skelfresearch/on-device-auction-audit

On-Device Auction Audit

Data for When Privacy Moves ML-Mediated Decisions On Device: Information and Incentive Misalignment in Auctions, Dipankar Sarkar, Skelf Research.

Paper: https://arxiv.org/abs/2609.33312

Code and full artifact: github.com/sarkar-dipankar/on-device-auction-audit (this package is built from commit 1d1a05b; the data files are unchanged since 5a6ad86).

The question

Privacy pushes ad decisions onto the device. The auction runs locally, next to the context it scores.

But the budget lives on a server.

So each device bids against a balance it last saw at its most recent sync. Between syncs, campaigns can spend money they no longer have. And the local auction can reward a bidder for misreporting.

The paper asks how much that costs. Two misalignments, measured side by side:

• Information: stale budget state. Measured as overspend_pct and underspend_pct against a frozen budget.
• Incentive: auctions the simulator flags as manipulable or misallocated. Measured as manipulable_auction_rate, manipulable_rival_auction_rate, and misallocation_rate.

What the simulation does

Every primary run uses the same frame, read from the manifests:

• 50 devices, 36 campaigns, 100 ticks, 3,000 expected arrivals per tick (about 300,000 auctions per run).
• Each arrival samples one ad context uniformly from the 50,808 labelled windows in the contexts config.
• Budget sync intervals of 0, 1, 2, 5, 10, 25, and 50 ticks. Interval 0 is the zero-lag reference.
• Pacing strategies asap, even, even_bernoulli, and pid. Payment rules second_score and critical_bid.
• Budgets are calibrated on 30 pilot seeds (30260527 to 30260556) to a target oversubscription of 20, then frozen. Results come from 30 disjoint evaluation seeds (20260527 to 20260556).

Two demand constructions share that frame. They are the two splits of most configs:

• synthetic: a fixed value draw (bid_distribution kind fixed, 100 score units). The release directory is track_b_smoke, a legacy name.
• ipinyou: a log-normal draw with mu 3.8713 and sigma 0.85, set so its mean equals the 68.89 mean market price reported for iPinYou campaign 1458 (Zhang et al. 2014). A calibrated synthetic value proxy, not raw bidding logs.

One cut you can recompute from primary_replicates in a few lines. Mean overspend_pct over 30 seeds, even pacing, second_score payment:

Sync interval (ticks)	0	1	2	5	10	25	50
synthetic	0.53	17.77	50.43	189.94	479.28	1346.13	1669.31
ipinyou	1.00	31.87	72.14	109.44	124.08	402.79	905.03

Lag alone moves total spend from about 1% over the frozen budget to 10.1 times the budget (ipinyou) or 17.7 times (synthetic) at 50 ticks. The paper's bounds, paired contrasts, and confidence intervals are in raw/results/track_b_paper_numbers.json.

Configs

Twelve viewer configs, all JSONL, built by build_hf.py (in this repo, stdlib only, deterministic). Rows keep their source order. The original files are unchanged under raw/results/.

Config	Splits (rows)	Source	What it is
primary_replicates (default)	synthetic 900, ipinyou 900	track_b_{smoke,ipinyou}/replicates.jsonl	The primary result. One row per strategy, payment rule, sync interval, and seed: spend, overspend, underspend, fill, clearing price, delivered value, manipulability, misallocation. 30 cells x 30 seeds per split.
pressure_balance	train 1,050	track_b_pressure_balance/replicates.jsonl	840 pressure_sensitivity rows (budget pressure 2, 5, 10, 20) and 210 visible_balance_guard rows (devices reject bids above their visible balance, pressure 20). even pacing, 7 sync intervals, 30 seeds.
bursty_heterogeneous	train 210	track_b_bursty_heterogeneous/replicates.jsonl	Robustness sweep. Two-state Markov-modulated demand over 10 heavy and 40 light devices, stationary mean load unchanged. 7 sync intervals x 30 seeds.
replicate_summary	synthetic 30, ipinyou 30	replicate_summary.jsonl	One row per primary cell: mean, sd, min, max, and Student-t 95% CI of each metric over its 30 seeds.
per_campaign	synthetic 216, ipinyou 216	per_campaign.jsonl	Per-campaign budget, spend, wins, and delivered value for even pacing at sync intervals 1 to 50.
calibration	synthetic 744, ipinyou 744	calibration.jsonl	The budget calibration trail: 720 pilot runs and 24 frozen per-vertical budgets, for the second_score and critical_bid (reserve 100) designs.
factorial_ablation	synthetic 240, ipinyou 240	factorial_ablation.jsonl	Payment rule x budget policy (calibrated or floored at 500) x reserve (0 or 100), zero lag, 30 seeds per cell.
factorial_calibration	synthetic 384, ipinyou 384	factorial_calibration.jsonl	Calibration rows for the factorial budget vectors.
sensitivity	synthetic 180, ipinyou 180	sensitivity.jsonl	Device count (10, 50, 200) and aggregate arrival load (1,500, 3,000, 6,000) varied one at a time, 30 seeds per level.
label_perturbation	synthetic 210	track_b_smoke/label_perturbation.jsonl	Combined label stress test on the context labels, 7 sync intervals x 30 seeds.
label_perturbation_calibration	synthetic 372	track_b_smoke/label_perturbation_calibration.jsonl	Budget calibration for the perturbed labels (360 pilot runs, 12 frozen summaries).
contexts	train 50,808	data/opportunities_teacher_50k.jsonl	The frozen context labels every run samples from. Two shards.

The contexts config

One row per WildChat-1M conversation window (window_id, all unique). No conversation text.

• topic: 24 values. technology is the largest (18,394 rows).
• intent_stage: none, informational, comparison, purchase_ready.
• category: 19 ad categories, null on 41,247 rows.
• safety_class: 8 values, ok on 45,886 rows.
• ad_eligible: true on 6,370 rows.
• subtopic is null on every row. locale (en-US) and device_type (desktop) are constant. They are kept so the schema matches the simulator input.

The file is 9.9 MB, so it is split on line boundaries into data/contexts/part-0000.jsonl and part-0001.jsonl (25,404 rows each). The shards are byte-exact slices. Concatenated in order they reproduce the source file, and its sha256 matches the one every manifest records:

cat data/contexts/part-0000.jsonl data/contexts/part-0001.jsonl | sha256sum
# 16ad2a17e55bb2185bfe4633521c70c9c03962131bc430358cc13df018ecbbf8

The label stress test in label_perturbation changes 1,558 of the 50,808 records: 525 false positives, 508 false negatives, and 525 wrong-vertical relabels (raw/results/track_b_smoke/label_perturbation_manifest.json). The manifest calls it a sensitivity analysis, not an estimated label-error model.

Column notes

• Accounting uses dimensionless integer score units. The *_cents identifiers are legacy Rust field names. They carry no currency semantics. The newer runs (pressure_balance, bursty_heterogeneous) already use *_score_units.
• Nested values are serialised as JSON strings so each column has one type. That covers every metric in replicate_summary (a {mean, sd, min, max, ci95_halfwidth, ci95_method} object) and campaign_exhaustion_fractions in primary_replicates (one value per campaign). Parse them with json.loads.
• sensitivity holds two designs in one file. aggregate_arrivals_held_fixed is set only on n_devices rows and budget_vector_held_fixed only on aggregate_arrival_load rows. Each is null elsewhere.
• pilot_index in label_perturbation_calibration is null on the 12 frozen_summary rows.

raw/

raw/results/ is an exact copy of the release results/ directory. raw/scripts/ holds the paper's offline generators and tests. raw/data/ipinyou_1458.json holds the published iPinYou aggregates. raw/SHA256SUMS and raw/LICENSE are the release's checksum and license files. No config points at raw/.

It also holds files with no config:

• track_b_paper_numbers.json and track_b_revision_numbers.json, the statistical summaries the paper cites.
• Every manifest.json, with seeds, budgets, commands, input hashes, and source hashes.
• budgets, oracle, pacing_report, and variants JSONL files (5 to 36 rows each).
• track_b_controllers/, a legacy floored-budget controller export. Its manifest has status superseded by track_b_smoke and track_b_ipinyou and says it is not an input to the paper. It is kept for completeness and given no config.

Withheld simulator source

The simulator is not public. Each manifest names a source_snapshot.tar.gz holding the code that produced its rows. Those archives are not in the GitHub release and not in this package. Their SHA-256 values stay in the manifests as a record of the exact code state:

• primary and iPinYou runs: dabe08d17b18aa6ba8816b99b4ed164dc866c3de12b9eee3439e55d5da72a2a4
• pressure/visible-balance and bursty/heterogeneous runs: 207a49793ebeab890dd6821deebde1138d569d8c0f55c32c7961f95597a1cebc

So the simulations cannot be replayed from this repo. The statistics can.

Reproduce the statistics

From a clone of the GitHub repo:

python3 -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
python scripts/generate_track_b_paper_numbers.py
python scripts/generate_track_b_revision_numbers.py
python scripts/test_track_b_analysis.py

We re-ran this on a copy of commit 5a6ad86 (Python 3.12.3, SciPy 1.18.0). Both generators passed their input checks. All 5 tests passed.

The regenerated summary JSONs differ in bytes from SHA256SUMS. Every non-float value is identical. The float differences are rounding noise: at most 1.03e-13 relative in track_b_paper_numbers.json and 2.2e-14 in track_b_revision_numbers.json. Compare them numerically, not by hash.

To rebuild this package from a checkout:

python3 build_hf.py /path/to/on-device-auction-audit .

Provenance

• Contexts derive from WildChat-1M (Zhao et al., ICLR 2024, arXiv:2405.01470). Each window was labelled by a single gpt-oss:120b teacher, described in the paper. The labels are teacher labels, not human ground truth. No conversation text is redistributed. See the WildChat paper and dataset card for collection and consent details.
• iPinYou parameters are published aggregates for campaign 1458 from Zhang et al. 2014, arXiv:1407.7073, Table 3. No raw bidding logs are included. The market price is a value proxy, not advertiser private value.
• Replicates are simulator outputs produced by the withheld source above.

License

The data is released under CC BY-NC 4.0.

The context labels derive from WildChat-1M and inherit its CC BY-NC 4.0 terms. Every simulation result here is computed from them, so the whole dataset carries those terms. Attribute both WildChat and this derivative.

The GitHub repo's code is MIT (raw/LICENSE). That covers raw/scripts/ and build_hf.py. It does not cover the WildChat-derived data, and neither does the paper's arXiv license.

Citation

@misc{sarkar2026ondevice,
  title = {When Privacy Moves ML-Mediated Decisions On Device: Information and Incentive Misalignment in Auctions},
  author = {Dipankar Sarkar},
  year = {2026},
  eprint = {2609.33312},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.33312}
}

Please also cite WildChat (Zhao et al., ICLR 2024) and, if you use the ipinyou split, Zhang et al. 2014.

Links

• Paper: https://arxiv.org/abs/2609.33312
• Code and full artifact: https://github.com/sarkar-dipankar/on-device-auction-audit
• WildChat-1M: https://huggingface.co/datasets/allenai/WildChat-1M