# Historical benchmark provenance

These are the result and figure artifacts checked into the repository at
`8793c49ee4f9f8b07649e8db6571ed53746a9a06`. A Git commit below identifies
when an artifact entered the repository; it does **not** establish the commit
used to execute the run. No artifact is rewritten here. The vision benchmark
code used final test data for early stopping and sleep rollback, and its heads
and backbones were unmatched. Treat its rankings as legacy/test-informed until
the corrected protocol in [DEVELOPMENT_PLAN.md](../DEVELOPMENT_PLAN.md) is run.

| Artifact | Recoverable provenance | Unknowns and interpretation |
|---|---|---|
| [benchmark_pareto_hard_results.json](../benchmark_pareto_hard_results.json) | Added `6ecfcff`. JSON declares synthetic hard images, noise 0.08, 2,500 train/700 test, 10 classes, 96 px, CUDA, 20 epochs, and seeds 7/13/29. It contains 10 backprop, 12 PC, and 12 circadian trials with three seed reports per trial. `scripts/run_pareto_hard_tuning.py` writes this path and at that commit sets an ImageNet backbone, batch size 64, and further base settings. | Full resolved config, pretrained-weight hash, actual execution commit, GPU model, dependency versions, and dataset identities are absent. Its test-accuracy-based selection is legacy/test-informed. |
| [benchmark_circadian_policy_sweep_results.json](../benchmark_circadian_policy_sweep_results.json) | Added `6ecfcff`. JSON has 18 trial overrides and declares synthetic hard images, noise 0.08, 2,500/700 samples, 10 classes, 96 px, CUDA, ImageNet weights, and **14 epochs**. `scripts/run_circadian_policy_sweep.py` writes this path and sets seed 7. | The producer script already specified **20 epochs** at `6ecfcff`, so the exact executing config/code is unknown. Seed 7 is inferred from the script, not recorded in the JSON. Test-ranked, legacy/test-informed. |
| [benchmark_tuning_hardest_results.json](../benchmark_tuning_hardest_results.json) | Added `6ecfcff`. It contains 6/8/10 backprop/PC/circadian trials; reports show 14 epochs, synthetic hard images, noise 0.08, 2,500/700 samples, 10 classes, 96 px, and CUDA. | No checked-in producer script, seed list, backbone identity, execution commit, or complete configuration was found. |
| [benchmark_pareto_hard_summary.md](../benchmark_pareto_hard_summary.md) | Added `6ecfcff`; reports 14 epochs and best accuracy about 0.2157. | It is **not** a summary of the checked-in Pareto JSON, which reports 20 epochs and best accuracy about 0.9524. It also does not match the tuning JSON's best accuracy of 0.19. Its source data are unknown. |
| [benchmark_master_cifar100_subset_2026-02-28.txt](benchmarks/benchmark_master_cifar100_subset_2026-02-28.txt) | Added `77c0d07`. Output states CUDA, CIFAR-100, 20,000 train/5,000 test subset, 96 px, batch 32, augmentation, and 12 epochs. README gives a candidate command with ImageNet weights and a frozen backprop backbone; it describes a single-seed result. | Seed 7 is inferred from the CLI default rather than embedded. Exact execution commit, subset IDs, weight hash, hardware/version, and run manifest are absent. Test-informed stopping/rollback applies to the source benchmark. |
| [benchmark_continual_shift_strength_case_2026-02-28.txt](benchmarks/benchmark_continual_shift_strength_case_2026-02-28.txt) | Added `ed569aa`; output embeds seeds 3/7/11/19/23/31/37, 40° shift, translation (0.9, −0.7), and phase-B training fraction 0.14. The same-revision `strength-case` profile defaults to 500+500 samples, 110+80 epochs, width 12, phase noise 0.8/1.0, and sleep intervals 40/8. | Exact command and per-seed raw results are absent. It is a tuned historical profile; independent confirmation provenance is unknown. |
| [benchmark_continual_shift_hardest_case_2026-02-28.txt](benchmarks/benchmark_continual_shift_hardest_case_2026-02-28.txt) | Added `54c0247`, then updated at `6af5cda` and `8793c49`. Current text embeds seeds 3/7/11/19/23/31/37; 700+700 samples; widths (24, 24, 24); 120+180 epochs; phase noise 0.8/1.45; 68° shift; translation (1.6, −1.3); and phase-B fraction 0.05. The `8793c49` profile specifies sleep intervals 40/6. | Exact command and per-seed raw results are absent. The profile changed across commits; use the final text with the final profile only. Independent confirmation provenance is unknown. |

## Figures and dashboard

| Files | Recoverable source | Limitation |
|---|---|---|
| [benchmark_accuracy.png](figures/benchmark_accuracy.png), [benchmark_train_speed.png](figures/benchmark_train_speed.png), [benchmark_inference_latency_p95.png](figures/benchmark_inference_latency_p95.png) | Added `7289dbb`; generated by `scripts/generate_readme_figures.py` from `benchmark_multiseed_cifar100_summary.csv`. | The source CSV and per-seed CSV/JSON are not checked in or present in this checkout. README documents a three-seed runner command, but the chart input and exact run settings cannot be verified. |
| [interactive_benchmark_accuracy.html](figures/interactive_benchmark_accuracy.html), [interactive_benchmark_train_speed.html](figures/interactive_benchmark_train_speed.html), [interactive_benchmark_inference_latency_p95.html](figures/interactive_benchmark_inference_latency_p95.html) | Added `42a4aa5`; interactive counterparts to the CSV-derived charts. | Embedded aggregate values do not recover the missing source CSV or per-seed records. |
| [benchmark_overview_compact.png](figures/benchmark_overview_compact.png), [interactive_benchmark_overview.html](figures/interactive_benchmark_overview.html) | Added `87674fc`; also produced from the missing summary CSV. | Source run and configuration are not independently verifiable. |
| [circadian_sleep_dynamics.gif](figures/circadian_sleep_dynamics.gif) | Added `7289dbb`; `scripts/generate_readme_figures.py` defaults to width 384, 24 cycles, 12 splits, and 2 prunes. | `docs/figures/README.md` calls it illustrative. It is not observed sleep telemetry. |
| [hardest_mode_dynamics.gif](figures/hardest_mode_dynamics.gif) | Added `0179a99`; the generator at that commit defaulted to seed 7, 500+500 samples, width 8, 90+120 epochs, phase-B noise 1.2, and 44° rotation. | It predates the current 700+700, three-hidden-layer hardest-case profile. The exact invocation was not saved. Intermediate frames scored phase-B test labels, so this is a legacy/test-informed visualization. |
| [interactive_hardest_mode_dynamics.html](figures/interactive_hardest_mode_dynamics.html) | Added `b4ea123`, when the generator still used the earlier 500+500, width-8, 90+120 profile. | It predates the current hardest-case profile. The exact invocation was not saved. Intermediate frames scored phase-B test labels, so this is a legacy/test-informed visualization. |
| [docs/index.html](index.html) | Added `42a4aa5` as a dashboard linking/embedding figures. | It is presentation, not an independent experiment record. |

There are **no checked-in CSV files**. The missing
`benchmark_multiseed_cifar100_summary.csv` prevents regeneration of the chart
set from the repository alone. Across historical artifacts, execution commits,
resolved configs, model and split hashes, dependency versions, and exact
hardware are not consistently recoverable. Preserve these files as historical
evidence; new reports should identify their protocol and full provenance
separately.