On-Device Commercial Intent Retrieval Under Size, Latency, and Privacy Constraints: A 3 MiB Retrieval System with Typed Egress Boundaries

1.1 From Tracking to Reasoning

• Stable identifiers and centralized behavioral histories are eroding as assumptions, motivating inference from the user's present on-device activity.The paper focuses on processing current queries, reading, or taps locally rather than sending that information to a server.Source: §1.1 · p. 2

1.2 Three Design Constraints and the Position of This Paper

• The design treats payload, latency, and privacy as falsifiable constraints, with the 3 MiB budget covering data and library code rather than only a binary.The prior build exceeded the 3 MiB gate by 206,555 bytes, prompting a design change; Table 1 identifies constraints alongside measured values.Source: §1.2 · p. 2, Table 1 · p. 3

1.3 Contributions

• The system retrieves across 6,020 commercial leaf categories within a 3 MiB download without an inference runtime, addressing a footprint-versus-quality gap in client-side semantic retrieval.The paper measures the costs of dimensionality and quantization, including the 8-bit-to-4-bit reduction forced by the size ceiling.Source: §1.3 · p. 3
• The personal federated slice has no path into the upload type, and the closed OpenRTB request gives identifiers no field to occupy.The paper also accounts for the campaign-wide privacy budget over the remaining channel.Source: §1.3 · p. 3
• The paper reports system-changing failure analyses and distinguishes demonstrated results from unevidenced claims, including synthetic-dominated taxonomy supervision and unmeasured personalization lift.Its contributions also report that the OS perception channel was closed after harming leaf ranking and that an initial purchase-probability result reflected a degenerate evaluation.Source: §1.3 · p. 3

2.1 Privacy-Preserving Advertising and On-Device Targeting

• Earlier client-side systems target browser audiences or coarse topics, while this work addresses commercial-leaf retrieval and carries an identifier-free typed boundary through the exchange request.The comparison frames the contribution as deployable ranking under a fixed download budget, not client-side inference alone.Source: §2.1 · pp. 3–4, Table 2 · p. 4

2.2 Retrieval Under Resource Constraints

• Lexical scoring misses semantic matches and dense encoders require an unavailable runtime, so the paper uses a static table and inner product for 6,020 leaves.A flat scan takes 0.8 ms on-device, while scalar quantization avoids the codebook state required by product quantization.Source: §2.2 · p. 4

2.3 Static Embedding Distillation

• Static embedding distillation replaces runtime transformer inference with token-row lookup and mean pooling, trading context-dependence for representations derived from a sentence teacher.The paper reports that the static encoder retains 78–87% of the teacher on real-data classification and outperforms same-size public alternatives.Source: §2.3 · p. 4

2.4 Federated Learning

• FedPer separates a shared trunk from a personal head, and this system enforces that separation by making the personal slice unavailable to upload APIs.The adopted structure is supported by a non-IID per-client holdout BCE improvement of +0.382; the aggregation implementation remains a masked prototype without dropout recovery.Source: §2.4 · p. 5

2.5 Differential Privacy

• The privacy guarantee depends on blind aggregation: each device adds Gaussian noise before upload, while the claimed aggregate mechanism relies on independent draws summing to σ√n.The paper says its prototype has not yet met the blindness condition; DP-SGD separately relies on gradient clipping, and accounting choices change the resulting numbers.Source: §2.5 · p. 5

2.6 On-Device Perception

• OS-provided perception options differ by platform: Apple supplies image feature vectors without an additional SDK cost, while Android's dense path requires bundling a model.The categorical labeling and OCR path is available through ML Kit's unbundled distribution at roughly 200 KB.Source: §2.6 · p. 5

3.1 One Contradiction Determined the Structure

• A tiny static embedding floor avoids a resident inference model, while an optional on-device OS model adds inference depth where available.The paper rejects the claim that a 3B model can run within a 5 MB SDK; it reports measuring payload and latency targets separately.Source: §3.1 · p. 5

3.2 Three Tiers

• Tier 0 fixes a dependency-free processing path, while Tier 1 is an optional on-device OS model rather than a server.The wasm and native implementations must produce bit-identical vectors; Tier 1 output enters the Tier 0 vector space through a trained projection.Source: §3.2 · p. 6, Figure 1 · p. 6

3.3 The Federated Layer and FedPer

• Federated learning updates the projection and combiner, not the fixed embedding table, keeping each round's delta small.The base federated vector has 4,096 entries and a 64-entry personal portion; the personal head has no serialization path, and a previously reported 6.9% reduction was withdrawn as inconsistent with the shipped structure.Source: §3.3 · pp. 6–7, Table 4 · p. 6

3.4 The Transmission Whitelist and the Egress Inspector

• The transmission contract allows exactly six top-level keys and excludes the intent vector because it can be recovered by embedding inversion.The egress check detected zero of seven canaries in a 79,968-byte payload, below the 307,200-byte ceiling.Source: §3.4 · p. 7

3.5 The Consent Layer

• Behavioral events carry no text, while a content sketch requires both consent types and is discarded when behavioral-data consent is withdrawn.The event is a 12-byte record; the consent tiers are ordered so higher tiers require lower-tier consent first.Source: §3.5 · p. 7, Table 5 · p. 7

4.1 The Five-Stage Pipeline

• The encoder pipeline selects a 768-dimensional Korean sentence teacher, distills a static table, pretrains it against sentence averages, contrastively trains query–leaf pairs, then truncates and quantizes to 4 bits.The 40,000-document C4-ko pretraining and fp32-before-quantization order are specified; quantizing earlier breaks the fine-tuner.Source: §4.1 · p. 7
• Training used 26,294 synthetic utterances plus capped real Japanese and English ESCI queries; using all 10,748 English queries significantly regressed creative matching and KLUE-STS/YNAT.The real English arm was capped at 2,000 queries.Source: §4.1 · p. 8

4.2 The Promotion Gate

• The promotion gate requires significant improvement on unseen leaves with no regression across nine axes; the upstream pipeline is not verified as bit-for-bit reproducible.The final supervised contrastive stage is deterministic for a fixed artifact and seed, but upstream stages and training inputs drifted.Source: §4.2 · p. 8, Table 6 · p. 8

4.3 Dimensionality and Quantization

• 128 dimensions raises taxonomy top-1 by 27.3 pp, from 0.2410 to 0.5137; PCA truncation leads native-128 by +4.13 pp [2.77, 5.57] but loses at 64 dimensions by −9.95 pp [−11.50, −8.37].The 128-dimensional gain costs 784 KB of gzip; the truncation result reverses with dimensionality, so it cannot be generalized.Source: §4.3 · p. 8, Table 7 · p. 8
• MRL beats PCA on YNAT by +2.08 pp [1.28, 2.90] but trails it on taxonomy by −1.17 pp [−2.33, −0.03], so the paper adopts PCA for the taxonomy product path.The two product paths move in opposite directions, making this a trade-off rather than a free improvement.Source: §4.3 · p. 8
• 4-bit quantization requires per-row f32 scales: a global scale collapses small rows to zero, while per-row scales add 126 KB at 31,497 vocabulary rows.The 126 KB overhead is vocabulary × 4 bytes.Source: §4.3 · p. 9

4.4 Why the 512-Dimensional Candidate Failed

• At 512 dimensions, the measured L1 size was 6,115,038 bytes, exceeding both the 4 MiB line and 5 MiB promise; the earlier 4,002,327-byte projection measured a model that would not ship.The calculation omitted taxonomy and catalog artifacts that grow with dimensionality, while unpacked f32 resident memory was not considered.Source: §4.4 · p. 9, Table 8 · p. 9

5.1 Taxonomy and the Product Surface

• The taxonomy has 28 top-level, 256 mid-level, and 6,020 leaf categories; L2 top-5 is the product surface, derived from leaf rankings by each mid-category’s best leaf rank.Leaf identifiers are their positions in the taxonomy, so additions must go at the end to preserve profile-history alignment.Source: §5.1 · p. 9

5.2 A Hybrid of Dense and Lexical Signals

• 0.9483 leaf top-1 is achieved when a query literally contains the leaf name, while the benchmark excludes such queries and therefore reports a lower bound.The deployment path combines dense cosine similarity with lexical overlap; catalog matching uses pre-downloaded ad vectors and requires no per-user ad request.Source: §5.2 · p. 9, Table 9 · p. 10

5.3 Comparison on Real Queries

• The teacher comparison depends on query regime: the deployed model beats the generic teacher on translated search queries, while results on product names vary with the lexical anchor and teacher training.The comparison used Japanese production queries translated into Korean; its labels were LLM mappings, and its taxonomy had 2,960 leaves.Source: §5.3 · p. 10, Table 10 · p. 10

5.4 Utterance Coverage

• 5,927/6,020 leaves (98.5%) have utterances at deployment time, across a corpus of 36,007.At the earlier 2,960-leaf size, 169 leaves were silently dropped during merging; the recoverable coverage was 2,931/2,960 (99.0%).Source: §5.4 · p. 10

5.5 What the 3 MiB Constraint Cost

• 0.991 (taxonomy) and 0.993 (catalog) are the mean reconstruction cosine values after per-row 4-bit quantization of the category and catalog vectors.The product surface changes by −0.02 pp, while catalog-matching leaf top-1 drops a full point; the embedding table remains at 1.79 MB.Source: §5.5 · p. 11, Table 11 · p. 11

6.1 Turning On the Categorical Channel Hurt

• Turning on the categorical perception channel improves block-level accuracy but scrambles leaf order within blocks, so the ranking-channel flag remains false.Channel dropout avoids regression on devices without Tier 1, but the channel’s leaf-ranking effect is harmful.Source: §6.1 · pp. 11–12, Table 12 · p. 11

6.2 The Resolution of the OS Dense Embedding

• The image embedding exceeds the block-prior baseline by +56.15 pp [+54.28, +57.91] on within-block resolution.The block prior assumes the block is already known and selects its most common leaf, making it a favorable ceiling for a block-level signal.Source: §6.2 · p. 12, Table 13 · p. 12

6.3 The Asymmetry Between the Two Platforms

• Android’s zero-byte perception path is categorical; its dense path requires the app to carry a model, breaking the 3 MiB envelope.On iOS, the OS dense embedding adds zero SDK bytes; vector dimensionality also differs by OS version and requires a revision tag.Source: §6.3 · p. 12, Table 14 · p. 12

6.4 The Ceiling Is the Target, Not the Projection

• Translating the titles into Korean raised the target ceiling by 17.2 pp, and the image arm reached 0.3015, essentially the old ceiling.The English-title arm had only 0.344 category purity; after translation, categories in use widened from 102 to 137, and the image-to-ceiling ratio improved from 0.585 to 0.687 across 456 verticals.Source: §6.4 · pp. 12–13, Table 15 · p. 13

6.5 Wired In but Kept Disabled

• The projection path is wired into the SDK but disabled by default; on the test device, the OS language-model path was unavailable at runtime, blocking Layer 3.The projection reaches ranking only when both content consent and the user-embedding toggle are on; the default state keeps all three required switches off.Source: §6.5 · p. 13

7.1 Round Protocol

• At 30% dropout, round aggregation reaches 50/50 and holdout BCE falls from 0.6936 to 0.2125.The dropout simulation omits masking, while mask cancellation is demonstrated separately under full participation; the two properties are not yet demonstrated together.Source: §7.1 · p. 14

7.2 What It Means to Close the Loop

• A regression test checks application of the aggregate, movement of the personal head, primary-ranker intervention, and survival across refresh.The earlier implementation accepted uploads but sent zero-vector-based parameters, discarded the personal head, and ignored the returned aggregate.Source: §7.2 · p. 14

7.3 On-Device Labels

• The listwise combiner learns which displayed candidate was selected from taps, while the personal head uses binary cross-entropy with positives and negatives.Tap labels retain implicit-feedback bias, predictions are not reused as labels, and taps are rare.Source: §7.3 · p. 14

7.4 Differential Privacy Accounting

• A 50-round campaign composes to ε = 17.67 by hand-computed zCDP, outside the cited Gboard production range of 0.99–13.7.The proposed rollout keeps federated learning off in Stage 1 and enables it only after the device-population precondition is met.Source: §7.4 · p. 15

7.5 Attribution Channel

• The composed campaign ε is 1.7295, including 1.5422 from the federated-delta channel and a +0.187 contribution from attribution.The attribution channel is bounded; at σ = 0.3, billing error is about 2.7% cumulatively, while 19% per-round error rules out per-round settlement.Source: §7.5 · p. 15

7.6 The Personalization Axis Is the Weakest

• Sub-category top-1 falls below the shipping build in every corrected-data arm, so the personalization artifact fails the criterion requiring both surfaces to improve.The runtime default remains off; the reported blockers include only 3 real labels.Source: §7.6 · p. 16
• A measured sign reversal gives mid-level (L2) top-5 +0.80 pp and sub-category top-1 −0.91 pp within the same trained-combiner family.The shipped α = 1.0 yields +6.02 / +3.75, whereas α = 2.0 yields +7.02 pp on the surface and −3.35 pp for sub-category.Source: §7.6 · p. 16

8 The Ad Call — OpenRTB 2.6 and Query-less Placements

• The system embeds its final ad-call leg in OpenRTB 2.6 and uses types, rather than documentation alone, to enforce the privacy boundary.This connects the on-device ranking to an ad call.Source: §8 · p. 16

8.1 Closed Request Types

• The request’s closed user and extension types omit identifier fields and reject unexpected keys during serialization and deserialization.Measured requests are 1,721 bytes per call with five intents, and all seven raw-text canaries are undetected.Source: §8.1 · p. 16

8.2 Leaf Granularity and Sensitive Categories

• The standard segment carries the mid-level category, while the optional leaf identifier is confined to the extension and is omitted for sensitive categories.Leaf matching is accepted only within the same mid-level category; otherwise matching falls back to the mid-level pool.Source: §8.2 · pp. 16–17

8.3 Korean Placement Specifications and the Creative Pipeline

• The creative pipeline registers 11 Naver GFA and Kakao placements with exact pixel specifications and size caps.Synthetic SVG fallbacks are explicitly marked, and unregistered raster requests return 404.Source: §8.3 · p. 17

8.4 Query-less Placements — Ads Served Without an Utterance

• Without an utterance, ranking uses a consented content sketch, profile prior, and behavior signals; history-only fallback is used when confidence falls below the floor.No ranking is fabricated when consent or history is absent, and the article text itself is not included in the request.Source: §8.4 · p. 17

8.5 Validation and Limitations

• The web end-to-end test validates the round trip against a mock exchange, but no real exchange is connected and mock matching quality says nothing about a real exchange.The authors also note that nurl/burl reveal wins to the exchange and that the SDK cannot erase IP.Source: §8.5 · p. 17

9.1 A Closed-Form Hazard, Not a Learned Model

• The paper models repurchase as a closed-form hazard using a per-category cycle prior and count term, avoiding an inference runtime.It calls this formulation mPG and uses the closed-form property as a device-budget requirement.Source: §9.1 · pp. 17–18

9.2 A Degenerate Evaluation Task Looked Like a Weak Model

• The same formula scores 0.7425 at w = 3d and 0.6915 at w = 7d on the revised harness, but these values are not claimed as improvements over the earlier 0.53–0.59 task.The harnesses ask different questions; the authors conclude that half of the apparent weakness came from evaluation.Source: §9.2 · p. 18, Table 21 · p. 18

9.3 Where the Signal Comes From, and Its Ceiling

• Free weights for base rate, frequency, and recency account for almost all the gain; adding the Weibull logit contributes only +0.0002 / +0.0008.The fitted Weibull coefficient is negative, and the best-fitting shape parameter is m < 1 (0.85–0.9), unlike the recently-purchased suppression rule.Source: §9.3 · p. 18, Table 22 · p. 18
• The deployed closed form scores 0.7425/0.6915 against LightGBM’s 0.7316/0.6996 on the same features, a tie that places the arithmetic hazard at the feature-supported ceiling.The authors identify more information or free weights, rather than a larger model, as ways to lift that ceiling.Source: §9.3 · p. 18

9.4 Does It Survive Time, Our Own Taxonomy, and the Deployed Code?

• Across time, ③ retains 88–89% of its gain, while ⑤’s marginal contribution collapses from +0.0108 / +0.0087 to +0.0009 / +0.0032.On the authors’ taxonomy, the lift is ∆+0.0040 at w = 3d and +0.0016 at w = 7d; the deployed Rust path reproduces the Python measurements.Source: §9.4 · p. 19

9.5 Combining It With Intent

• With competing categories, λp = 2.0 is significantly worse than the deployed default (−1.55 pp), while λp = 1.0 is significantly better (+1.73 pp).The fine-grid optimum is 1.0 with a 0.9–1.1 plateau; the within-category-only setting would have selected the rejected value.Source: §9.5 · p. 19

9.6 Learning the Combiner Federated — and What Is Missing

• In a 50-round simulated campaign, the rank head converges from AUC 0.5 to 0.7734, with the four added coordinates leaving the whole-campaign privacy budget unchanged at ε ≈1.54.The coefficients still require a pilot cohort of real people; without them, the ranker falls back to the hazard path and forgoes the measured gain.Source: §9.6 · p. 19

10.1 Real-Data Korean Benchmarks

• On KLUE-YNAT and NSMC, the public model trails by 24.2 and 5.2 pp, tokenlearn gains 11.7 and 0.8 pp, and shipping retains 82.4% and 88.8% of teacher performance at 196.6× smaller size.The tokenlearn NSMC confidence intervals overlap; the shipping retention figures are distinct from the stage-2 training-ablation figures.Source: §10.1 · p. 20, Table 23 · p. 20

10.2 Real Korean Commerce Text Against Our Own Taxonomy

• 0.7504 mid-category top-5 on 1,174 real product names exceeds the 0.1843 permutation baseline, rising to 0.8350 when a leaf name appears as a substring.The permutation baseline severs the query-label link while preserving the mapping and label distribution; scrambling query characters lowers top-5 to 0.4736, confirming the path reads the query.Source: §10.2 · p. 21, Table 24 · p. 21
• With source-category fixed effects, the model-code estimate attenuates from −35.47 to −4.01 pp, while the lexical-anchor effect survives; failures cluster in titles with no category noun.Two source categories fall to their own chance level or below (0.035 versus 0.162; 0.036 versus 0.040), including systematic misrouting toward digital-device categories.Source: §10.2 · pp. 21–22
• Out-of-distribution reviews score 0.4552 mid-category top-5 against 0.1241 baseline, while full customer utterances score 0.5897 top-category against 0.3813.These are stress tests rather than bounds on short-query accuracy; exact leaf-name matches cover 32.5% of annotated product mentions in customer utterances.Source: §10.2 · pp. 23, 25
• Narrowing the hand-built crosswalk lowers accuracy from 0.7504 to 0.5571 while lift over its recomputed null falls from +56.6 to +42.1 points; category-macro accuracy is 0.6473.The narrowed mappings are descendants of a crosswalk written while reading evaluation names, so this sensitivity check does not eliminate evaluation-definition leakage.Source: §10.2 · pp. 23, 25, Table 28 · p. 23
• The deployed path beats BM25 by 5.11 points overall, and its margin over BM25 is 12.74 pp without an anchor versus 2.95 pp with one.Without an anchor, BM25 scores 0.3243 versus dense-only at 0.4247; the deployed path’s paired margins are significant in both regimes.Source: §10.2 · p. 24, Table 29 · p. 24
• 0.8586 matched-teacher accuracy beats dense-only by 10.6 pp with an anchor and 20.9 pp without one.Task adaptation adds 10.5 pp with an anchor but 0.8 pp without; the generic teacher’s corresponding gains are 0.1 and 20.1 pp, respectively.Source: §10.2 · pp. 24–25, Table 29 · p. 24
• Declining the least confident fifth of product names using the ranker’s score margin lifts accuracy on the rest to 0.8296; declining 40% reaches 0.8977.The margin is the deployed ranker’s own first-to-fifth mid-category score gap, so the failure regime can be detected on-device without a teacher or network call.Source: §10.2 · p. 25, Table 30 · p. 25
• 29 of 200 (14.5%) order-corpus items and 11 of 200 (5.5%) call-centre lexicon terms were suffixes of at least two leaf names, indicating skipped hypernyms.The proposed next lever is vocabulary expansion, not more generation, and is described as concrete and cheap.Source: §10.2 · p. 26

• The current 128-dimensional, 4-bit build improved the product surface but paid for that gain on general language.The promotion gate compares it with the previous build on real-data axes.Source: §10.3 · p. 26, Table 31 · p. 26

• The product commitment is L1 downloaded data plus L2 library code; L3 resident runtime bytes has a separate cap.L1 is byte-identical across targets, while linked library code differs.Source: §10.4 · p. 26, Table 34 · p. 26
• 2,942,652 bytes is the shipping payload, 93.5% of the ceiling, using the largest measured target link.All three targets now have measured links; the Android link is largest at 65,827 bytes.Source: §10.4 · p. 26, Table 34 · p. 26
• L1 accounts for 98.3% of payload and leaves 268,903 bytes of headroom for library code, roughly five times either measured link.The L1 data is byte-identical across targets.Source: Table 34 · p. 26, §10.4 · p. 27
• The earlier budget omitted taxonomy and affinity artifacts and had no line for library code, understating the app-size increase publishers experience.The missing artifacts left 206 KB uncounted.Source: §10.4 · p. 27

10.5 Latency

• The p95 budget is 20 ms on a 512-character input; Tier 1 is excluded because an OS model call takes hundreds of milliseconds.The OS call runs asynchronously to improve the vector for a later request.Source: §10.5 · p. 27
• The user-embedding switch adds similar latency for sparse and dense histories (+0.329 vs. +0.342 ms), with cost in two taxonomy score computations.The paired difference transfers across taxonomy changes, but 1.421 ms is not the shipping-build latency.Source: §10.5 · p. 27, Table 35 · p. 27

10.6 On a Handset

• The handset test runs the gate’s Rust function on installed iOS and Android devices, using nine fresh processes and reporting the median and range.The artifacts are checked against deployed artifacts on device.Source: §10.6 · p. 27, Table 36 · p. 27
• All three handsets meet the 20 ms budget at 22.2%, 18.3% and 25.4% of its allowance, and each is slower than all three server CPUs.The budget Android tablet measures 5.080 ms; the server CPU p95 values span 2.00–3.37 ms.Source: §10.6 · p. 27, Table 36 · p. 27, Table 37 · p. 28

10.7 Why We Do Not Quote a Reference-Mac Number

• The single-machine reference-Mac figure was withdrawn because its source run lacked an artifact and a later measurement was contaminated by another process.The contaminated p95 was 3.02 ms, 25% above the previously quoted 2.404 ms; neither is claimed as quiet-machine truth.Source: §10.7 · p. 28

10.8 The Same Binary on Server CPUs

• Across three server CPUs, p95 ranges from 2.0 to 3.4 ms across a 16× thread-count spread, and larger machines are not faster in these observations.The authors do not instrument the cause of the ordering.Source: §10.8 · p. 28, Table 37 · p. 28
• Server observations do not establish lower-end mobile performance or address thermal throttling, cold caches, or a busy host application.They do show the budget does not rest on one unusually fast development machine.Source: §10.8 · p. 28

10.9 Gate Summary

• +6.51 pp [+5.20, +7.89] is the leaf figure over 1,750 people; the top-level +18.23 pp [+16.34, +20.17] rollup would nearly triple the displayed gain.In the conservative host-area-only condition, gains are +14.23 and +3.71 pp.Source: Table 38 · p. 29, §10.9 · p. 29
• The attribution-axis consistency check is red: its ℓ2 sensitivity of 1.7321 is within target, but the budget declaration and shipped 60-class intent head disagree, blocking deployment.The gate reports a consistency failure; it does not correct previously reported numbers.Source: §10.9 · p. 29

10.10 A Claim Retracted on the Behavioral Axis

• The behavioral headline is retracted because four measurements of the same axis diverged 25-fold, and its supporting run had the fewest samples and smallest teacher margin.Matching teacher dimensionality yielded only +1.29 pp [+0.26, +2.58]; reducing samples from 5,250 to 1,187 alone changed the teacher gain from +9.31 to +3.02 pp, though the latter comparison lacks a ledger.Source: §10.10 · pp. 29–30
• Behavior was complementary rather than redundant: on the synthetic LLM run, behavior alone was more often right when text was wrong, with teacher lift −0.118.The small fusion gain was attributed to a classifier that had not learned when to trust behavior.Source: §10.10 · p. 30
• Three runs retained a behavioral gain for the weak on-device text model: +15.1 pp on real logs, +17.8 pp synthetic, and +18.23 pp on the shipping path; the shipping-path gain was +27.92 pp for ambiguous versus +6.34 pp for unambiguous queries.The paper frames the retained claim as access rather than accuracy: the same events would improve a server, but the signal stays on-device.Source: §10.10 · p. 30

10.11 Promotion Signal

• At margin threshold 0.1648, promoting the bottom 25.0% recovered 49.2% of errors, a 1.97× lift over random promotion; this promotion result is based on synthetic data.The authors identify it as the first result to remeasure when real usage utterances arrive.Source: §10.11 · p. 30

Data limitations

• Taxonomy supervision remains dominated by synthetic Korean utterances with no deployment-derived Korean labels, and the evaluation corpus is real commerce text rather than device traffic.The Korean labels were mapped by the authors while viewing evaluation product names; no Korean marketplace query log is available.Source: §11 · pp. 30–31

Real-data evaluation

• Real Korean commerce evaluation supports mid-category top-5 of 0.8350 on real product names carrying a leaf-name anchor, but results are crosswalk-conditional and vary widely across source categories.The hand-built AI-Hub crosswalk remains a limitation; the 0.7504 average spans category accuracies from 0.035 to 1.000, and query-log proportions remain unknown.Source: §11 · p. 31

Distribution probe

• The real-text distribution probe uses unlabeled reviews rather than queries, so its prediction concentration cannot distinguish real category imbalance from model bias.Reviews are four times longer than queries and may not name a product; the probe establishes that real-text distribution differs from the uniform-by-construction evaluation set, not deployment accuracy.Source: §11 · p. 31, Table 40 · p. 31

Evaluation scope

• The sub-category top-1 of 22.0% excludes leaf-name queries, while 0.9259 applies only to queries literally naming a leaf; their prevalence in actual use is unknown.The reported 0.9259 is not product accuracy because the query distributions differ and no real usage log establishes their proportions.Source: §11 · p. 31

Platform measurements

• Low-end latency coverage is one budget Android tablet tested as a bare process, without sustained-load thermal measurements, host-app load, or weaker-phone coverage.Android verification checks used by the iOS app were not exercised, and resident memory on low-end Android was not measured.Source: §11 · p. 32

Personalization

• Personalization lacks a real update path: the user-embedding weights use synthetic personas, only three real labels exist, and 800 training plus 1,750 evaluation labels are needed.The labels reside on-device, while central collection is forbidden; the upper-bound combiner result also flips sign at another granularity.Source: §11 · p. 32

Artifacts and perception

• Several deployment components remain unvalidated: the projection artifacts are untrained, the catalog and taxonomy are synthetic, and perception measurements do not establish performance on real-device iOS or the 6,020-leaf product taxonomy.The dense-embedding agreement metrics compare against the title arm rather than ground truth, except purity.Source: §11 · p. 32

Deployment and market scope

• The paper cannot establish market-relative accuracy or bid-price impact without a pilot, and its federation, legal, and OpenRTB pathways retain unresolved implementation or review conditions.Secure Aggregation dropout recovery is unimplemented, the 10,000-device cohort may require a multi-publisher agreement, consent copy has not had legal review, and the OpenRTB path has had zero real round trips.Source: §11 · pp. 32–33

Availability

• The published repository contains ledgers, gate and evaluation harness sources, and artifact fingerprints, but no SDK crates, model weights, taxonomy, ad catalog, or datasets.The repository is a measurement record, not the system or data.Source: Data and Code Availability · p. 33

Data availability

• The authors redistribute none of the datasets underpinning proprietary-taxonomy experiments because source terms restrict derived materials and the authors lack rights to their own constructed materials.The public-domain Naver review corpus is an exception and is cited at its original source rather than copied.Source: Data and Code Availability · p. 33

Reproducibility

• Readers can audit figures against ledgers and artifact fingerprints, but cannot regenerate results from the bundle alone without independently obtaining or re-deriving underlying sources.Gates report UNMEASURED when required artifacts are absent rather than fabricating a verdict.Source: Data and Code Availability · p. 33

12 Conclusion

• The system reports a 2.14 MB static encoder retaining 78 ˜ 87% of a 196.6× larger teacher on real-data classification, with mid-level top-5 of 77.6% on unseen-category colloquial queries over 6,020 leaves.The reported deployment payload is 2,942,652 bytes, and Tier-0 p95 is 4.431 and 3.670 ms on two iPhones and 5.080 ms on a budget Android tablet; all three devices are slower than the server CPUs.Source: §12 · p. 33
• The categorical perception channel was a net loss for leaf ranking, so the authors closed it and pivoted to a rectangular projection.The dense-embedding probe measured the gap at 56.15 pp.Source: §12 · p. 33
• A projection from images into the system’s space cannot exceed its target vector, and translating English titles into Korean raised that ceiling by 17.2 pp.The projection rose with the target, supporting the conclusion that data constrains both the training set and target vector.Source: §12 · pp. 33–34
• The paper frames its claim as access rather than accuracy: behavior can improve a server by the same amount, while the egress gate and type system measure whether the signal stays on-device.The authors do not claim the system is smarter than the server.Source: §12 · p. 34
• Real usage utterances and labels remain the blocker: personalization has three real labels, taxonomy supervision has no deployment-derived Korean labels, and evaluation uses public corpus text rather than SDK traffic.The authors identify a pilot, not a new model, as the next step.Source: §12 · p. 34

Reproducibility

• The reproducibility audit checks 386 figures against ledgers and makes an unledgered figure fail rather than silently exempting it.The final contrastive stage is deterministic for a fixed artifact and seed, but upstream token-learning and corpus preparation are not.Source: Reproducibility · p. 34
• Real-data experiments can be run locally by readers with their own AI-Hub access; the authors neither receive nor redistribute the text.The workflow downloads validation archives and runs the evaluation scripts locally.Source: Reproducibility · p. 34