FOUNDATION PROJECT · EVIDENCE AND TRANSFER

Diagnose, repair and report a training failure.

Make a comparison you can explain, record an artifact, then solve new assessment cases. Choose the kind of evidence that fits your goal.

Saved on this device with My Work export/import. Playground values are temporary until captured. Report submission, objective assessment and human review are separate. Learner pilots remain pending.

Read the full lesson Open experiment Write report Try assessment directly

Your brief

A colleague says a flat loss means the network is too small. Test that explanation before increasing the model. Record a matched failure and repair, then turn the trace into a reproducible incident report.

Run Healthy start for 100 accepted updates and capture it.
Run Blocked gradient for 100 accepted updates with the same optimizer, rate and scale; capture it.
Compare both losses and the gradient/update path. Optionally capture a saturated or divergent run without hiding its failure.

Project evidence variant

Rubric: trace one calculation, identify the intervention, compare outcomes, and state a limitation. The report is learner-authored; the separate transfer check scores specific answers.

Run and capture evidence

Loading local workbench…

Captured evidence (0/4)

Scores are recalculated from your saved experiment. This is a learner record, not independent proof of how the experiment was run. Removing a run reopens the report submission but keeps assessment history.

No evidence captured yet. Use Capture run for project after a completed experiment.

Required comparison: not yet captured. This checks the experiment structure, not the quality of your reasoning.

Your experiment report

Symptom and diagnosisState the failure and the diagnostic that distinguishes your explanation from at least one alternative.Controlled repair and resultsCompare matched healthy and blocked runs. State optimizer, rate, scale, accepted updates, training and validation loss. Explain the actual update mechanism.Limits and acceptance decisionState what the synthetic curve cannot establish. For a decision brief, specify a repair acceptance criterion and what requires new validation or human review.

Submission needs the required comparison and nonempty report fields. Presence checks do not assess writing, correctness of the artifact, or decision quality. Any edit reopens the submission for review.

Back up all work

Independent transfer assessment

These new cases differ from the guided flower and neural-network examples. Work unaided: use paper or a calculator, and submit all answers before viewing feedback. Numerical answers accept a tolerance of 0.000001. Each case has four criteria; all four must pass.

First submissions measure this objective check. Up to two retries are labelled practice and cannot replace the first result. Case exposure and submissions are saved locally and retained when importing older backups. Client-side cases are inspectable; this is not a proctored credential or human review of reasoning.

First-attempt cases passed: 0 / 2. Latest attempts passed: 0 / 2. Written report and implementation review: pending.