If I want to predict noise for a wing shape under an unfamiliar condition, it matters whether the model has already seen another frequency from that condition. A test row can be absent from training while closely related measurements are still present. The public dataset has no experiment IDs, so I used four recorded non-frequency features as condition proxies and kept that limitation visible.
I built two training pools with the same 843-row budget and the same validation and test rows. Pool A sees none of the test proxies. Pool B sees other frequencies from all 27 test proxies, replacing 185 rows from A's pool. I trained three fixed seeds for each pool. This lets a reader inspect the change in task, while retaining the simultaneous changes in coverage and training normalizers.
B's mean test MAE across the three seeds was lower: about 2.32 dB versus A's 2.64 dB. The first displayed case goes the other way for the default seed. Across the 603 paired comparisons, B has larger absolute error in 237. All of them remain in the package; the three seeds reuse the same 201 test cases.
I kept the worse default example because a lower aggregate score does not tell a reader which individual predictions improved. The app shows training exposure and exports all six model outputs together. Since replacing training rows also changes coverage and normalization, this experiment cannot isolate a causal effect of overlap alone. It does not establish physical experiment independence or predict engineering noise compliance.
The six original 120-epoch CPU fits and independent rerun were completed on November 4, 2026. This release reuses the frozen models and all results. It adds no new training, test sample or claim of adoption. The data and targets were already inspected before the protocol.
假设我想预测一种陌生翼型条件下的噪声,模型有没有见过这个条件的其他频率,会影响这个问题的含义。测试那一行没进训练,不代表相关测量都没进去。原数据没有实验编号,我只能用四个非频率特征组成条件代理,并把这个缺口保留。
我做了两组各843条的训练池,共用验证和测试记录。A组没见过测试条件代理,B组见过全部27个测试代理的其他频率,同时替换了A组里的185条记录。每组用三个固定种子训练。这样可以检查任务如何改变,也能看见覆盖和归一化统计一起变化的事实。
B组三个种子的平均测试MAE约2.32 dB,低于A组约2.64 dB。但默认种子的第一个展示案例,B反而更差。603次配对比较里有237次如此,全部保留;三个种子反复使用的是同一批201个测试案例。
我保留这个更差的默认案例,让读者能从平均分往下看具体预测。页面同时显示训练条件是否见过,导出保留六个模型的全部输出。替换训练记录也改变了覆盖和归一化,不能把差异全算成重叠带来的单独因果效应,更不能据此声称真实实验独立或工程噪声合规。
原六次120轮CPU训练和独立重跑在2026年11月4日完成。本次发布沿用冻结模型与完整结果,不是今天又训练了六次,也没有新测试样本或真实采用。原数据及目标值在制定协议前已经看过。下面保留完整划分、单位、失败和许可说明。
Two training pools have the same 843-row budget and the same validation/test cases. One pool never sees the test condition proxies; the other sees different frequencies of those conditions. Compare their fixed model predictions, inspect training exposure, and download all six outputs with complete input and provenance.
This is one experiment with six actually trained networks, not six projects.
Observed condition tuples are not verified experiment IDs. The study does not
establish physical independence, a causal effect of overlap alone, or validated
acoustic predictions. Work identity: work:airfoil-split-lab:v1.
From this directory run python3 -m http.server 17891 --bind 127.0.0.1, then open
http://127.0.0.1:17891/. A modern browser needs JavaScript, WebCrypto and Blob
downloads. File URLs are unsupported. Browser inference needs no Python model
dependency and makes no model API call.
The fixed first common-test source row 29 loads only if editing has not started. Enter the five unit-labelled fields or the exact-key JSON object below, select paired seed 17/41/73, and Compare. Seed 17 is the predeclared default, not a choice made from the test scores. Changing inputs or seed clears old results and links.
{
"frequency_hz": 200,
"attack_deg": 0,
"chord_m": 0.3048,
"velocity_m_s": 39.6,
"thickness_m": 0.00310138
}
All five values must be finite real numbers within the displayed original per-feature ranges; booleans, null, strings, missing/extra keys, blank fields, nonfinite values, BOM and JSON exceeding 4,096 UTF-8 bytes reject. A blank angle is not silently converted to 0. Angles remain real-valued. Frequency may be a positive real within bounds; an arbitrary frequency/tuple combination is not thereby validated by the source data. No units are converted automatically.
JSON edits are authoritative until valid comparison updates the numeric fields. An explicit field edit switches back to the fields and updates the mode message. Invalid JSON never falls back to a previous prediction. The report's input hash uses canonical numeric values, not the original JSON text bytes; JSON 1/1.0 and negative-zero/zero are the same numeric value. SHA checks consistency, not identity or authenticity of a user-supplied experiment.
The two visible cards use the selected paired seed. The report always retains all six point outputs, baselines, selected epochs, five input values/units, condition lookup, complete training normalizers, source/model/catalog hashes and attribution. Inputs stay in the tab; refreshing loses edits. The app adds no upload, storage, cookies, telemetry or external fonts. Browser extensions and hosting have their own behavior; this is not a universal privacy guarantee. Full reports contain your parameter values. User inputs keep their original terms.
The original 1,503 source rows contain 106 exact tuples of the four non-frequency features: angle, chord, velocity and displacement thickness. These are label-blind condition proxies. UCI does not supply run IDs or an official train/test split. The original features and targets were audited before the protocol, so this is not a historically blind study.
Canonicalize original decimal tokens without unnecessary trailing zeros or exponents, with zero represented as 0. SHA256(condition key)%10 selects 59 base groups (remainders 4–9),20 validation groups (2–3),27 test groups (0–1). Within each test group sort by frequency then original ordinal: even ranks give 201 common test rows and odd ranks give 185 context rows. Validation uses all 274 rows in its groups.
Pool A uses all 843 base rows. Pool B uses all 185 context rows and 658 base rows,
ranking base rows by SHA256(row-context-replace: + full feature key), tie by
original ordinal. Thus 185 base rows are displaced. Both pools share 658 rows and use
the same 843 training-row budget, 274 validation rows and 201 test rows. The full
membership file preserves all 1,503 rows, including A's unused 185 context rows and
B's unused 185 base rows. No identical test row is in either fit.
Pool B sees other frequencies of all27 test proxies; A sees none. B's numbers describe a seen-condition task and do not estimate unseen-condition performance. Replacing 185 training rows also changes coverage and fitted normalizers, so the metric difference is not an isolated causal effect of overlap. Unrecorded source dependencies may remain even between different tuples. The catalog lookup checks exact source tuples, not real-world OOD safety or experimental independence.
Every test result below uses all 201 common rows. Error is in the original target's scaled sound-pressure dB units, not dBA, pressure or energy. The three seed comparisons reuse the same test cases and are not 603 independent examples.
| Seed | A MAE / RMSE dB | B MAE / RMSE dB | B smaller / equal / larger absolute error |
|---|---|---|---|
| 17 | 2.637440 / 3.433836 | 2.268305 / 2.979579 | 115 / 0 / 86 |
| 41 | 2.685900 / 3.475715 | 2.301926 / 2.995912 | 135 / 0 / 66 |
| 73 | 2.610386 / 3.387189 | 2.399225 / 3.197637 | 116 / 0 / 85 |
Arithmetic mean of the three seed MAEs: A 2.644575 dB, B 2.323152 dB. This is a mean
of metrics, not an ensemble prediction, confidence band or significance estimate.
All 237 paired rows where B has larger absolute error remain in
evaluation-001/row-context-worse.csv. The default recorded case itself is worse
under B for seed 17; it was selected by source ordinal before model evaluation.
No winning seed or favorable case replaces the complete denominator.
Two train-mean baselines have test MAE 5.853341/5.871245 dB. They are fitted once per
pool, not six additional neural fits. predictions.csv preserves all 3,800 method
predictions: six networks and two means on 274 validation+201 test rows. Full 392
overall/condition metrics,603 paired rows and the complete 1,503-row membership are
retained. Micro test MAE uses all 201 rows; macro-proxyMAE separately averages 27 group
MAEs equally and is not silently substituted. Both are in results.json.
split-study.png shows all six 120-epoch validation curves, selected epochs,
six test MAEs, all603 paired error differences and condition exposure. The separate
six-panel figure shows every 1,206 neural test prediction without removing failures.
Both full-size figures are linked in the app.
Python 3.14.7 / NumPy 2.5.4 CPU float64. Each network is 5 → 24 tanh → 1 linear, 169 parameters. Two pools × fixed seeds 17/41/73 give six actual fits. For a given seed the initial arrays are identical across pools; training-only input/target normalizers differ. All input means/population standard deviations, target mean/std and the constant baseline use only that pool's 843 training rows. Source-wide marginal ranges are descriptive input guards, not fitted normalizers or learned features.
Fixed Adam .003, coupled weight-only L2 .0001, batch 32 and 120 epochs. Each fit uses 3,240updates and 101,160 row visits; total 720 epochs / 19,440 updates / 606,960 visits. All three initial weight sets shared across paired seeds, 720 trace rows, selected weights, validation predictions and normalizers are saved. Earliest minimum common-validation MAE selects epochs A 116/110/118 and B 119/117/116. No test-based retuning, restarts or seed selection. PRETEST binds all selected artifacts/source/code before current-model test predictions; the evaluator and its PRETEST binding were separately frozen. Local hashes do not provide a third-party historical timestamp attestation.
Use a copy and new output directories; keep the frozen candidate unchanged.
Create a venv and install requirements.txt. Set PYTHONDONTWRITEBYTECODE=1.
python test_core.py
python train.py --out /absolute/new/run
Fresh timing changes PRETEST. Only in your review copy, bind EVALUATION-FREEZE.pretest_sha256 to the new run's PRETEST digest before evaluation, keeping the evaluator code digest unchanged and preserving before/after gates.
python evaluate.py --run /absolute/new/run --out /absolute/new/evaluation
python verify_results.py
python verify_portable.py
node test_inference.mjs
python verify_browser.py validation/actual-first-report.json
The supplied verifiers use bundled run-001 and evaluation-001. To verify fresh
artifacts, place them under those names in another isolated copy, preserving the
originals elsewhere. No verifier silently rewrites gates. Results are locally
reproducible in the recorded environment, not a universal cross-runtime byte
identity promise. Timed metadata excludes source/membership setup and is not
cold-start or general speed evidence.
The separate scalar oracle checks all 2,850 neural forwards, all 3,800 method rows, 392metrics,603 paired rows and237worse rows, with complete membership/normalizers. NumPy only decodes NPZ containers in that oracle; no author core math is imported. All 475 validation/test contexts × six portable JS models agree within 2.843e-14 dB. These checks are consistency evidence, not additional training or new scores.
Plots use headless MatplotlibAgg. To regenerate them, additionally install
requirements-plot.txt and put MPLCONFIGDIR outside a frozen candidate, then run
python plot_study.py. See VALIDATION.md for exact native UI scope.
Brooks, Pope and Marcolini (1989), Airfoil Self-Noise, UCI, https://doi.org/10.24432/C5VW2C, CC BY 4.0. The original ZIP/dat and unchanged values are in data/ with exact provenance. The UCI page describes NACA 0012 sections at varied flow conditions with the same observer position/span. Its variable table calls attack angle binary, while the actual file has 27 numeric angles; values are preserved as real numbers. No upstream metadata was changed.
NASA source record https://ntrs.nasa.gov/citations/19890016302 describes NASA-RP-1218 (1989), a semiempirical self-noise method. I read its record/abstract and metadata; this study does not reproduce the report's algorithm or validate aircraft, blade design, sound measurements or noise compliance. No paper, code or figures from that report are copied or relicensed here.
Original implementation/docs ©2026 Renchao Wu, Apache-2.0. Original data and derivative weights, normalizers, results and figures retain Brooks/Pope/Marcolini/ UCI CC BY 4.0 attribution. See LICENSES.md, NOTICE and data/SOURCE.json. Arbitrary user inputs keep their original terms. No employer, private or competition material was used. Local acceptance is not hosted HF operation, publication or adoption; later free Static/build/model/App/native input/report verification remains required.
这个工具比较两组同样843条训练记录、共用验证和测试样本的小模型。B组见过测试条件 下的其他频率记录,A组没见过这些条件;条件只是四个非频率特征组成的代理,不是真实 实验编号。B平均误差更低,但603条配对比较里仍有237条更差,原始失败和完整分母都保留。
六个模型来自两训练池×三个固定种子,不当六个作品。B组还替换掉185条其他记录,覆盖 和归一化统计都变化了,所以不能把差异完全归因于分组重叠,也不能把B的分数当全新条件 的泛化能力。输出单位是scaled SPL的dB,不是dBA、声压/能量降幅、置信区间或工程保证。
本地HTTP打开后可输入带单位的五个字段或JSON,选种子并对照;报告保留全部六个输出、 训练统计、输入/模型/源hash与署名。空字段不能变成0,坏JSON不能复用旧结果,修改会清 旧下载。刷新丢失输入,用户参数原条款不变。后续公开仍须独立重训验收及HF真实免费云端 App/模型/交互/原生下载核验。