version: 2 article_bundle: content/curriculum/003-checkpoint-resume-equivalence article: index.md visuals: - path: cover.svg type: cover_svg asset_type: cover_svg purpose: Establish the recurring split-timeline object and show that a successful load can still branch on the next update. alt_text: Two training timelines coincide only when the checkpoint carries every input needed by the next transition. caption: A checkpoint is tested by the future transition it reproduces, not by whether deserialization succeeds. provenance_or_license: Original local SVG generated for this article; no third-party visual content; copyright Anshuman Sahoo. generation_method: Generated by code/generate_visuals.py from the experiment schema and editorial contract. source_data_path: data/checkpoint-schema.json article_section: Opening units: optimizer-step position and qualitative branch path assumptions: - The uninterrupted path is the reference computation. - Branch distance is illustrative and does not encode loss or parameter-error magnitude. accessibility: title_and_desc_present: true color_not_only_carrier: true - path: figures/terminal-01-fixture.svg type: terminal_transcript_svg asset_type: runtime_and_model_terminal_snapshot purpose: Show the actual model, parameter shapes, runtime, deterministic settings, training configuration, and checkpoint payload keys before interpretation begins. alt_text: Actual terminal transcript showing Python and PyTorch versions, the TinyDropoutClassifier architecture, parameter shapes and counts, optimizer and scheduler settings, and boundary versus mid-accumulation checkpoint keys. caption: The inspect command exposes the exact fixture used by every intervention rather than asking the reader to infer it from prose. provenance_or_license: Original local SVG rendered from an actual retained command transcript generated in the packaged environment; no third-party visual content; copyright Anshuman Sahoo. generation_method: Command output captured in data/terminal-01-fixture.txt and rendered without content changes by code/generate_terminal_snapshots.py. source_data_path: data/terminal-01-fixture.txt article_section: Build the fixture before naming the state units: runtime versions, tensor shapes, parameter counts, configuration values, checkpoint key names assumptions: - The transcript reflects the packaged CPU environment at revision time. - SVG styling changes presentation only; transcript text is retained verbatim. accessibility: title_and_desc_present: true color_not_only_carrier: true - path: figures/terminal-02-seed-11-matrix.svg type: terminal_transcript_svg asset_type: single_seed_intervention_terminal_snapshot purpose: Walk through all seven interventions on one representative seed before presenting aggregate results. alt_text: Actual terminal table for seed 11 showing whether each scenario resumed exactly, its first divergent optimizer step, whether the next batch matched, learning-rate error, and final maximum parameter error. caption: Seed 11 exposes two exact controls and five incomplete checkpoint signatures in one terminal run. provenance_or_license: Original local SVG rendered from an actual retained command transcript generated by the packaged harness; no third-party visual content; copyright Anshuman Sahoo. generation_method: Command output captured in data/terminal-02-seed-11-matrix.txt and rendered without content changes by code/generate_terminal_snapshots.py. source_data_path: data/terminal-02-seed-11-matrix.txt article_section: Run the control, then remove one field at a time units: exactness flag, optimizer step, batch identity, learning-rate error, maximum absolute parameter error assumptions: - Seed 11 is used for walkthrough clarity, while aggregate claims use all five seeds. - The table reports final error and first model divergence as produced by the harness. accessibility: title_and_desc_present: true color_not_only_carrier: true - path: figures/terminal-04-tests.svg type: terminal_transcript_svg asset_type: pytest_terminal_snapshot purpose: Show the direct correctness, equivalence, counterexample, and walkthrough-interface tests executing successfully. alt_text: Actual terminal transcript showing twenty-one Pytest checks passing for the checkpoint equivalence harness. caption: The tests cover complete recovery, planted omissions, delayed scheduler divergence, data-versus-RNG diagnosis, and the inspection commands shown in the article. provenance_or_license: Original local SVG rendered from an actual retained Pytest transcript generated in the packaged environment; no third-party visual content; copyright Anshuman Sahoo. generation_method: Pytest output captured in data/terminal-04-tests.txt and rendered without content changes by code/generate_terminal_snapshots.py. source_data_path: data/terminal-04-tests.txt article_section: Run the control, then remove one field at a time units: test progress and pass count assumptions: - The transcript corresponds to the packaged test file and code revision. - A passing suite supports only the claims encoded by those tests. accessibility: title_and_desc_present: true color_not_only_carrier: true - path: figures/terminal-03-scheduler-trace.svg type: terminal_transcript_svg asset_type: scheduler_differential_trace_terminal_snapshot purpose: Show the delayed causal sequence in which the scheduler learning rate differs one step before model parameters diverge. alt_text: Actual terminal trace for seed 11 with scheduler state omitted. Control and resumed learning rates differ at optimizer step 8 while model parameters still match; parameters first differ at step 9. caption: The scheduler trace demonstrates why a healthy checkpoint load and first resumed update are insufficient tests. provenance_or_license: Original local SVG rendered from an actual retained command transcript generated by the packaged harness; no third-party visual content; copyright Anshuman Sahoo. generation_method: Command output captured in data/terminal-03-scheduler-trace.txt and rendered without content changes by code/generate_terminal_snapshots.py. source_data_path: data/terminal-03-scheduler-trace.txt article_section: Follow the first branch, not the successful load units: optimizer step, learning rate, learning-rate error, batch comparison, model comparison, maximum absolute parameter error assumptions: - The trace is for seed 11 and the omit_scheduler intervention. - Scheduler stepping occurs after optimizer stepping in the packaged loop. accessibility: title_and_desc_present: true color_not_only_carrier: true - path: figures/figure-02-intervention-traces.svg type: explanatory_svg asset_type: failure_trace_diagram purpose: Show the first optimizer step at which each omission changed the representative resumed run. alt_text: Representative seed 11 paths show RNG and data omissions branching at step 6, scheduler omission at step 9, and complete recovery overlapping the control. caption: Different omissions can produce different branch timing, which turns the trace into a diagnostic rather than a binary pass or fail. provenance_or_license: Original local SVG generated for this article; no third-party visual content; copyright Anshuman Sahoo. generation_method: Generated by code/generate_visuals.py from data/step-traces.json. source_data_path: data/step-traces.json article_section: Follow the first branch, not the successful load units: optimizer step; vertical distance is categorical branch visibility assumptions: - Seed 11 is representative of branch timing across all five seeds. - Vertical displacement does not encode numerical error magnitude. accessibility: title_and_desc_present: true color_not_only_carrier: true - path: figures/figure-03-results.svg type: explanatory_svg asset_type: results_chart purpose: Summarize exact-run counts and median final parameter error across all 35 scenario-seed runs. alt_text: Complete boundary and mid-accumulation payloads match in five of five seeds; every planted omission matches in zero of five seeds. caption: Complete payloads reproduced the control in this deterministic fixture; every selectively incomplete payload failed. provenance_or_license: Original local SVG generated for this article; no third-party visual content; copyright Anshuman Sahoo. generation_method: Generated by code/generate_visuals.py from data/results.csv and data/run-summary.json. source_data_path: data/results.csv article_section: Follow the first branch, not the successful load units: exact runs out of five; median maximum absolute parameter error on a log scale assumptions: - Exact means all compared model, optimizer, scheduler, stream, RNG, gradient, and next-batch checks pass. - Error magnitudes are fixture-specific and are not estimates of production convergence impact. accessibility: title_and_desc_present: true color_not_only_carrier: true - path: figures/figure-01-state-surface.svg type: explanatory_svg asset_type: state_surface_diagram purpose: Locate the update, stochastic, data, and partial-step components that the fixture selectively saves or omits. alt_text: Four panels list update state, stochastic state, data state, and partial-step state that may cross an interruption boundary. caption: Boundary checkpoints and mid-accumulation checkpoints have different payload requirements because only the latter contains unfinished gradient work. provenance_or_license: Original local SVG generated for this article; no third-party visual content; copyright Anshuman Sahoo. generation_method: Generated by code/generate_visuals.py from data/checkpoint-schema.json. source_data_path: data/checkpoint-schema.json article_section: Explain the failures from the checkpoint payload units: named state components assumptions: - The component list is specific to the packaged CPU fixture. - Distributed and mixed-precision systems may own additional state. accessibility: title_and_desc_present: true color_not_only_carrier: true