Files
paper-how-to-use-this-techn…/research/06-applications-matrix.md
T

62 KiB
Raw Blame History

title, status, date, scope
title status date scope
Applications matrix: six challenges and one decision-grade urban sensing stack research-note 2026-11-14 Future City Hackathon 2026 challenges, datasets, and the PulseCast solution in OKF_Vault

Applications matrix: six challenges and one decision-grade urban sensing stack

Purpose. This note tests the paper's central claim: that a privacy-first sensing + spatial field + forecasting core generalizes across urban problems. It maps all six hackathon challenges and the PulseCast solution onto one stack, identifies where the abstraction holds, and states where it breaks. It is deliberately conservative: a method that is transferable in principle is not presented as empirically validated.

Evidence labels. [D] documented in the knowledge bundle; [M] measured in the PulseCast baseline; [P] proposed or planned, not evaluated; [G] gap or assumption; [X] external data not supplied in the bundle. Every number below is traceable to a concept file or mission file. No number is estimated here. Where the bundle is silent, the text says so.

One-paragraph answer. The stack generalizes as a contract, not as a single model. The contract is: declare the observation operator and its privacy class; infer a latent state at the resolution the decision needs; forecast only when the decision is forward-looking; fuse context that is actually available at decision time; separate the decision rule from the predictor; and evaluate against a baseline, and where possible against decision regret. The contract is demonstrated for footfall-to-staffing and is plausible, with caveats, for advertising and waste routing. It becomes a sibling spatial-design problem for sensor placement and LoRaWAN planning, where the decision changes the observation process. It does not hold for cable-car sentiment or the learning guide, which have no physical field, no forecast target, and no resource-allocation decision.

Definition. For this note, a stack is decision-grade when (i) the observation operator and its error/bias are stated; (ii) spatial and temporal resolution match the decision unit; (iii) forecasts are validated on a time-based holdout against a naive baseline; (iv) the decision rule is explicit, constrained, and separable from the predictor; (v) evaluation includes a decision-relevant metric, not only accuracy; and (vi) privacy and provenance are auditable. The matrix in Section 3 assesses each challenge against these conditions.


1. Challenge summaries

The six challenges are not six versions of the same problem. They differ in what is observed, whether a spatial field exists, whether the decision is forward-looking, and whether the decision changes the observation process. The summaries keep those differences explicit.

C1 — Pulse of the City (Schwarz Digits)

Decision. The challenge names three decisions but does not choose one: where and when to place physical advertising, where to locate a store, and how to staff a store. The decision-maker is a retailer or the city. The challenge's stretch goal is explicitly "estimations and predictions, not only historical visualisation." [D]

Data. PAX counters passively sniff Bluetooth MAC advertisements from nearby smartphones and smartwatches, then transmit aggregated device counts over a LoRaWAN network. The counts are estimates, not a census: not everyone carries a discoverable device. The challenge brief calls the data privacy-preserving because the counters aggregate devices rather than identify people. The raw CSV is in data/raw/ in the workspace, not in the knowledge bundle; the download and schema are documented in mission/data.md. [D]

Spatial resolution. The challenge describes sensors in Heilbronn's inner city. The delivered data contains 12 device IDs; after merging the two density-01 IDs and excluding three sensors, 8 usable sensor locations remain in a ~580 m x 280 m cluster around Hagenbuchersee, Neckarufer, Stadtgalerie, Marahaus and Götzenturm. The three excluded sensors are maker-fair (73 readings, located in Hannover), niklas-01 (8 readings, outside the inner city), and hn-bollwerksturm-01 (5 readings). This is an inner-city micro-area, not a city-wide network. [D]

Temporal resolution. The 8-sensor network covers 2026-06-12 to 2026-11-07, about five months. Cadence is ~15 minutes (median 15.3 min; some sensors 30 min). Gaps reach up to 6 days per sensor in the 8-sensor network; density-01, which has 13 months of history, has a 35-day gap. It is a single location and cannot establish network-wide annual seasonality. The full file covers 351 days overall. [D]

Key uncertainty. Device penetration is unknown; the values are an index of detected devices, not people. The spatial coverage is not city-wide. Gaps, missing events, observed rather than forecast weather, and a short multi-sensor history all limit inference. There is no POS, sales, queue, labour-cost, or event data. [D]

Heatmap / spatial field. The challenge slide illustrates a heat map over a map of Heilbronn plus a multi-line time-series chart. With eight points in a 580 m cluster, only a local pulse map with measured points and a clearly labelled interpolation is honest. A city-wide heatmap is not supported by the data. [D]

Prediction target. Footfall or device density per location and time. The challenge does not fix a horizon; "estimations and predictions" are the stretch goal. [D]

External covariates. The brief suggests weather, holidays, events, or other real-world signals that explain patterns. It does not supply them. PulseCast adds Open-Meteo weather and Baden-Württemberg public holidays. [D]

What the challenge leaves open. The unit of densityData, the sensors' catchment areas, the device-penetration correction, the cause of 4–6 day gaps, whether an event log exists, and whether a live API exists. [D]

S1 — PulseCast (our solution to C1)

Decision. A weekly shift plan for one inner-city store: how many staff per hour next week, with the manager able to accept, edit, or reject the plan. The forecast and scenario engine is the reusable core; the staffing app is its first consumer. [D]

Data. The same PAX stream: 192,464 CSV rows, 96,232 density readings, 12 device IDs, 8 usable sensors, 3 excluded test sensors, ~15-minute cadence, about five months of multi-sensor history. Weather comes from Open-Meteo; holidays from the Baden-Württemberg calendar. [D]

Spatial resolution. Eight fixed points in the ~580 m x 280 m cluster. The product shows a pulse map of measured points and a labelled interpolation, not a city-wide surface. [D]

Temporal resolution. Four horizons are exposed: 15 min, 1 h, 48 h, and 7 d. The 7-day plan is the hero. The documented evaluation uses a 28-day time-based holdout; 15 min and 1 h are nowcast/replay because the CSV has no live feed. [D]

Key uncertainty. Device penetration, 4–6 day gaps, observed weather in training, no event log, no POS/labour data, a four-week test period, one store, and no annual seasonality. [D]

Heatmap / spatial field. A "pulse map" of eight measured points with a clearly marked interpolation. It is evidence for a local staffing decision, not a city-wide measurement. [D]

Prediction target. Hourly footfall index per sensor, then staff count via a transparent heuristic: required staff = ceil(footfall / visitors-per-staff-hour x safety factor), bounded by minimum staff, shift length, opening hours, and cost. [D]

External covariates. Open-Meteo temperature and rain are the useful model features; cloud and wind did not improve the short test. Public holidays are a feature and a dashboard badge, but added little in the 28-day test. The what-if scenarios are counterfactual inputs, not observations. [D]

Claim discipline. The model predicts estimated devices, not people. The staffing plan is decision support with explicit assumptions, not an optimizer with proven ROI. The deployment is a hackathon deployment, not a multi-tenant production system. [D]

C2 — WasteRouting.AI (PreZero)

Decision. Which glass containers to empty, in what order, with which fleet, under capacity, shift, and depot constraints. The starting point is a fixed area-by-area schedule that ignores actual fill levels. The challenge asks for a configurable route engine, not only a dashboard. [D]

Data. 11,701 fill-level readings for 449 anonymized containers, covering 2025-10-17 to 2026-09-30 (almost one year). Three glass colours: white (4,432 readings), green (4,169), brown (3,100). Numeric fill levels run from 0 to 100; the string overfull appears in 1,192 rows. The file has only four columns: Location_id, material_type, date, fill_level_perc. [D]

Spatial resolution. No coordinates. Container IDs are stable and anonymized, so per-container history and grouping work, but route geometry, inter-point distances, and depot distances are absent. The spatial object is a graph or attribute field over containers, not a geographic field. [D]

Temporal resolution. About one year, but readings are event-like and irregular, not a complete daily grid. Dates only, no time of day. Values cluster on round percentages, suggesting reported or estimated levels rather than continuous sensor measurements. [D]

Key uncertainty. Reporting error and discreteness; how to encode overfull; missing container capacity and type; no fleet, depot, or shift data; no ground-truth routes; unknown reading protocol; holidays and school vacations are suggested as drivers but not supplied as a structured calendar. [D]

Heatmap / spatial field. A fill-level map would be the natural visual, but without coordinates it can only be a container graph or zone-level attribute map. Any geographic heatmap requires an external location mapping or an explicit synthetic layout. [D]

Prediction target. Future fill level per container, or time-to-threshold / overflow risk, then a route. The brief emphasizes pattern analysis and route optimization; a forecasting target is implicit rather than specified. [D]

External covariates. Material type, public holidays, school vacations, seasonal shifts, and the operational variables named in the brief: waste stream and density, container specifications, fleet characteristics, inter-point distances, disposal depot proximity, depot distances, and shift constraints. [D]

Cross-challenge link. The Pulse of the City brief explicitly names PreZero as a user: historical footfall heatmaps could inform waste collection routes around high-traffic zones, and predictions could support dynamic routing. This is a documented conceptual bridge between C1 and C2, but it is a use-case suggestion, not a shared dataset or a validated pipeline. [D]

C3 — Measure where it matters (AI TRAQC)

Decision. Where to place stationary heat sensors, which routes mobile units take and when, and how many measurement locations are needed, given budget, mission time, and priority areas such as day-care centres or care homes. The deliverable is positions and routes, not only a corrected map. [D]

Data. Open satellite data: Sentinel-3 land surface temperature at roughly 1 km per pixel, daily; Landsat 8/9 at roughly 100 m, every eight days, around midday, cloud-free; optional ECOSTRESS at different times of day, but only up to about 52 degrees north. Suggested open inputs: the global Local Climate Zone map at about 100 m and OpenStreetMap for the path network. No challenge dataset is stored in the bundle; the deck describes data sources and a simulation workflow. [D]

Spatial resolution. A coarse 1 km baseline, a 100 m target/reference grid, and city-scale placement. Example constraints include 5+5 or 10+2 sensors. [D]

Temporal resolution. Satellite snapshots around midday; Landsat repeats every eight days; there is no continuous time series. The "when" needs a separate time axis, via ECOSTRESS scenes or a modelled diurnal cycle. [D]

Key uncertainty. Land surface temperature is not air temperature; clouds and revisit limits; the simulation samples a same-day Landsat reference, so sim-to-real transfer is untested; priority-area weighting versus overall map error; how much time of day changes the optimal placement; transfer to another city. [D]

Heatmap / spatial field. This is a literal heat map: a land surface temperature field. The task is to reconstruct a fine-grained field from sparse ground/mobile samples plus a coarse baseline, and to choose the sampling design that minimizes reconstruction error. [D]

Prediction target. A reconstructed 100 m land surface temperature field, optionally at different times of day. This is spatial prediction, not a temporal forecast of future demand. [D]

External covariates. Local Climate Zones, OpenStreetMap, built environment and surface characteristics, satellite bands, time of day. The bonus asks whether the approach also works for humidity or noise. [D]

What the challenge tests. The algorithm must choose placements from coarse data only, without peeking at the fine reference, then be evaluated honestly against Landsat. An error-versus-number-of-measurement-points curve is the suggested justification. [D]

C4 — LoRaWAN network planning (HNVG)

Decision. How many gateways a city needs, where they should go, and what coverage and redundancy can be expected. The output is a first planning tool that lets a user select candidate sites and assess coverage and gaps. The brief explicitly says this is not a replacement for professional radio planning or field measurements. [D]

Data. 8,009 reception rows with 31 columns, from 448 distinct devices and 9 distinct gateway IDs, over 2026-11-04 to 2026-11-10 (six days). Transmitter metadata has 593 meters; gateway metadata has 8 rows; 71 candidate gateway sites; 23,127 addresses. RSSI ranges from -127 to -70 dBm; SNR from -21.5 to 12 dB. Spreading factors: 7 (3,494 rows), 8 (372), 9 (691), 10 (733), 11 (609), 12 (2,110). [D]

Spatial resolution. Link-level point-to-point receptions between meters and gateways, plus a coverage field over candidate sites and addresses. Transmitter GPS coordinates are of uncertain quality; gateway and candidate-site coordinates use a decimal comma in the delivered files. [D]

Temporal resolution. Six days of messages. This is enough for a coverage snapshot but not for a temporal forecast of reception, traffic, or interference. [D]

Key uncertainty. Received messages alone do not establish packet loss; a reliable estimate of expected transmissions is missing. One observed gateway (00007076FF03010C) has no metadata row, so the "8 gateways" and "9 gateway IDs" mismatch must be resolved. 145 mapped meters did not transmit during the six-day window. Radio settings, building/terrain obstructions, vegetation, interference, and gateway height are not fully supplied. [D]

Heatmap / spatial field. A coverage or reception-quality field: expected RSSI, SNR, or link success probability over space, plus redundancy where multiple gateways hear the same message. This is a planning heatmap, not a temporal demand forecast. [D]

Prediction target. Reception success or quality at a location or link, and coverage/redundancy for a candidate gateway set. This is spatial prediction and site ranking, not a time-series forecast. [D]

External covariates. Distance, buildings and terrain, spreading factor, bandwidth, transmit power, gateway height, vegetation, interference, and network traffic. The brief lists these as examples to test, not a checklist. [D]

What the challenge tests. Whether a city can get a useful first answer from real urban data rather than a perfect radio-planning model, and whether the approach transfers to other cities and regions. Partial solutions are explicitly welcome. [D]

C5 — Riding the Sentiment (Stadt Heilbronn)

Decision. What the city should understand and communicate about public debate around the planned cable car. The brief asks for three to five key insights and an explorable dashboard; it does not ask for a resource-allocation decision, a forecast, or a policy optimizer. [D]

Data. 39,361 records from 2024-10-19 to 2026-10-27, collected from publicly accessible content without login and without circumventing access restrictions. The export is about 210 MB. It was pre-filtered with the keywords IPAI, Seilbahn, and téléphérique. Fields include media type, source, language, sentiment (1, 0, -1, -99), reach, engagements, authors, title, and content. [D]

Spatial resolution. No physical spatial field. "Space" is the comparison between projects and cities: Heilbronn's planned cable car, the Câble C1 in the Paris region, the planned cable car in Bonn, and the IPAI campus in Heilbronn. [D]

Temporal resolution. About two years, event-driven and irregular. Publication timestamps support trend, peak, and change-point analysis, but not a regular forecasting grid. [D]

Key uncertainty. The export is a media-monitoring sample, not a representative survey. Keyword filtering captures tourist cable cars, toys, and unrelated projects. Sentiment is not stance: a negative-sounding statement may still support the project. The -99 value is a sentinel and should be treated as unknown, not as extreme negative. Duplicates, mixed languages, media-type imbalance, missing reach, and platform bias all limit inference. [D]

Heatmap / spatial field. None. The nearest analogues are a topic/stance landscape over time or a comparison matrix across projects and cities. Calling this a heatmap would be a metaphor, not a measurement. [D]

Prediction target. None specified. The task is descriptive and classificatory: topics, arguments, sentiment/stance, temporal peaks, and cross-project comparison. A forecast of attention is possible in principle but is not asked for and has no ground truth in the bundle. [D]

External covariates. Real-world developments: political decisions and announcements, new cost estimates, route or planning changes, visualisations, construction milestones, media reports, and the opening of a new system. Also media type, language, source, and platform. There is no structured event log in the dataset. [D]

Coverage details. 28,805 records have a non-null reach; total engagements are 1,308,843; German-language records are 30,812 and French-language records 7,999; news records are 15,594. Sentiment counts: 1 = 20,712, 0 = 9,928, -99 = 6,797, -1 = 1,924. [D]

C6 — Learning Guide (aim)

Decision. Which learning and leisure offers to present to a family, and how to match them to a child's age, interests, date, and location. The brief is a product-design and data-integration task, not a sensing or forecasting task. It also asks how the application stays current with low maintenance. [D]

Data. No dataset is supplied. The prototype is limited to four institutions: aim, experimenta, Haus der Familie, and IPAI. The thematic focus is AI, coding, and gaming. Teams must check for APIs, RSS feeds, iCalendar feeds, or CSV/JSON exports before scraping, and must check whether extraction is permitted. [D]

Spatial resolution. Institution locations and, optionally, distance from home or school via OpenStreetMap. There is no sensor field and no continuous spatial measurement. [D]

Temporal resolution. Event and course dates, opening hours, and content freshness. There is no forecasting target and no regular time series. [D]

Key uncertainty. Offer coverage and data quality; website structure changes; scraping legality and terms; child privacy and GDPR compliance; age-appropriate design; no attendance or demand ground truth; long-term maintenance burden. [D]

Heatmap / spatial field. None. A map of offer locations or distances is possible as a supporting view, but it is a catalogue map, not a measured spatial field. [D]

Prediction target. None. The relevant target is search or recommendation relevance, and secondarily demand/supply gaps from anonymous queries. Neither has ground truth in the bundle. [D]

External covariates. Age group, theme (AI, coding, gaming), date, event type, location/distance, language, accessibility, and institution calendars. [D]

Privacy regime. This is the highest-privacy challenge in the set. The brief asks for data minimisation, ideally no stored data; guest or incognito modes; no self-registration by underage users; parental consent under GDPR, where young people can consent from age 16 in Germany; child-friendly explanations; and caution with AI chatbots and third-party services. [D]


2. Shared abstraction

2.1 Formal core

The common structure is a partially observed decision problem, not a universal heatmap:

observation:   y_t      = H_t x_t + e_t
field:         x_hat_t  = I(y_<=t, H_t, c_t, G)
forecast:      x_hat_t+h = F(x_hat_<=t, c_t+h; theta)
decision:      a*       = argmax_a E[U(a, x_t+h)]  subject to constraints
evaluation:    L(y, x_hat) + R(a*, x_oracle) + privacy leakage
  • x_t is a latent state over spatial units S. S can be a grid (land surface temperature), a set of fixed points (PAX sensors), a graph (glass containers), a set of links or candidate sites (LoRaWAN), or a set of projects (cable-car debate).
  • H_t is the observation operator: point sampling, spatial aggregation, link reception, text collection, or catalogue extraction. In C1/C2/S1 it is fixed. In C3/C4 it is a design variable: the decision changes what is observed.
  • c_t is exogenous context available at decision time: weather, holidays, events, built environment, radio settings, platform, age/theme.
  • G is geometry or graph structure. It may be a metric space, a container graph, a radio link graph, or empty.
  • I is spatial inference: interpolation, kriging, Gaussian-process regression, physics-based propagation, graph smoothing, or simple aggregation.
  • F is temporal forecasting. It is only meaningful when the decision is forward-looking.
  • U is the decision utility. It must be stated explicitly; accuracy alone is not a decision metric.

The key distinction the paper must make is state estimation versus design. In C1/C2/S1, the pipeline estimates a state and then acts on it. In C3/C4, the pipeline predicts the value of a design and then chooses where to measure or transmit. Calling both "forecasting" hides the inversion.

2.2 The seven-layer stack

Layer Question PulseCast implementation Generalization test
L0 Sensing and sampling What is measured, where, how often, by whom? PAX counters: 8 fixed points, ~15 min, 5 months Is H_t fixed or chosen? Does the sampling design match the decision?
L1 Observation and privacy What does y mean, what error/bias, what privacy class? Aggregated device counts; no MAC/PII in the app; penetration unknown Is the observation model calibrated? Is the privacy claim formal or only architectural?
L2 Spatial inference How do sparse observations become a field? Local pulse map and labelled interpolation only Is the field type Euclidean, raster, graph, or link-based? Does resolution match the decision?
L3 Temporal forecast What will happen next? Global HistGB; 7-day hero; 15 min/1 h/48 h/7 d exposed Is the target forward-looking? Is there enough history for the claimed horizon?
L4 External context What explains the signal? Open-Meteo temperature/rain; BW holidays; scenarios Is context available at decision time? Is it observed or forecast?
L5 Decision What should be done? Staffing heuristic with visible assumptions; human accepts/edits Is the objective explicit? Are constraints real? Is uncertainty propagated?
L6 Evaluation Did it work, for the decision? MAE/WAPE/R² vs seasonal naive; plan latency; load test Is there a decision-relevant metric, not only accuracy? Is there decision regret or outcome data?

2.3 Where the abstraction holds

The stack holds when five conditions are met:

  1. A latent state exists with spatial or temporal structure. Footfall, fill level, land surface temperature, and radio coverage all have autocorrelation; a sparse sample is informative about unobserved places or times.
  2. The observation operator is known and stable enough to model. PAX point sampling, container readings, satellite pixels, and radio links each have a defensible H_t, even if it is noisy.
  3. Context explains a meaningful share of variance and is available at decision time. Weather is the clearest example in PulseCast; built environment and radio settings are the analogues in C3/C4.
  4. The decision can be written as an optimization or rule over forecasts with explicit constraints. Staffing, ad allocation, routing, sensor placement, and gateway placement all qualify; "understand the debate" does not.
  5. Evaluation can use a time-based holdout and a decision-relevant metric. PulseCast has the first; it lacks the second.

S1 satisfies conditions 1–4 locally. C1 as posed has a latent state and a plausible observation operator, but it does not supply context or a decision rule. C2 has temporal structure and a routing decision, but its observation model is coarse, its context is only suggested, and its spatial part is missing because there are no coordinates. C3 and C4 satisfy 1, 2, and 4 for predictions, but they invert the arrow: the decision is the sampling design.

2.4 Where the abstraction breaks

Break Mechanism Challenges affected Consequence
B1 — No physical field The latent state is opinion, relevance, or a catalogue, not a spatial quantity C5, C6 The "heatmap" layer is a metaphor; spatial inference is absent
B2 — Adversarial or non-representative observation The sample is selected by keywords, platforms, or self-selection C5 H_t is not a neutral sampling operator; sentiment is not stance
B3 — Decision changes the observation Placement or gateway choice determines what can be measured next C3, C4 The pipeline becomes active learning or optimal experimental design, not forecast-then-act
B4 — Spatial units are anonymized or absent Container IDs without coordinates C2 The field is a graph or attribute vector, not geography; routing needs an external layout
B5 — Resolution mismatch Eight points in 580 m cannot rank city-wide store locations or ad sites C1 ads/store, S1 The forecast may be valid locally but not at the decision's spatial scale
B6 — Privacy regime prohibits individual-level state Children's data, social-media authors, meter/address data C5, C6, C4 Aggregation or data minimisation is required; the modelling unit changes
B7 — No decision outcome data No POS, labour, ad cost/reach, fleet, packet-loss, or stance labels All Accuracy is a proxy; decision regret and ROI cannot be claimed

2.5 What "multimodal" can and cannot mean

The bundle contains at least five modality families:

Modality family Examples Used empirically by PulseCast?
Point/aggregate time series PAX device counts, glass fill levels Yes, PAX only
Raster/field Sentinel-3 and Landsat land surface temperature No
Link/event records LoRaWAN receptions, RSSI/SNR, spreading factor No
Text News, social posts, YouTube comments No
Structured catalogue Institution offers, dates, age groups, themes No

PulseCast empirically exercises one sensor modality (PAX counts) plus tabular context (weather, holidays). That is enough to claim a context-fused, modality-agnostic pipeline contract. It is not enough to claim a single multimodal model trained across modalities. A defensible paper sentence is: "The core is modality-agnostic at the interface, and we demonstrate it on one sensing modality; extending it to raster, link, and text modalities is a research agenda, not a result." [D]


3. Matrix table

The matrix is split into two halves for readability. C1 is the challenge as posed; S1 is our implementation. They are separate rows because the challenge leaves the decision open, while PulseCast fixes it. "Ads" and "store location" are decision layers inside C1, not separate challenges.

3.1 Matrix A — sensing to field

ID Sensing modality Privacy risk Spatial inference (field type) Temporal forecast
C1 Pulse of the City (as posed) Passive Bluetooth PAX counts, aggregated over LoRaWAN; delivered CSV P1: aggregate counts, but MAC sniffing at the edge; penetration unknown; presence/re-identification risk Point field: 8 usable sensors in ~580 x 280 m; local interpolation only; city-wide heatmap unsupported ~15 min cadence; 5 months multi-sensor; prediction requested, horizon unspecified
S1 PulseCast Same PAX stream + Open-Meteo + BW holidays P1: aggregate only in the app; no MAC/PII; no geolocation permission; sensing externality remains Point field: local pulse map; measured points plus labelled interpolation Temporal forecast: global HistGB; 15 min/1 h/48 h/7 d; 7 d hero; 28-day holdout
C2 WasteRouting.AI Fill-level scalars per glass container; event-like readings P1: anonymized container IDs; low personal risk; operational sensitivity Graph/attribute field: no coordinates; container IDs only; route geometry absent Implicit temporal forecast: irregular, ~1 year, date-only; fill level or time-to-threshold
C3 Measure where it matters Satellite LST (Sentinel-3, Landsat, ECOSTRESS) + ground/mobile sensors P0/P1: open satellite data; priority-area maps may reveal vulnerable sites Raster field: 1 km baseline to 100 m target; LCZ/OSM; literal heat field No temporal forecast: snapshots/diurnal; spatial reconstruction is the target
C4 LoRaWAN planning LoRaWAN uplink receptions (RSSI/SNR/SF) from water meters P1: meter/address data; occupancy inference; device IDs Link/coverage field: link-level; 71 candidate sites; 23,127 addresses No temporal forecast: 6-day window; spatial prediction and site ranking
C5 Riding the Sentiment Text/media mentions: news, social, YouTube, forums P2: author handles and quotes; platform terms; high re-identification risk No field: project/city comparison; no physical space No forecast: 2 years, event-driven; trends and change points
C6 Learning Guide Institution offers via web/API/RSS/iCal P2: children's data; GDPR age 16; highest risk No field: institution locations and optional distance only No forecast: event dates and freshness only

Privacy classes. P0 = open/no personal data (satellite, OSM, weather). P1 = aggregated or pseudonymous data with residual re-identification risk (PAX counts, container IDs, meter/address data). P2 = personal or special-category data (social-media authors and quotes, children's data). The classes are a paper device, not a legal determination. [G]

3.2 Matrix B — forecast to decision

ID Prediction target External context Decision layer Evaluation metric
C1 Pulse of the City (as posed) Footfall/device density per location and time; horizon open Weather, holidays, events suggested but not supplied Ads, store location, staffing; no rule specified None specified; business KPI absent
S1 PulseCast Hourly footfall index per sensor; next-week staff plan Open-Meteo temp/rain (cloud/wind no gain); BW holidays (little gain); counterfactual scenarios Staffing heuristic: required staff, minimum staff, shift blocks, cost, coverage; human accept/edit MAE/WAPE/R² vs seasonal naive; plan latency; coverage/cost vs flat schedule; load test target
C2 WasteRouting.AI Fill level per container / time-to-threshold; route Holidays, school vacations, season, material; fleet, depot, shift variables VRP + fleet selection under capacity, shift, and depot constraints; configurable Fill MAE/overflow rate; cost/km; vehicles; no ground-truth routes
C3 Measure where it matters 100 m LST field; optionally diurnal field LCZ, OSM, time of day, satellite bands; humidity/noise bonus Sampling design: sensor positions, mobile routes, number of locations; budget/priority Reconstruction RMSE vs Landsat; error-vs-count curve; priority-area error
C4 LoRaWAN planning Reception quality/success at link or location; coverage/redundancy Distance, buildings/terrain, SF/BW/TX power, height, vegetation, interference Gateway planning: number and sites; candidate-site ranking Link AUC/MAE; coverage % of addresses; redundancy; cost
C5 Riding the Sentiment Stance/topic/time; no forecast Political decisions, cost estimates, route changes, construction, opening; platform/language Communication/insight: no resource allocation Stance F1/accuracy; topic coherence; reproducibility; insight quality
C6 Learning Guide Offer relevance/search; no forecast Age, theme, date, location/distance, language, accessibility Matching/search/product maintenance: no resource allocation Task success; search relevance; accessibility; maintenance effort; privacy compliance

3.3 Matrix C — layer status

Symbols: ● present and evidenced (data and/or documented result); ◐ partial, prototype, or assumption; ○ absent or unsupported; – not applicable.

ID L0 sensing L1 observation/privacy L2 spatial L3 temporal L4 context L5 decision L6 evaluation
C1 as posed ● ◐ ◐ ◐ ◐ ◐ ○
S1 PulseCast ● ● ◐ ● ● ● ◐
C2 waste ● ● ○ ◐ ◐ ◐ ○
C3 sensor placement ◐ ● ● ○ ● ● ●
C4 LoRaWAN ● ◐ ● ○ ◐ ● ◐
C5 cable car ● ◐ ○ ○ ◐ ○ ◐
C6 learning guide ○ ○ ○ ○ ◐ ◐ ○

Reading the matrix. Only PulseCast has an evidenced end-to-end path from sensing to a decision layer, and even there the spatial layer is local and the decision evaluation is a proxy. C3 and C4 are strong on spatial inference and decision design but have no temporal forecast. C5 and C6 have neither a field nor a forecast. C2 has a strong temporal/operational story but no spatial geometry. The matrix therefore supports a conditional generalization claim, not a universal one.


4. Transferability assessment

4.1 Verdict scale

Verdict Definition What it permits in the paper
Direct Same sensing modality, same field, new decision; no new data required A re-analysis or extension of PulseCast
Conditional Same pipeline pattern, but a new dataset, modality, or geometry is required A method-transfer claim with an explicit data condition
Sibling Shares spatial inference and evaluation, but the decision changes the observation process A separate "spatial design" contribution, not the forecast core
Non-application No physical field, no forecast target, no resource-allocation decision At most a context source or a different problem class

4.2 Transfer verdicts

Challenge / decision Verdict What transfers What must change Evidence
C1/S1 staffing Direct Observation handling, local field, weather-aware forecast, staffing heuristic, deployment pattern Store-specific assumptions; city-scale coverage; outcome data [M] MAE 3.36 vs 5.40 naive; WAPE 27.8% vs 44.7%
C1 advertising Direct (local exposure only) Same field and forecast; allocation objective replaces staffing objective Ad inventory, cost, reach, dwell, conversion; city-wide sites for a real campaign [P] no result in bundle
C1 store location Conditional (data-bound) Forecast/context pipeline; ranking logic City-wide field, POI/rent/competition/sales data; eight points cannot rank zones [G] explicitly out of scope in mission/mvp.md
C2 waste routing Conditional Temporal forecasting, clustering, configurable VRP, holiday/season context Coordinates or zone map, capacity/type, fleet/depot/shift, route costs, overflow ground truth [P] no result; dataset documented
C3 sensor placement Sibling (spatial design) Observation model, spatial inference, error-vs-cost evaluation Decision changes H_t; active learning/optimal design; no temporal forecast; no delivered dataset [P] protocol only
C4 LoRaWAN planning Sibling (spatial design) Link/coverage prediction, candidate-site ranking, address-weighted coverage Packet-loss estimate, buildings/terrain, longer window, GPS calibration; no temporal forecast [P] no result; dataset documented
C5 cable car sentiment Non-application At most an event/attention signal for another model No physical field, no forecast target, no allocation decision; different NLP stack and privacy regime [D] descriptive task
C6 learning guide Non-application At most a location/context consumer No sensing, no field, no forecast; matching/recommendation and child-privacy regime are different [D] product task

4.3 What is truly transferable

Direct or same-data transfers.

  • Staffing. The PulseCast path documented in the bundle is the only end-to-end, evidence-backed instance. It can be reused for another store in the same cluster; another city would require retraining and revalidation on that city's sensors and weather.
  • Advertising. The same field and forecast can rank the eight locations and time windows by expected exposure. This is a decision-layer substitution, not a new sensing study. It is transferable as an analytical extension, but it cannot support city-wide ad planning without more sites and ad-economics data.
  • City operations. The same pulse map can support crowd management, event planning, or dynamic street infrastructure, but this is a secondary use case and is not evaluated in the bundle.

Method transfers that require new data or a new geometry.

  • Waste routing. The temporal-forecast and decision-rule pattern transfers, but the spatial object changes from a metric field to a container graph. The route result is conditional on assumed geometry, capacity, and fleet data. Without coordinates, a geographic heatmap or a credible route-saving claim is not possible.
  • Store location. The method can rank sites only if the field is observed at city scale. With eight sensors in 580 m, the correct verdict is not supported, not "promising." This is a data-acquisition proposal, not a case study.
  • Sensor placement. The spatial-inference layer transfers, and the evaluation protocol (error versus number of measurement points) is strong. But the decision is a design decision, and the temporal forecast layer is absent. It is a sibling, not the same core.
  • LoRaWAN planning. Same verdict as sensor placement: a spatial-design sibling with real data, a link-level observation model, and no temporal forecast. It is the best available test of the "spatial field" part of the thesis because the data are real and the candidate sites are explicit.

Not transferable.

  • Cable-car sentiment. There is no physical field, no forecast target, and no resource-allocation decision. The dataset is a media-monitoring sample, not a sensor. It can be an exogenous context signal (for example, an event/attention feature for footfall or waste), but it is not an application of the sensing-to-field-to-forecast core. Its privacy regime (author handles, quotes) is also different from aggregate counting.
  • Learning guide. There is no sensing, no field, no forecast, and no allocation. It is a data-integration, search, and child-privacy product problem. It is the clearest example of why "urban" does not imply "sensing stack."

4.4 The six challenges are not six instances of one problem

A useful taxonomy for the paper:

Problem class Challenges What the core can offer
State estimation + forward decision C1 (staffing, ads, city ops), C2 (waste) The main pipeline: observation, field, forecast, decision rule
Spatial design / active sensing C3 (heat sensors), C4 (gateways) A sibling pipeline: observation model, spatial prediction, optimal design
Text/opinion analysis C5 (cable car) Nothing directly; at most a context signal
Catalogue/matching C6 (learning guide) Nothing directly; a different product class

This taxonomy is more defensible than "six challenges, one stack." It says the stack covers one challenge directly (C1, which contains three decisions), one conditionally (C2), two as siblings (C3, C4), and two not at all (C5, C6). That honesty strengthens the paper: it shows the authors know where the method stops.

4.5 What would falsify the generalization claim

The paper should state these falsifiers explicitly:

  1. Waste forecast fails a naive baseline after accounting for irregular sampling and overfull encoding, or the routing result is driven entirely by assumed geometry.
  2. LoRaWAN link model cannot beat a distance-only baseline on held-out gateways, or candidate-site rankings are unstable under GPS uncertainty.
  3. Advertising allocation is unstable under plausible weather scenarios or measurement error, so the ranking is noise.
  4. Store-location ranking with eight points is no better than random at the city scale.
  5. Decision regret is large despite low MAE, meaning forecast accuracy does not translate into better staffing.
  6. Aggregate counts can be re-identified or the privacy claim fails a threat model.

A paper that reports these tests, even if some fail, is stronger than one that asserts six-way generalization.


5. Case studies

Four case studies are enough for a 4–5 page report. They are ordered by evidence strength, not by ambition. A fifth candidate (AI TRAQC) is described as a protocol because the bundle contains no delivered dataset.

5.1 Case study 1 — Footfall to staffing (PulseCast, empirical anchor)

Research question. Can eight sparse, privacy-preserving PAX counters plus weather support a weekly staffing plan better than a seasonal-naive baseline?

Why it tests the thesis. It is the only case with an end-to-end path: real sensor data, a time-based holdout, a baseline, a forecast, a decision rule, and a deployed application. It establishes the upper bound of what the current evidence supports.

Data we have. 192,464 CSV rows; 96,232 density readings; 12 device IDs; after merging the two density-01 IDs and excluding three sensors, 8 usable sensor locations; 2026-06-12 to 2026-11-07; ~15-minute cadence; ~580 m x 280 m cluster; values 0–292.6; multi-sensor mean ~14.4; median 7.7; p95 46.9; p99 78.4. Peak 18:00–21:00 local, top at 19:00; Saturday busiest (mean 19.1). [D]

Results we have. Global HistGradientBoostingRegressor; train 2026-06-12 to 2026-10-10; test last 28 days. Seasonal naive: MAE 5.40, WAPE 44.7%, R² 0.462. Day-ahead, no weather: MAE 4.20, WAPE 34.8%, R² 0.678. Day-ahead with temperature/rain: MAE 3.36, WAPE 27.8%, R² 0.779. Seven-day-ahead, no weather: MAE 4.37, WAPE 36.2%, R² 0.672. Holidays added little: 4.15 vs 4.20 without weather; 3.44 vs 3.36 with weather. [M]

What else we have. A documented three-service deployment on SKE (Nuxt SSR, Go API/SSE, FastAPI) behind a Cloudflare Tunnel, with no Ingress, cert-manager, or Jobs; build-time batch; cached fallback; four forecast horizons; a staffing heuristic with configurable assumptions. Plan generation is documented as under 2 seconds; the k6 load-test number is still [TODO]. [D/P]

What is missing. POS/sales/queue data; actual manager decisions; labour cost and schedule data; forecast weather rather than observed weather; an event log; school holidays and bridge days; multi-store and multi-city validation; annual seasonality; per-sensor MAE table (required by mission/paper.md, not present in the bundle); a calibrated uncertainty interval; a decision-regret evaluation; the k6 result. [G]

Claim supported. In one 580 m cluster, over a 28-day holdout, a weather-aware global model reduces day-ahead MAE by about 20% versus a no-weather model and about 38% versus seasonal naive, which is enough to support a relative staffing plan. [M]

Claim not supported. A city-wide heatmap; a census of people; annual seasonality; causal weather effects; ROI or productivity gains; a general multi-city model; live sensor data. [G]

Effort and risk. Low: the work is already done. The risk is overclaiming. The paper should present this as a single-site, short-horizon decision-support result, not as proof of generalization.

5.2 Case study 2 — Footfall to advertising allocation (same field, new decision)

Research question. Can the same forecast rank the eight locations and time windows for physical advertising exposure, and how sensitive is the ranking to weather scenarios and measurement error?

Why it tests the thesis. It isolates the decision layer. The sensing modality, field, and forecast are unchanged; only the objective changes from staffing to exposure allocation. If the core is modular, this should be a cheap and credible extension. If it is not, the "generalizes across decisions" claim is weak.

Data we have. The same PAX data and PulseCast forecast. No ad inventory, cost, reach, frequency, dwell time, audience composition, or conversion data. [D]

Analysis we can do now. Rank the eight locations by forecasted footfall and stability; compute exposure-weighted allocations under an explicit equal-cost/equal-reach assumption; run the sunny-Saturday, rainy-Saturday, and weekday-event scenarios; report how the ranking changes. This is a re-analysis of the same field, not an ad-effectiveness study. [P]

Results we have. None. The bundle contains no advertising result. Any lift, CPM, or revenue number here would be invented. [G]

What is missing. Ad inventory (sites, formats, prices); reach and frequency; audience segmentation; dwell time; store conversion; city-wide sites; a control/treatment design for causal inference. [G]

Claim supported. The forecast can support a local, exposure-based ranking of the eight measured locations and hours, and the decision layer can be swapped without retraining the sensor model. [P]

Claim not supported. Ad effectiveness, incremental revenue, city-wide optimal placement, or causal impact of advertising. [G]

Effort and risk. Low for the exposure ranking; high for any credible effectiveness claim. The paper should label this as an analytical extension, not a second empirical validation.

5.3 Case study 3 — Fill level to waste routing (new modality, graph field)

Research question. Can fill-level histories produce container-level forecasts and a fill-level-driven routing policy that reduces overflow or cost relative to a fixed area schedule?

Why it tests the thesis. It is the strongest test of modality transfer: the sensor stream is scalar and event-like rather than a 15-minute count, and the decision is a vehicle-routing problem rather than a staffing heuristic. It also exposes the spatial gap: the data have no coordinates, so the "field" must become a graph.

Data we have. 11,701 readings; 449 anonymized containers; 2025-10-17 to 2026-09-30; white 4,432, green 4,169, brown 3,100; overfull 1,192; numeric 0–100; four columns; no coordinates, capacity, fleet, depot, or shift data. [D]

Analysis we can do now. Per-container fill trajectories; clustering by fill rate and seasonality; time-to-threshold or survival models; holiday/school-vacation effects; a synthetic or assumed geometry for routing; a VRP baseline versus a fixed schedule under explicit assumptions. The forecast part is defensible; the routing part is conditional on assumed geometry. [P]

Results we have. None. The bundle documents the dataset but contains no model result, and the raw CSV is not present in this workspace; only the concept file is. [G]

What is missing. Coordinates or a zone mapping; container type and capacity; fleet capacity/weight limits; depot locations; shift rules; route costs; overflow ground truth; the reading protocol; the raw file for reproduction. [G]

Claim supported. The temporal-forecast and decision-rule pattern transfers to a second scalar sensor stream, provided the spatial layer is redefined as a container graph or zone-level attribute field. [P]

Claim not supported. Route savings, reduced emissions, or overflow reduction without real geometry and fleet data. A geographic heatmap of fill levels without coordinates. [G]

Effort and risk. Medium-high. The forecast is feasible; the routing evaluation is the bottleneck. If the paper includes this case, it must show an explicit synthetic-geometry sensitivity analysis and label it as such.

5.4 Case study 4 — LoRaWAN reception to gateway planning (inverted arrow)

Research question. Can reception data from 448 observed devices and 8 gateways (9 observed gateway IDs) predict coverage and rank 71 candidate gateway sites, compared with a distance-only baseline?

Why it tests the thesis. It is the best available spatial-design case with real data. It tests whether the observation model, spatial prediction, and decision layer transfer when the decision is "where should the infrastructure go?" rather than "how many staff next week?" It also exposes the limit: there is no temporal forecast.

Data we have. 8,009 reception rows; 31 columns; 448 devices; 8 gateway metadata rows and 9 observed gateway IDs; 593 meters; 71 candidate sites; 23,127 addresses; six days (2026-11-04 to 2026-11-10); RSSI -127 to -70 dBm; SNR -21.5 to 12 dB; spreading factors 7–12; 145 mapped meters silent. [D]

Analysis we can do now. A link-level model of reception quality or success using distance, radio settings, and building/terrain proxies; leave-one-gateway-out validation; candidate-site coverage and redundancy scoring; address-weighted coverage; a distance-only baseline. [P]

Results we have. None. No model metrics are documented in the bundle. [G]

What is missing. A packet-loss or expected-transmission baseline; building heights and terrain; gateway heights; GPS calibration; resolution of the ninth gateway; a longer window; field validation of predicted coverage. [G]

Claim supported. Spatial prediction and candidate-site ranking transfer as a sibling of the forecast core, with a different observation model and a different decision objective. [P]

Claim not supported. Temporal forecasting of reception; a replacement for professional radio planning; coverage guarantees without packet-loss data. [G]

Effort and risk. Medium-high. The data are real and the candidate sites are explicit, but the missing packet-loss baseline is a serious limitation. A leave-one-gateway-out protocol is the minimum credible evaluation.

5.5 Candidate protocol, not a case study — Satellite LST to sensor placement (AI TRAQC)

Research question. Where should k stationary sensors be placed, and which mobile routes should be run, to reconstruct a 100 m land surface temperature map from a 1 km baseline?

Why it matters. This is the purest test of the literal heatmap claim and of the spatial-inference layer. It also makes the design inversion explicit: the decision is the sampling operator.

Data we have. No delivered dataset in the bundle. Open Sentinel-3, Landsat 8/9, optional ECOSTRESS, Local Climate Zone, and OpenStreetMap are described. [D]

What a minimal protocol would need. A city boundary and priority areas; cloud-free Landsat reference scenes; a Sentinel-3 baseline; an LCZ/OSM feature layer; a placement algorithm (greedy, GP-based, or Bayesian optimal design); a mobile-route planner; a reconstruction method; and an error-versus-number-of-points curve. [G]

Why it is not a full case study here. No dataset is supplied, no field deployment exists, and the paper's time budget is already tight. Including it as a protocol is honest; presenting it as a result would not be.


6. Gaps

6.1 Data gaps

Gap Challenges affected Consequence for the paper What would close it
Spatial coverage C1/S1, ads, store location Eight points in 580 m cannot support city-wide heatmaps, ad-site ranking, or store-location ranking More sensors or city-wide mobility data; explicit local-only scope
Temporal coverage C1/S1, C2 Five months of multi-sensor data; no annual seasonality; 28-day test; waste readings irregular Multi-season data; longer holdout; forecast weather; event calendar
Observation calibration C1/S1 Device penetration unknown; counts are devices, not people Penetration study or calibration assumption with sensitivity analysis
Spatial geometry C2, C4 No container coordinates; uncertain meter GPS; no building/terrain Zone mapping; cadastral/OSM/DEM; GPS validation
Decision outcomes All No POS/labour, ad cost/reach, fleet/route, packet-loss, or stance labels Operational data sharing; pilot or A/B design; labelled validation set
Privacy evaluation C1/S1, C4, C5, C6 No formal threat model, DP/k-anonymity, or leakage metric Privacy audit; aggregation thresholds; formal guarantee where possible
Multimodal data C1/S1 vs C3/C4/C5 Only one sensor modality is empirically fused with context Joint dataset or a clearly scoped interface-level claim
Text labels C5 No stance gold set; keyword noise; -99 sentinel Manual annotation; relevance classifier; inter-annotator agreement
Offer data C6 No dataset, no attendance, no demand ground truth Institution APIs/exports; legal review; anonymous query analytics
Reproducibility C2, C3, C4 Raw files are not all in the workspace; PAX raw is git-ignored Documented data-access procedure; derived aggregates; model card

Documentation inconsistency to reconcile. mission/data.md reports the cluster as ~580 m x 280 m; mission/paper.md reports 584 m x 284 m for the same cluster. Both are approximations, but the paper should use one number consistently. [G]

6.2 Method gaps

  • No formal observation model. PulseCast treats the PAX count as a direct signal; there is no measurement-error model or detection-function correction. C4 has the same issue with packet loss.
  • No spatial-inference baseline. PulseCast does not compare interpolation methods; with eight points, a GP/kriging baseline and a leave-one-sensor-out test would make the local field claim much stronger.
  • No probabilistic calibration. The forecast exposes an uncertainty band, but the bundle does not document calibration (coverage of intervals, CRPS, or quantile loss).
  • No decision regret. The staffing heuristic is evaluated by forecast accuracy and plan latency, not by whether the plan improved coverage or cost in practice.
  • No causal design. Weather coefficients are predictive, not causal. The what-if scenarios are counterfactual inputs, not identified effects.
  • No active-learning baseline. C3 and C4 need a comparison between the proposed placement and a random or greedy baseline.
  • No VRP baseline. C2 needs a fixed-schedule baseline and a distance-only or fill-threshold baseline.
  • No link-budget/physics baseline. C4 needs a distance-only or free-space-path-loss baseline before a learned model can claim value.
  • No cross-city validation. Every claim is single-city, and most are single-site.
  • No multimodal fusion experiment. Cloud/wind did not improve the short test; no second sensor modality was fused.

6.3 Claim gaps

Claim Risk Safer formulation
"City-wide heatmap" Eight sensors in 580 m cannot support it "Local pulse map with measured points and labelled interpolation"
"Generalizes across six challenges" Only one end-to-end empirical case "Transfers as a contract across a subset; two challenges are siblings and two are out of scope"
"Privacy-first" Aggregation is architectural, not a formal guarantee; PAX sensing has externalities "Aggregate-only handling with no MAC/PII in the app; formal privacy evaluation is future work"
"Multimodal forecasting" Only PAX + tabular context is empirically fused "Modality-agnostic interface, demonstrated on one sensing modality"
"Production-ready" Hackathon deployment; no auth/HA/live ingestion "Deployed hackathon system with health checks, limits, and documented gaps"
"Real-time" CSV, no live feed; 15 min/1 h are replay/nowcast "Replay/nowcast layer; no live sensor ingestion"
"Decision-grade" No decision outcome data "Decision-support with explicit assumptions; decision-regret evaluation is future work"
"Weather improves accuracy" Observed weather, 28-day test, one site "Observed weather improves short-horizon accuracy in this test; forecast-weather validation is needed"

6.4 Evidence status

Claim Evidence in the bundle Status What would settle it
Sparse footfall + weather beats seasonal naive MAE 3.36 vs 5.40; WAPE 27.8% vs 44.7%; R² 0.779 vs 0.462; 28-day holdout Supported, single site, short test Multi-season, multi-site, forecast-weather replication
Weather is a real signal 4.20 to 3.36 MAE (~20%); 5.40 to 3.36 (~38%) Supported in test Ablation across seasons and sites
Holidays improve accuracy 4.15 vs 4.20; 3.44 vs 3.36 Not supported; little effect Larger sample; event model; school holidays
The pipeline transfers to waste No result; dataset documented Untested Fill forecast + routing with real geometry
The pipeline transfers to ads No result; same data Untested analytical extension Ad inventory/cost/reach; decision evaluation
The pipeline transfers to store location No; eight points Not supported at city scale City-wide field and commercial data
The pipeline transfers to sensor placement No result; protocol only Untested sibling Satellite pipeline + error-vs-count experiment
The pipeline transfers to LoRaWAN No result; data documented Untested sibling Link model + leave-one-gateway-out site ranking
Cable car / learning guide are applications No Not supported Different stack; at most context/consumer
Privacy-first Aggregate counts, no MAC/PII in app Partially supported Threat model; formal privacy audit
Production-ready Deployed design, health checks, limits Partially supported HA/auth/live ingestion/ops
Real-time CSV, replay/nowcast Not supported Live feed

6.5 Minimum additions for a credible paper

P0 — must do before submission.

  1. Scope the thesis. Replace "generalizes across six challenges" with "a decision-grade sensing contract transfers across a subset of urban problems; we show where it stops."
  2. Keep the PulseCast numbers exact and baseline-anchored. Report MAE/WAPE/R², the 28-day holdout, the weather ablation, and the honest holiday result.
  3. Relabel the heatmap. Use "local pulse map" or "measured points plus labelled interpolation"; never "city-wide heatmap."
  4. Include the transferability matrix with evidence status. The matrix in this note is designed to be adapted directly into the paper.
  5. Add one decision-relevant evaluation. At minimum, compare the staffing plan to a flat schedule on coverage and cost under the stated assumptions; label it as a heuristic comparison, not ROI.

P1 — should do if time allows.

  1. Add one additional empirical case. The cheapest is the advertising exposure re-analysis (same data, new objective). The most informative is waste (new modality) or LoRaWAN (spatial design), but both need substantial data work.
  2. Add a per-sensor error table and a leave-one-sensor-out check. This tests whether the global model is learning shared structure or memorising sensor IDs.
  3. Add uncertainty calibration. Report interval coverage or a proper scoring rule, not only a band on a chart.
  4. Add a privacy paragraph with a threat model. State what is aggregated, what is not stored, and what residual risk remains.
  5. Fill the k6 load-test number. Do not claim 120 clients without the measured result.

P2 — future work, not paper claims.

  1. Waste forecast + synthetic-geometry routing sensitivity.
  2. LoRaWAN link model with leave-one-gateway-out validation.
  3. AI TRAQC placement protocol with an error-versus-count curve.
  4. Cross-city replication of the PulseCast model.
  5. A joint multimodal dataset that would justify the word "multimodal" empirically.

6.6 Suggested thesis sentence for the paper

We do not claim that one model solves six challenges. We claim that a decision-grade sensing contract — explicit observation model, resolution-matched spatial inference, context-fused forecast, separated decision rule, and baseline-plus-regret evaluation — transfers across a subset of urban problems, and we show precisely where it stops.

This sentence is consistent with every number and every gap in the bundle. It is stronger than an unsupported six-way generalization because it makes the boundary conditions part of the contribution.


Source map

Source Used for
OKF_Vault/knowledge/challenges/*.md Challenge decisions, data descriptions, resolutions, uncertainties, deliverables
OKF_Vault/knowledge/data/*.md Dataset counts, schemas, coverage, data-quality notes
OKF_Vault/mission/data.md PAX counts, sensor profile, model metrics, weather/holiday findings, explicit limitations
OKF_Vault/mission/mvp.md PulseCast decision, staffing heuristic, out-of-scope decisions
OKF_Vault/mission/architecture.md Horizons, services, deployment, privacy handling, failure modes
OKF_Vault/mission/paper.md Paper structure, required figures/tables, facts and numbers, verification checklist
OKF_Vault/mission/README.md Product thesis, scope, non-negotiables
OKF_Vault/mission/demo.md Claim discipline, Q&A boundaries, "do not say" list
OKF_Vault/mission/questions.md Settled facts, deployment constraints, open decisions
OKF_Vault/knowledge/index.md Challenge and dataset inventory