# Results comparison: provenance and interpretation

Platform: **Privacy as Commodity**. Snapshot date: 2026-10-20.

The comparison page transcribes the current manuscript's reported tables.
It does not claim that the manuscript tables are the latest completed
experiment batch, independently verified empirical conclusions, or evidence
generated by the interactive website. Ongoing experiments are not read.

## Frozen source identity

| Source | SHA-256 |
| --- | --- |
| `main_final2/main_m_tdsc.tex` | `15cdbe8474d779b7241b94c8e0b823d1bb6154fa9c5bdc5be58cc47d57e4b2b0` |

Source paths are repository-relative. The website data are in
`results-data.js`, exported with `provenance: manuscript_table_snapshot`.
Changing a teaching example never changes these results. Downloading the
source data exports all snapshot tables and the current display selection.

### Historical cross-check: different versions

The older archive `main_final2/revisions/rq4_redo_20261014/results/rq4_summary.json`
(SHA-256 `e4a10b109381a18d7272e83c6388a360526cc31f7c3c9af14ee749012fd4d358`)
does **not** reproduce the current manuscript table exactly: 16 of the 36
accuracy mean/SD pairs differ at the displayed precision. All four settings
differ for MFG-RegretNet, RegretNet, DM-RegretNet, and Uniform-DP.
For example, the archive gives MFG-RegretNet MNIST alpha=0.5 as 94.71±0.78;
the current manuscript prints 94.65±0.80.

The archive is therefore **not used as display data**. All displayed numbers
are transcribed and checked directly against the current LaTeX table.
This version difference is recorded in the downloadable snapshot metadata.
The webpage does not resolve the experimental origin of subsequent manuscript
updates or combine the two versions into a new result.

## Accuracy: RQ4

- Source: `tab:rq4`, approximately lines 2784–2818 of the inspected manuscript.
- MNIST and CIFAR-10; Dirichlet alpha = 0.5 and 0.1.
- N = 10, financial budget B = 50, 100 FL rounds, three seeds.
- Final accuracy is the per-seed mean over rounds 96–100, then summarized
  as mean and standard deviation across seeds. Units are percent.
- Values on the website preserve the manuscript's displayed precision.
- For historical context, the older archived summary was generated by
  `main_final2/revisions/rq4_redo_20261014/exp/rq4_make_figs.py`.
  The table generator is `rq4_make_main_table.py` in the same directory.
- That historical generator takes mean per-round privacy and payment from the
  `MNIST_a0.1` summary. Its cumulative epsilon column is the minimum/maximum
  of median-client cumulative values over the four settings, rounded to one
  decimal. Since its summary differs from the current manuscript, it is not
  asserted to be the complete provenance of the current table. The website
  preserves the manuscript's single privacy/payment summary per row. These
  are **not four distinct per-setting privacy columns**.
- Private arms use DP-SGD and RDP accounting as described by the manuscript;
  the cumulative range uses delta = 1e-5. No-DP has no corresponding privacy
  or payment value and is rendered with a dash.
- PAC and VCG share a reported row because their releases coincide in this
  protocol. Their separate process pages do not imply independent accuracy
  measurements for that table.
- Uniform-DP matches a mean privacy resource, not individual trajectories.
  It has no compensation mechanism. No-DP is a learning reference rather
  than a privacy-preserving procurement alternative.

The point plot uses an explicitly labeled focused axis with mean ±1 SD.
SD is not a confidence interval. Differences between displayed means are
descriptive; the webpage does not compute significance tests or infer
causal effects of the mean-field input.

The snapshot does **not** support a claim that MFG-RegretNet has the highest
accuracy in every setting. For example, MNIST alpha=0.5 reports 94.65±0.80
for MFG-RegretNet, while MFG-Pricing and DM-RegretNet report 94.71±0.77.
The website preserves those numbers and their ordering.

## Incentive security: RQ1

- Source: `tab:rq1`, approximately lines 2561–2581, plus the adjacent RQ1
  paragraph for RegretNet (approximately lines 2623–2626).
- B = 25; N = 10; 10,000 held-out market profiles.
- Searched regret is shown in units of **10^-3 utility**. Reported seed
  standard deviations are preserved; missing SD is null, not zero.
- The RegretNet paragraph reports 1.75±0.12 and no IR violations but does
  not give its budget use there. The page therefore leaves budget use
  missing and labels that row's prose source.
- The reported MFG-RegretNet value is 2.43±0.15: it is higher than the
  RegretNet value in this snapshot, while both have zero reported IR
  violations. No universal incentive ranking is invented.
- Search outcomes are lower bounds on the supremum of deviation gain;
  this chart is not the separate certified upper-bound analysis.
- Report-domain and execution-utility assumptions matter. A found zero is
  not, by itself, a new proof of truthfulness.

## Budget and participant welfare: RQ3

- Source: `tab:rq3-efficiency`, approximately lines 2745–2767.
- B = 25, 50, 75, 100; 10,000 market rounds per budget.
- Welfare is the sum of participant utilities, sum(p_i - v_i*epsilon_i).
  It is not accuracy and does not include the buyer's utility.
- Budget-use ranges and maximum IR violation rates span the entire sweep;
  they do not become per-budget values when the selector changes.
- Cell-level SD is not given in the table. The plot supplies no fabricated
  error bars. The manuscript states the proposed method's SD is at most
  0.11 across the sweep; this is not copied into every cell as an exact SD.
- RegretNet and Lin are absent from the source table, so no welfare values
  are imputed for them.
- MFG-RegretNet has the largest reported participant welfare among the
  listed rows at all four budgets. This statement is limited to this table.

## Baseline source mapping

Each standalone page has a four-stage process, a synthetic numerical lab,
an explicit comparison of design choices, and a scope/source disclosure.

| Page | Numerical example / source |
| --- | --- |
| `pac.html` | Five-client implementation of `_pac_single` in `baselines/pac.py`; proposed, pre-check contracts. |
| `vcg.html` | Five-client implementation of `_vcg_single` in `baselines/vcg.py`; labeled budget-constrained VCG adaptation, not a generic VCG solver. |
| `csra.html` | Positive-cost one-shot `_csra_single` in `baselines/csra.py`; honesty history fixed at one; no dishonest-client detection. |
| `mfg-pricing.html` | Full-cap participation and payment B*epsilon_i/sum(epsilon), matching `baselines/mfg_pricing.py`. |
| `regretnet.html` | Head arithmetic for synthetic logits; per-client item/abstain softmax and N+1 payment softmax. No checkpoint is loaded. |
| `dm-regretnet.html` | Corrected item-plus-abstention argmax from `models_scaled.py`; not a claim to rerun the historical table's checkpoint. |
| `lin.html` | Five-type discrete menu using `baselines/lin2023/mechanism.py`, jointly scaled for B and cap ceiling; best feasible positive-surplus choice. |
| `uniform-dp.html` | Equal and heterogeneous allocations with exactly matched mean; arithmetic only, no privacy or accuracy inference. |
| `no-dp.html` | Synthetic scalar FedAvg with data-size weights, corresponding to `plain_round`; no training or privacy claim. |

Base synthetic client unit costs are [0.2,0.35,0.55,0.7,0.95], capacities
[0.5,0.8,1.2,1.6,2], and data sizes [80,160,100,240,420]. Relevant sliders
modify only the quantity described on their page. Numerical labs do not
write to any source experiment or manuscript.

The Lin lab uses types [0.1,0.3,0.5,0.7,0.9], equal probabilities, alpha=1,
Z1=Z2=1, five possible buyers, and public cap ceiling 2. Zero-surplus clients
exit, matching the strict-positive choice in the inspected RQ4 adapter.
The neural labs expose synthetic logits; changing them is not a model
training step. The corrected deterministic path is described separately
from the legacy single-item abstention-removal issue.

## Evidence review

- Contribution: preserve population-aware neural procurement as the proposed
  method; baseline pages explain their distinct design choices.
- Clarity: separate the rule, its synthetic worked example, and empirical
  results. Each method is reachable as a standalone page.
- Experimental strength: display SD and resource quantities; do not infer
  significance or superiority from a favorable single setting.
- Completeness: include all seven named procurement baselines plus both
  controls; leave unreported cells missing and label shared PAC/VCG data.
- Method soundness: identify implementation adaptations and limit claims
  about truthfulness, cumulative privacy, and dynamic equilibrium.

This is a read-only presentation of reported evidence, not a new experiment
audit or a resolution of previously documented protocol/code issues.

## Verification completed for the website revision

- JavaScript syntax checks pass for all scripts.
- PAC, VCG, and CSRA browser calculations match their Python reference
  functions on 45 synthetic market settings, including payment feasibility.
- The Lin menu matches the native Python solver; 27 combinations of budget,
  true cost, and capacity pass menu-budget and optimal feasible-choice checks.
- Neural-head checks cover softmax budget conservation, the unspent share,
  fractional allocation, and deterministic item/abstain boundary behavior.
- Proportional pricing preserves its total payment, Uniform-DP matches the
  displayed mean exactly, and scalar FedAvg matches its analytic value.
- All 36 accuracy mean/SD pairs and all privacy/payment cells match the
  current LaTeX table exactly. RQ1/RQ3 table entries also match the source.
  This is transcription verification, not experimental reproduction.
- Results controls, comparator differences (including unfavorable signs),
  four welfare budgets, missing values, and exported provenance pass.
- All 13 HTML pages pass checks at 360, 390, 768, 1024, and 1440 px:
  dark scheme, English visible text, and no document-wide horizontal overflow.
  Wide data tables/charts may scroll inside their own containers.
- Local links, library cards, baseline selection, preselected result links,
  and mobile navigation work. Browser smoke checks report no script or
  failed-resource errors. The original workflow also passes its misreport
  refusal and budget smoke check after the shared-navigation change.
- Both new main entry points return HTTP 200 on the existing Tailscale-bound
  service when checked from this host. No network, proxy, or experiment
  service configuration was changed.

Current previews use the `preview-*-v3-*.png` filenames. Earlier previews
remain as historical artifacts and are not linked as the current interface.