, "target_prob": 0.50, "target_model": 0.64, "logged_model": 0.60},
    {"reward": 0.55, "propensity": 0.50, "target_prob": 0.50, "target_model": 0.57, "logged_model": 0.52},
    {"reward": 0.73, "propensity": 0.50, "target_prob": 0.50, "target_model": 0.66, "logged_model": 0.62},
]

weights = [e["target_prob"] / e["propensity"] for e in events]
ips = sum(w * e["reward"] for w, e in zip(weights, events)) / len(events)
snips = sum(w * e["reward"] for w, e in zip(weights, events)) / sum(weights)
direct = sum(e["target_model"] for e in events) / len(events)
dr = sum(
    e["target_model"] + w * (e["reward"] - e["logged_model"])
    for w, e in zip(weights, events)
) / len(events)
ess = sum(weights) ** 2 / sum(w * w for w in weights)

print(round(ips, 3), round(snips, 3), round(direct, 3), round(dr, 3), round(ess / len(events), 3))

The witness prints several estimates rather than one magic number. If overlap is poor, the right response is often to collect better randomized logs or run a guarded holdout, not to celebrate the largest estimate.

Interactive Demo

Use direct manipulation to connect the explanation to a moving system.

Section prompt

Use the evaluation lab to predict the right policy-evaluation move before the proof unlocks:

Estimator: decide which estimator or logging fix fits the evidence.
Promotion: decide whether to promote, hold out, block, or collect more logs.
Guardrails: decide whether learner-risk metrics override positive lift.
Coverage: decide whether overlap and subgroup coverage are strong enough.

Before reveal, exact propensities, estimator values, effective sample size, selected action, confidence intervals, guardrail verdicts, and promotion decisions stay locked. After reveal, inspect the evaluation ledger and ask whether the decision follows from the evidence.

Live Concept Demo

Explore Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion

The stage is code-native and interactive. Use it to test the explanation against the mechanism.

difficulty 4/5graduatecode-aligned

Demo Prediction Checkpoint

Manipulate one control and predict the visible change.

Commit to what Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion should make visible before reading the result.

After The First Pass

Turn the concept into an inspected object.

Once the invariant is visible in the intuition, math, code, and demo, use these panels to inspect the mechanism visually, check source support, practice the idea, and attach a grounded research question.

Mechanism Storyboard

See the idea move before the page explains it

Evaluate whether an adaptive learning policy should ship by inspecting logged propensities, IPS/SNIPS/DR estimates, holdout evidence, guardrails, coverage, and promotion thresholds.

Prediction open01 / Intuition

Prediction lens

Start with the picture, metaphor, or geometric mechanism.

Commit first

Before reading further, choose the kind of change Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion should make visible.

Visual Inquiry

Make the image answer a mathematical question

Evaluate whether an adaptive learning policy should ship by inspecting logged propensities, IPS/SNIPS/DR estimates, holdout evidence, guardrails, coverage, and promotion thresholds.

4/4 stages readyLive demo connected

Visual cueWhich visible object should carry the first intuition?

Inspection depth2/4

Prediction

Which visible object should carry the first intuition?

Commit first

Pick the cue that should make Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion easier to reason about before the page gives the answer.

Source Grounding

Canonical references for the mechanism on this page.

paper · 2011Unbiased Offline Evaluation of Contextual-bandit-based News Article Recommendation AlgorithmsLi, Chu, Langford, and Wang

Primary support for replay/offline evaluation from randomized logged bandit data.

Open source

paper · 2011Doubly Robust Policy Evaluation and LearningDudik, Langford, and Li

Primary support for combining reward models with importance weighting in doubly robust policy evaluation.

Open source

paper · 2013Counterfactual Reasoning and Learning Systems: The Example of Computational AdvertisingBottou et al.

Support for treating system changes as counterfactual questions over logs, interventions, and policy consequences.

Open source

paper · 2009Controlled Experiments on the Web: Survey and Practical GuideKohavi, Longbotham, Sommerfield, and Henne

Support for randomized online experiments, A/B testing practice, metrics, and trustworthy promotion decisions.

Open source

paper · 2010A Contextual-Bandit Approach to Personalized News Article RecommendationLi, Chu, Langford, and Schapire

Support for contextual bandit policy learning and offline evaluation using recorded random traffic.

Open source

Claim Review

Evaluate whether an adaptive learning policy should ship by inspecting logged propensities, IPS/SNIPS/DR estimates, holdout evidence, guardrails, coverage, and promotion thresholds.

Status2 substantive reviews recorded

Claims without a substantive review badge still need exact source-support review.

Sources5 references

li-2011-unbiased-offline-evaluation, dudik-2011-doubly-robust, bottou-2013-counterfactual-learning-systems, kohavi-2009-controlled-experiments, li-2010-contextual-bandit-news

Local checks4 local checks

Use equations, runnable code, and demos to check whether the source support is operational.

Substantively reviewedLearning-system policy evaluation needs logged context, action, reward, and propensity before replay, IPS, SNIPS, or doubly robust estimates can be trusted.Claim metadata: source checked

The references support logged propensities, partial-label/counterfactual evaluation, replay-style evaluation, importance weighting, and model-assisted doubly robust estimates.

Sources: Unbiased Offline Evaluation of Contextual-bandit-based News Article Recommendation Algorithms, Doubly Robust Policy Evaluation and Learning, A Contextual-Bandit Approach to Personalized News Article Recommendation, Counterfactual Reasoning and Learning Systems: The Example of Computational AdvertisingThe page teaches finite logged-policy estimators for a contextual learning-policy setting; it does not prove fairness, long-term retention, or full RL policy value.A bounded review summary is present; still check caveats and exact reference scope.

Checked replay/offline contextual-bandit evaluation, doubly robust evaluation, contextual-bandit logged-random-traffic framing, and counterfactual learning-system reasoning. Oracle/GPT Pro review remains pending because the correct Chrome/Oracle lane is not attachable from this workstation.

Reviewer: codex-local-primary-reference-audit; reviewed 2026-07-05

Equation 1 Equation 2 Runnable code 1 Interactive demo

Substantively reviewedA candidate curriculum policy should not be promoted from an offline estimate alone; online holdout evidence, guardrail metrics, coverage checks, and clear promotion thresholds must be inspected separately.Claim metadata: source checked

The references support randomized holdouts, counterfactual caution, metric inspection, and the need to separate evaluation evidence from deployment decisions.

Sources: Controlled Experiments on the Web: Survey and Practical Guide, Counterfactual Reasoning and Learning Systems: The Example of Computational Advertising, Unbiased Offline Evaluation of Contextual-bandit-based News Article Recommendation AlgorithmsThe demo uses small synthetic logs and simple intervals; it is not a production experimentation platform, ethics review, accessibility audit, or causal guarantee.A bounded review summary is present; still check caveats and exact reference scope.

Checked randomized controlled experiment practice for online treatment/control comparisons, counterfactual-learning-system caveats for policy changes, and bandit offline-evaluation limitations around overlap and logged propensities. The local demo exposes promotion as a separate gate from estimator ranking.

Reviewer: codex-local-primary-reference-audit; reviewed 2026-07-05

Equation 2 Runnable code 1 Interactive demo

Source support candidates

paper 2011Unbiased Offline Evaluation of Contextual-bandit-based News Article Recommendation Algorithms

Primary support for replay/offline evaluation from randomized logged bandit data.

paper 2011Doubly Robust Policy Evaluation and Learning

Primary support for combining reward models with importance weighting in doubly robust policy evaluation.

paper 2013Counterfactual Reasoning and Learning Systems: The Example of Computational Advertising

Support for treating system changes as counterfactual questions over logs, interventions, and policy consequences.

paper 2009Controlled Experiments on the Web: Survey and Practical Guide

Support for randomized online experiments, A/B testing practice, metrics, and trustworthy promotion decisions.

Mechanism checks

Equation 1

q_i = q(a_i \mid x_i).

Equation 2

w_i = \frac{\pi(a_i \mid x_i)}{q(a_i \mid x_i)}.

Runnable code 1events = [ {"reward": 0.62, "propensity": 0.50, "target_prob": 0.50, "target_model": 0.58, "l...Demo stateLive mechanism probe

Practice Loop

Try the idea before it explains itself

Evaluate whether an adaptive learning policy should ship by inspecting logged propensities, IPS/SNIPS/DR estimates, holdout evidence, guardrails, coverage, and promotion thresholds.

Readiness0/3 checks ready

Predict

Before touching the demo, predict one visible change that should happen in Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion.

Hint 1

Reveal when your model needs a nudge.

Hint 2

Reveal when your model needs a nudge.

Hint 3

Reveal when your model needs a nudge.

Your answer canvas

Local checks

Claim

A concrete answer is on the canvas.

Mechanism

The answer names why the claim should hold.

Bridge

It touches the page context or a neighboring idea.

Misconception check

Grounded research drawerClose

ConceptLearning System Policy Evaluation: Logs, Estimators, Guardrails, PromotionMachine Learning

Runnable code comparisonLearning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion runnable code 1events = [ {"reward": 0.62, "propensity": 0.50, "target_prob": 0.50, "target_model": 0.58, "l...Prediction before revealLearning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion interactive...Manipulate one control and predict the visible change.

Grounded room questionWhat is the smallest example that makes Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion click without losing the math?Local snapshot ready

Research Room

Attach the question to a claim, equation, code, or demo

Pick the concept, equation, source, runnable code, claim, misconception, or demo state before asking for help. The handoff keeps that page item in context.

Next local actionNo local draft saved yet

Open the draft below to save one note and next action in this browser.

conceptMachine Learning

Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion

Attached question

What is the smallest example that makes Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion click without losing the math?

Local action draftNo local draft saved yetExpand only when ready to capture one local next action

Local action draft

This draft stays in this browser, attached to the selected learning item.

Draft noteNext action

No local draft saved.

Evidence to inspect

References to inspect: attached references on this page.
Definition, prerequisite, and contrast concept links
The equation or runnable code that makes the concept operational
One demo state that shows the invariant instead of a slogan

What would resolve this

The learner can state the mechanism in their own words
The learner can name the prerequisite that would repair confusion
The learner can predict how the mechanism changes under one perturbation

Grounded AI handoff

I am working in Continuous Function's research reading room. Object: concept - Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion Selected item key: recorded for copy. Context: Machine Learning Page anchor: recorded for copy. Open question: What is the smallest example that makes Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion click without losing the math? Evidence to inspect: - References to inspect: attached references on this page. - Definition, prerequisite, and contrast concept links - The equation or runnable code that makes the concept operational - One demo state that shows the invariant instead of a slogan What would resolve this: - The learner can state the mechanism in their own words - The learner can name the prerequisite that would repair confusion - The learner can predict how the mechanism changes under one perturbation Answer as a careful research tutor: stay source-grounded, separate verified evidence from assumptions, name the relevant math objects, and end with one next action.

View it in context

concept/concept-notebook/machine-learning/learning-system-policy-evaluation
concept:machine-learning/learning-system-policy-evaluation

Learning Map

Before / Now / Try / Next

BeforeAdaptive Curriculum Optimization: Policy, Utility, Exploration, Audit (review)

NowIntuition -> Math -> Code -> Demo

TryManipulate one control and predict the visible change.

NextLearner Outcome Experiment Design: Outcomes, Power, Retention, Guardrails (review)

Intuitionready
Mathready
Codeready
Interactive Demoready

AI Companion

Ask beside the selected object

Evaluate whether an adaptive learning policy should ship by inspecting logged propensities, IPS/SNIPS/DR estimates, holdout evidence, guardrails, coverage, and promotion thresholds.

Your question

GoalComfortStyleStuck on

Context prompt

You are my AI learning companion for Continuous Function. Current context: Machine Learning concept. Learning surface: Learning System Policy Evaluation: Logs, Estimators, Guardrails, Promotion. What this page says: Evaluate whether an adaptive learning policy should ship by inspecting logged propensities, IPS/SNIPS/DR estimates, holdout evidence, guardrails, coverage, and promotion thresholds. Current section: Intuition, math, code, and interactive demo. Suggested next step: Manipulate one control and predict the visible change.. Learner goal: Understand the idea. Learner comfort level: New to this. Preferred explanation style: Visual first. Task: Explain the central idea in plain language, then restate it with the exact math objects from the page. Answer in a way that helps me learn: ask one clarifying question only if needed, use intuition before notation, and end with one thing I should try on the page.

Domain

Machine Learning

machine-learningadaptive-learningpolicy-evaluationcounterfactual-estimationexperimentsguardrails

Prerequisites

Adaptive Curriculum Optimization: Policy, Utility, Exploration, Audit (review)Evaluation Pipelines Causal Inference Basics: Correlation versus Intervention (review)

Leads To

Learner Outcome Experiment Design: Outcomes, Power, Retention, Guardrails (review)

Adaptive Curriculum Optimization: Policy, Utility, Exploration, Audit (review)Evaluation Pipelines Causal Inference Basics: Correlation versus Intervention (review)Bandits (review)

Within this domain