# Methodology for explainable outage-risk intelligence Canonical: https://geogridiq.com/docs/methodology/ Methodology version: 2.2.0 Published: 2026-06-07 Last modified: 2026-08-04 Facts reviewed: 2026-08-04 Next review: 2026-10-30 Responsible team: GeoGridIQ Engineering and Research Scope: Public outage-risk evidence, inference, interpretation, and accountability contracts Capability audit state: verified Capability audited as of: 2026-09-08T04:50:43.243576Z Editorial body: 6375 words (glossary, FAQ, sources, and table alternatives excluded) See how GeoGridIQ organizes time-bounded weather, vegetation, outage, geographic, and exposure evidence into reviewable risk signals—and how source health, model trust, uncertainty, and human judgment determine what can be shown. ## Direct answer GeoGridIQ estimates localized outage risk for a declared geography and future time window. It freezes only evidence available by the forecast issue time, aligns that evidence in space and time, and evaluates the resulting snapshot through the exact verified region-and-horizon serving path. Each public result carries its validity window, source freshness, confidence or evidence-quality context, prediction mode, and explanatory drivers; when that complete path cannot be verified, availability fails closed. The output supports monitoring, briefing, and prioritization; it is not an official outage notice, an emergency warning, or a guarantee that a particular asset or customer will lose power. ## Key takeaways - Hazard evidence is an input; it is not an outage observation or an outage forecast by itself. - Probability, relative risk, confidence, severity, consequence, and risk level answer different questions. - Only evidence genuinely available at issue time may enter the frozen forecast snapshot. - Trusted artifact, runtime load, current batch, valid row, display eligibility, and public availability are separate gates. - Missing, failed, stale, partial, unsupported, valid zero, and fallback are distinct states. - Drivers explain score associations; they do not prove physical cause. - GeoGridIQ supports human review and later accountability, not autonomous action or guaranteed outcomes. ### Figure — From evidence to accountable review **Evidence class:** Conceptual illustration **Purpose:** Orient readers to the complete evidence, trust, review, and learning chain. **Long description:** Source evidence flows through spatial and temporal validation into a frozen snapshot. The path then resolves an exact verified public-serving path or fails closed, attaches an explanation, supports operator review, and finally compares the stored forecast with held-back outcomes. Separately contracted fallback scoring is not public-serving unless its own public contract is implemented and verified. No numeric prediction or observed risk map appears. - **1 — Source evidence**: Outage, weather, vegetation, GIS, exposure, and history - **2 — Align and freeze**: Provenance, freshness, space, time, and feature contract - **3 — Resolve path**: Exact verified public-serving path or unavailable - **4 — Explain**: Validity, uncertainty, mode, and associated drivers - **5 — Review**: People compare official information and apply procedure - **6 — Validate**: Held-back outcomes support later evaluation and monitoring **Caption:** Evidence becomes usable only after alignment, trust checks, explanation, review, and later validation. ## 1. Methodology at a glance GeoGridIQ starts by defining a qualifying outage outcome, geography, issue time, and future window. It then gathers permitted evidence, checks provenance, freshness, and completeness, aligns sources in space and time, and freezes a versioned feature snapshot. Runtime controls resolve the exact region-and-horizon model or separately governed scoring, persist the intended prediction population, attach explanation and uncertainty, and apply trust, validity, and display gates. Fallback scoring may exist under its own contract, but public capability remains unavailable unless a separately implemented public-serving fallback contract is verified. Later outcomes are matched under a separate validation contract. No source, map, alert, or artifact becomes a public claim without these intermediate contracts and gates. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### Figure — The complete methodology pipeline **Evidence class:** Conceptual illustration **Purpose:** Teach the complete chain between a forecast question and later accountability. **Long description:** Eleven ordered steps: define the forecast question; collect source evidence; check provenance, freshness, and completeness; align evidence in space and time; freeze an issue-time snapshot; resolve the exact verified serving path or fail closed; generate and persist; attach explanation and uncertainty; apply trust, validity, freshness, and display gates; compare with later outcomes; and feed reviewed results into monitoring and future model work. - **1 — Define**: Target, geography, issue time, valid window, horizon, population, and use - **2 — Collect**: Only approved source families for the exact contract - **3 — Check**: Provenance, freshness, completeness, coverage, and quality - **4 — Align**: Versioned spatial and temporal normalization - **5 — Freeze**: One reproducible issue-time feature snapshot - **6 — Resolve**: Exact trusted region/horizon model or authorized alternative - **7 — Generate**: Complete intended population persisted with provenance - **8 — Explain**: Mode, uncertainty, validity, and associated drivers - **9 — Gate**: Trust, source, batch, expiry, and display controls - **10 — Evaluate**: Held-back outcomes matched under a declared contract - **11 — Learn**: Reviewed results inform monitoring and future work **Caption:** No source goes directly to a public claim. Figure sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 2. What GeoGridIQ is actually estimating A forecast contract names the target outcome, scoring geography, issue time, valid-from and valid-to times, horizon, eligible observations, prediction population, decision use, scoring mode, and availability state. A number without that context is incomplete. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Observed outages describe events already reported; hazard forecasts describe expected conditions; vegetation signals describe screened context; separately authorized relative risk ranks eligible places; and calibrated probability estimates a declared event. Severity describes magnitude, consequence describes impact, and recommendations add policy and human judgment. These meanings do not substitute for one another. A regional or grid-cell score applies only to that unit and does not identify an exact line, pole, transformer, address, building, or customer. Sources: [Environment and Climate Change Canada — Weather alerts](https://www.canada.ca/en/services/environment/weather/severeweather/weather-alerts.html), [United States Geological Survey — Landsat Normalized Difference Vegetation Index](https://www.usgs.gov/landsat-missions/landsat-normalized-difference-vegetation-index), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### Figure — A complete forecast contract **Evidence class:** Conceptual illustration **Purpose:** Make a complete forecast distinguishable from a bare score. **Long description:** A sanitized two-column contract names a qualifying unplanned-outage outcome, an illustrative analysis cell, issue and valid times, a 24-hour horizon, eligible provider observations, eligible-cell population, monitoring use, prediction mode, and availability. Values are explanatory and are not a current forecast. | Contract field | Sanitized example and meaning | | --- | --- | | Target outcome | Qualifying unplanned outage under label contract vX | | Geography | Illustrative analysis cell; not an exact asset or address | | Issue time | 2030-01-15 12:00 UTC; inputs frozen at this cutoff | | Valid window | 2030-01-15 12:00 to 2030-01-16 12:00 UTC | | Horizon | 24H contract | | Eligible outcomes | Verified provider events meeting spatial, temporal, and coverage rules | | Prediction population | All eligible active cells in the declared region | | Decision use | Monitoring and briefing; not autonomous action | | Mode | Conceptual trusted-model example; fallback has different semantics | | Availability | Must pass artifact, batch, validity, source, and display gates | **Caption:** A forecast's meaning travels with its target, geography, time, population, mode, and availability. Figure sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 3. Core terminology The glossary is part of the methodology contract, not ornamental vocabulary. The same terms must retain the same meaning in page copy, status messages, tooltips, exports, schema, and related documentation. In particular, probability, confidence, severity, consequence, and relative risk are not interchangeable; missing is not zero; a promoted artifact is not a current forecast; and a top driver is not a cause. Essential definitions appear in the visible body and glossary rather than being available only on hover. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### Visible glossary - **Observed outage** — A provider-reported loss-of-service event with recorded observation or status times; it describes evidence about an event, not a forecast. - **Qualifying outage** — An observed event that satisfies the versioned label contract, including provider, planned or unplanned, spatial, temporal, coverage, and any documented threshold rules. - **Forecast** — A time-stamped estimate issued before a declared future valid window, for a stated outcome, geography, population, and prediction mode. - **Issue time** — The cutoff at which forecast inputs are frozen; only evidence genuinely available by this instant may be used. - **valid_from** — The beginning of the future interval to which a forecast applies. - **valid_to** — The end of the future interval to which a forecast applies. - **Horizon** — The forecast contract's lead-time interval, tied to issue time and/or the end of the valid window as defined by implementation. - **Region** — A canonical supported geographic scope; it is not automatically a utility service territory. - **Grid cell** — A bounded analysis unit used to align and score evidence; it does not identify an exact asset or address. - **Feature snapshot** — The frozen, versioned set of typed feature values, missingness and freshness flags, and identities for one issue time, region, horizon, and grid contract. - **Source contract** — The rules for a source's provider or product identity, timestamps, units, coverage, quality, freshness, and permitted use. - **Feature contract** — The ordered schema of feature names, types, units, transformations, requiredness, and compatibility used consistently in training and inference. - **Label contract** — The versioned rules that determine which observed events count as positive, negative, or unknown outcomes and how they match forecasts. - **Model artifact** — A serialized model file produced by a particular run; its existence alone does not make it trusted, promoted, loaded, or current. - **Model registry** — The authoritative record that binds a region, horizon, family, version, artifact, integrity metadata, trust state, and promotion state. - **Model card** — A reviewed evidence summary describing intended use, data and evaluation scope, metrics, limitations, and governance decisions for a model. - **Prediction batch** — A provenance-bound set of outputs generated together for a declared region, horizon, issue time, and valid window. - **Probability** — An estimated likelihood of the declared event for the declared geography and window, only for a track designed and calibrated to support that meaning. - **Relative risk** — A rank, score, or percentile comparing eligible units; it is not automatically a calibrated probability. - **Risk level** — A human-readable category assigned by current, verified thresholds and rules; it is not a universal utility standard. - **Severity** — The magnitude or extremity of relevant conditions or a verified classification; it is not automatically outage probability. - **Confidence** — The implemented assessment of evidence and prediction-path quality; it is not a statistical confidence interval or a second probability. - **Data completeness** — Whether all records or layers required by a contract are present, regardless of age. - **Data freshness** — Whether evidence is recent enough for its declared validity or use contract, regardless of completeness. - **Consequence** — The potential importance or impact if disruption occurs; it can affect priority without changing physical likelihood. - **Critical infrastructure exposure** — Generalized context indicating that important community functions may be affected; it is not proof of outage cause or increased likelihood. - **Top driver** — An input or feature family most associated with a particular score under the implemented explanation method; it is not proven causation. - **Fallback** — An explicitly authorized non-primary scoring path with its own source contract and labels; it is not a machine-learning probability unless independently validated as one. - **Trusted** — A status indicating that a specific model identity passed required evidence and governance gates; it does not guarantee a current batch. - **Promoted** — A human-reviewed decision making a trusted artifact eligible for the defined production scope; it does not prove successful runtime loading or current validity. - **Display eligible** — Allowed to appear in a current user interface after presentation gates; hidden rows may still be persisted for audit. - **False positive** — A flagged forecast with no matched qualifying outcome, only where observation coverage is sufficient to support that conclusion. - **False negative** — A qualifying observed outcome that was not covered by the declared forecast and evaluable population. - **Calibration** — The degree to which forecast probabilities align with observed event frequencies under a specified dataset and contract. - **Temporal leakage** — Use, direct or indirect, of information unavailable at issue time or reserved for later evaluation. - **Concept drift** — A change over time in relationships among inputs, outcomes, providers, infrastructure, or operations that can reduce model relevance. ## 4. Evidence and data-source families The active feature manifest determines which source families participate in each region, horizon, and model. Every used source needs provider, time, spatial, transformation, and limitation records; current capability and freshness records remain authoritative. Sources: [GeoGridIQ — Data Sources documentation](/docs/data-sources/), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Related: [Data Sources documentation](/docs/data-sources/) ### 4.1 Verified outage and reliability evidence Verified outage records can include provider and event identity, observed, updated, restored, retrieval, and availability times; active or cleared state; geometry class; supplied customer count or cause; planned status; raw identity; deduplication; and revisions. They serve distinct roles in operational context, historical features, labels, and validation. Only qualifying verified evidence is outage truth: failed or missing coverage is unknown, not zero; a non-match may be unevaluable; and planned outages cannot silently become unplanned positives. Sources: [Hydro-Québec — Power outage FAQ](https://pannes.hydroquebec.com/poweroutages/understand-and-prevent/faq.html), [Hydro-Québec — Understand and prevent outages](https://pannes.hydroquebec.com/poweroutages/understand-and-prevent/), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### 4.2 Forecast weather and hazard evidence Weather contracts select only verified variables and products, potentially including wind, precipitation, snow or ice proxies, temperature change, lightning, antecedent moisture, and separately approved hazards. Values retain product and cycle, issue, forecast-hour, valid and retrieval times, units, grid mapping, and deterministic or ensemble semantics. Late, partial, and superseded cycles are labelled. Official alerts remain authoritative hazard context, not outage observations or automatic outage predictions. Sources: [Environment and Climate Change Canada — Weather alerts](https://www.canada.ca/en/services/environment/weather/severeweather/weather-alerts.html), [Environment and Climate Change Canada — Meteorological Service of Canada free data service](https://www.canada.ca/en/environment-climate-change/services/weather-general-tools-resources/weather-tools-specialized-data/free-service.html), [Environment and Climate Change Canada — GDPS documentation](https://eccc-msc.github.io/open-data/msc-data/nwp_gdps/readme_gdps-datamart_en/), [Environment and Climate Change Canada — RDPS documentation](https://eccc-msc.github.io/open-data/msc-data/nwp_rdps/readme_rdps-datamart_en/), [Environment and Climate Change Canada — HRDPS documentation](https://eccc-msc.github.io/open-data/msc-data/nwp_hrdps/readme_hrdps-datamart_en/), [Environment and Climate Change Canada — GEPS documentation](https://eccc-msc.github.io/open-data/msc-data/nwp_geps/readme_geps-datamart_en/). ### 4.3 Vegetation and remote-sensing evidence Vegetation evidence may include dated NDVI, land cover, density or change, seasonal context, and approved canopy or corridor layers, with acquisition and processing dates, quality masks, cloud effects, coverage, and resolution. NDVI measures greenness—not tree height, conductor clearance, a dangerous branch, or corridor encroachment. Dense vegetation away from infrastructure is not exposure, and satellite screening does not replace field inspection or utility vegetation expertise. Sources: [United States Geological Survey — Landsat Normalized Difference Vegetation Index](https://www.usgs.gov/landsat-missions/landsat-normalized-difference-vegetation-index), [United States Geological Survey — NDVI — foundation for remote-sensing phenology](https://www.usgs.gov/special-topics/remote-sensing-phenology/science/ndvi-foundation-remote-sensing-phenology), [Matikainen et al. — Remote sensing methods for power-line corridor surveys](https://doi.org/10.1016/j.isprsjprs.2016.04.011), [GeoGridIQ — Vegetation provider and data-freshness implementation](/sources/ndvi-data/), [BC Hydro — Tree-management program](https://www.bchydro.com/safety-outages/trees-power-lines/pruning-removing-trees.m.html), [North American Electric Reliability Corporation — FAC-003-5 Transmission Vegetation Management](https://www.nerc.com/pa/Stand/Reliability%20Standards/FAC-003-5.pdf). Related: [Vegetation source methodology](/sources/ndvi-data/) ### 4.4 Geospatial, terrain, and regional context Approved boundaries and grid cells align terrain, land and water masks, authorized access or urban context, and proximity or overlap measures. Coastal, boundary, partial-cell, and varying-area effects retain their handling and resolution. A municipality is not necessarily a utility territory, and infrastructure presence does not establish a legal service area. Sources: [PostGIS Project — PostGIS 3.6 spatial data management](https://postgis.net/docs/manual-3.6/en/using_postgis_dbmanagement.html), [Open Geospatial Consortium — Simple Feature Access standard](https://www.ogc.org/standards/sfa/), [MapLibre — MapLibre GL JS documentation](https://maplibre.org/maplibre-gl-js/docs/). ### 4.5 Infrastructure and network exposure Network features influence a model only when authorized and present in its exact feature contract; sensitive topology remains protected. Generalized critical-infrastructure context usually changes consequence and review priority, not outage probability. Proximity to an important facility does not by itself increase physical likelihood. Sources: [Public Safety Canada — National Strategy for Critical Infrastructure](https://www.publicsafety.gc.ca/cnt/rsrcs/pblctns/srtg-crtcl-nfrstrctr/index-en.aspx), [Public Safety Canada — Current critical-infrastructure overview](https://www.publicsafety.gc.ca/cnt/ntnl-scrt/crtcl-nfrstrctr/index-en.aspx), [Public Safety Canada — Risk Management Guide for Critical Infrastructure Sectors](https://www.publicsafety.gc.ca/cnt/rsrcs/pblctns/rsk-mngmnt-gd/index-en.aspx). Related: [Critical Infrastructure feature](/critical-infrastructure/) ### 4.6 Historical vulnerability Dated prior-outage aggregates can describe recurrence only when their cutoff precedes issue time. Interpretation depends on lookback, provider coverage, spatial aggregation, infrastructure change, and operations. Recurrence is not causation, later events cannot leak backward, and coverage gaps cannot become assumed non-events. Sources: [Taylor et al. — Machine learning evaluation of storm-related transmission outage factors and risk](https://doi.org/10.1016/j.segan.2023.101016), [Taylor et al. — Community power outage prediction modeling for the Eastern United States](https://doi.org/10.1016/j.egyr.2023.10.073), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### 4.7 Operational and user-provided context Saved locations, account context, notes, and crew information are governed separately. Public content excludes crew positions, customer identities, addresses, authentication, billing, tenant-restricted, and private utility data. Suggestions are not autonomous dispatch or proven route optimization without separate validation. Sources: [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### 4.8 Seasonal and climate context Seasonal outlooks, climate signals, and El Nino can inform background planning, but they are not local outage forecasts. A climate index is not an active feature unless the exact manifest proves it, and its timescale differs from operational horizons. Sources: [Government of Canada-led assessment — Canada's Changing Climate Report, Chapter 4](https://changingclimate.ca/CCCR2019/chapter/4-0/), [Government of Canada-led assessment — Canada in a Changing Climate — National Issues, Chapter 2](https://changingclimate.ca/national-issues/chapter/2-0/). ### Figure — Evidence families have different jobs **Evidence class:** Observed public data **Purpose:** Compare each evidence family's purpose, timestamp needs, limits, and activation rule. **Long description:** Rows compare outage, weather, vegetation, GIS, infrastructure, history, operational, and seasonal context. Columns state what each family represents, how it may be used, its essential timing or provenance, what it cannot prove, and that current activation comes only from the exact feature and capability contract. | Family | Represents / permitted use | Required identity and time | Cannot prove alone | Activation | | --- | --- | --- | --- | --- | | Outage | Observed service-event context, labels, history, validation | Provider, event, observed/updated/restored/available/retrieved times | Complete coverage, cause, or a negative when feed health is unknown | Exact current source contract | | Weather | Forecast environmental stress and official hazard context | Product, cycle, issue, forecast hour, valid and retrieval times | That an outage will occur | Exact region/horizon feature manifest | | Vegetation | Greenness, land-cover, change, or authorized corridor screening | Acquisition, processing, mask, cloud, resolution, coverage | Tree height, clearance, dangerous branch, or field condition | Exact feature manifest and freshness state | | GIS / terrain | Common analysis frame, overlap, distance, terrain context | Authority, CRS, geometry version, resolution, transform | Legal service territory or exact failure location | Approved spatial contract | | Infrastructure | Authorized physical exposure or generalized consequence context | Source authority, date, precision, access class | Causation or increased likelihood from proximity | Feature authorization or public-safe context | | History | Pre-cutoff recurrence and vulnerability context | Coverage, lookback, cutoff, aggregation, provider version | Causation or unchanged present conditions | Time-bounded feature contract | | Operational | Private saved-location, crew-readiness, or account context | Tenant, access, timestamp, purpose, audit identity | Autonomous dispatch or public evidence | Separately authorized feature | | Seasonal / climate | Background planning assumptions | Product, release, period, geography, uncertainty | A local outage forecast | Only when exact manifest proves use | **Caption:** Source families contribute different evidence and cannot prove the same things. Figure sources: [Environment and Climate Change Canada — Weather alerts](https://www.canada.ca/en/services/environment/weather/severeweather/weather-alerts.html), [Environment and Climate Change Canada — Meteorological Service of Canada free data service](https://www.canada.ca/en/environment-climate-change/services/weather-general-tools-resources/weather-tools-specialized-data/free-service.html), [United States Geological Survey — Landsat Normalized Difference Vegetation Index](https://www.usgs.gov/landsat-missions/landsat-normalized-difference-vegetation-index), [PostGIS Project — PostGIS 3.6 spatial data management](https://postgis.net/docs/manual-3.6/en/using_postgis_dbmanagement.html), [Public Safety Canada — Risk Management Guide for Critical Infrastructure Sectors](https://www.publicsafety.gc.ca/cnt/rsrcs/pblctns/rsk-mngmnt-gd/index-en.aspx), [GeoGridIQ — Data Sources documentation](/docs/data-sources/). ## 5. Provenance, freshness, and data-quality states Every signal retains provider and product, version or raw identity, checksum where applicable, issue, valid, observed, available, fetched, and processed times, transformation version, spatial reference and resolution, coverage, quality, completeness, and freshness. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/), [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1). Fresh is within contract; delayed arrived late but may remain usable; stale is too old; missing is absent; failed did not process; unsupported lacks an approved source; partial lacks required evidence; and valid zero is a successful current observation of no qualifying records. The interface labels each state and shows zero only with its establishing source and time. Completeness and freshness remain separate: complete data can be stale, and fresh data can be partial. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 6. Spatial normalization and GIS methodology GeoGridIQ aligns weather grids, outage points or areas, vegetation rasters, terrain, and infrastructure context in a declared coordinate system using geometry-validity checks, versioned reprojection, containment or intersection, bounded distance, raster zonal summaries, and land-water handling. QA covers duplicate or overlapping geometry, cell area, coastal and partial cells, source resolution, and weather-to-analysis-grid mapping. Buffers must match source precision; a centroid is not a failure location. Sources: [Open Geospatial Consortium — Simple Feature Access standard](https://www.ogc.org/standards/sfa/), [International Association of Oil & Gas Producers — EPSG OGC URI definitions](https://epsg.org/ogc_uri.html), [PostGIS Project — PostGIS 3.6 spatial data management](https://postgis.net/docs/manual-3.6/en/using_postgis_dbmanagement.html), [PostGIS Project — ST_IsValid](https://postgis.net/docs/manual-3.6/en/ST_IsValid.html), [PostGIS Project — ST_Transform](https://postgis.net/docs/manual-3.6/en/ST_Transform.html), [PostGIS Project — PostGIS 3.6 spatial queries](https://postgis.net/docs/manual-3.6/en/using_postgis_query.html), [MapLibre — MapLibre GL JS documentation](https://maplibre.org/maplibre-gl-js/docs/). Numerical checks and map review are complementary: correct SQL can encode the wrong geography, while a plausible map can show the wrong concept. A verified GeoSQL pilot supports upstream investigation and map-in-the-loop review, not outage modeling, legal territory truth, or direct production writing. Sources: [GeoGridIQ — GeoSQL pilot status and governance audit](/blog/how-geogridiq-uses-geosql-map-in-the-loop/), [PostGIS Project — PostGIS 3.6 spatial queries](https://postgis.net/docs/manual-3.6/en/using_postgis_query.html). Related: [GeoSQL map-in-the-loop article](/blog/how-geogridiq-uses-geosql-map-in-the-loop/) ### Figure — Different geometries, one bounded analysis cell **Evidence class:** Schematic, not surveyed **Purpose:** Explain spatial alignment without implying asset-level precision. **Long description:** A coarse forecast-weather tile, generalized outage point or area, vegetation raster pixels, and generalized infrastructure context overlap a bounded GeoGridIQ analysis cell. The sequence names CRS transformation, geometry checks, spatial join, bounded distance, and zonal summary. The cell centroid is representative and not a failure location. - **Coarse grid — Weather tile**: Product grid and valid time mapped with bounded source distance - **Observed geometry — Outage evidence**: Generalized point or area with provider precision retained - **Raster — Vegetation pixels**: Dated raster values summarized under mask and coverage rules - **Context — Infrastructure context**: Generalized line or exposure context; protected detail withheld - **Output unit — Analysis cell**: Approved bounded unit; centroid is representative, not the failure site **Caption:** Different geometries are transformed and summarized into one bounded analysis cell. Figure sources: [Open Geospatial Consortium — Simple Feature Access standard](https://www.ogc.org/standards/sfa/), [International Association of Oil & Gas Producers — EPSG OGC URI definitions](https://epsg.org/ogc_uri.html), [PostGIS Project — ST_IsValid](https://postgis.net/docs/manual-3.6/en/ST_IsValid.html), [PostGIS Project — ST_Transform](https://postgis.net/docs/manual-3.6/en/ST_Transform.html), [PostGIS Project — PostGIS 3.6 spatial queries](https://postgis.net/docs/manual-3.6/en/using_postgis_query.html). ## 7. Temporal alignment and the no-future-information rule GeoGridIQ distinguishes source issue, valid, observation, available, retrieval, and processing times from forecast issue, valid-from, valid-to, and outcome times. Only evidence genuinely available by the issue-time cutoff may enter a forecast; later creations, revisions, and outcomes belong to validation or a later run. Sources: [Kapoor and Narayanan — Leakage and reproducibility failures in machine-learning science](https://doi.org/10.1016/j.patter.2023.100804), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Leakage includes using final observations as forecasts, valid-window outages in history, revisions without original availability, preprocessing fit on holdouts, random splits that mix related events or later periods, unarchived reconstructions, labels in features, and caches without source or feature time. A reproducible snapshot binds one issue time, region/horizon contract, bounded values and source versions, missingness/freshness flags, and source, feature, and grid fingerprints. Sources: [Kapoor and Narayanan — Leakage and reproducibility failures in machine-learning science](https://doi.org/10.1016/j.patter.2023.100804), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### Figure — Issue time separates evidence from outcomes **Evidence class:** Conceptual illustration **Purpose:** Make the no-future-information rule memorable and auditable. **Long description:** Before the cutoff, approved forecasts, prior observations, and pre-cutoff history may be eligible. At issue time, one feature snapshot is frozen. During and after the valid window, outage outcomes, final observations, restorations, and revisions are held back for later evaluation. Moving any of those later facts backward is prohibited leakage. - **Before — Allowed before cutoff**: Evidence whose defensible available_at is no later than issue time - **Cutoff — Frozen forecast**: One issue time, region/horizon contract, values, flags, and hashes - **After — Held-back valid window**: Qualifying outages and evolving conditions become potential outcomes - **Validation — Later revisions**: Restorations, final observations, corrections, and evaluation stay outside the snapshot **Caption:** Only evidence genuinely available by issue time belongs in the frozen forecast. Figure sources: [Kapoor and Narayanan — Leakage and reproducibility failures in machine-learning science](https://doi.org/10.1016/j.patter.2023.100804), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 8. Feature engineering Feature engineering turns raw evidence into a bounded, typed snapshot through verified window summaries, maxima, means, totals, change rates, exceedance durations, lagged history, spatial aggregation, documented interactions, encoding, or scaling. Missingness and freshness stay explicit. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). The feature contract fixes order, name, type, units, required, optional, or experimental status, transformation version, and exact region/horizon compatibility; training and inference must match it. Examples such as forecast gust, antecedent precipitation, dated vegetation, pre-cutoff outage history, terrain or exposure, and source-health flags must be labelled active, possible, conceptual, or planned. Schema presence alone does not prove current use. Sources: [Environment and Climate Change Canada — GDPS documentation](https://eccc-msc.github.io/open-data/msc-data/nwp_gdps/readme_gdps-datamart_en/), [United States Geological Survey — Landsat Normalized Difference Vegetation Index](https://www.usgs.gov/landsat-missions/landsat-normalized-difference-vegetation-index), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### Figure — Timestamped evidence becomes a feature snapshot **Evidence class:** Conceptual illustration **Purpose:** Show how values, units, missingness, freshness, and identity travel into typed inputs. **Long description:** Sanitized weather records, pre-cutoff outage history, dated vegetation screening, and static terrain context pass through bounded time-window and spatial summaries. The result is an ordered feature contract with typed values, missing and freshness flags, source versions, and feature and grid fingerprints. Example families are not represented as universally active. - **Inputs — Raw evidence**: Values retain provider, product, issue/valid/available times, unit, and geometry - **Rules — Bounded transforms**: Window summaries, spatial aggregation, unit conversion, and cutoff checks - **Contract — Typed fields**: Fixed names, order, types, units, and required/optional status - **Flags — Quality context**: Missingness, freshness, coverage, and source-role flags stay explicit - **Snapshot — Frozen identity**: Region, horizon, issue time, source, feature, and grid fingerprints **Caption:** A feature snapshot binds values, missingness, freshness, units, schema, and time. Figure sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/), [Environment and Climate Change Canada — GDPS documentation](https://eccc-msc.github.io/open-data/msc-data/nwp_gdps/readme_gdps-datamart_en/), [United States Geological Survey — Landsat Normalized Difference Vegetation Index](https://www.usgs.gov/landsat-missions/landsat-normalized-difference-vegetation-index). ## 9. Labels and observed outcome truth The label contract defines qualifying positives, authoritative providers, planned-outage handling, deduplication and revisions, spatial and valid-window matching, event grouping, any verified customer threshold, approximate geometry, negatives, missing coverage, censoring, provider transitions, versioning, and reproducible rebuilds. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/), [Hydro-Québec — Power outage FAQ](https://pannes.hydroquebec.com/poweroutages/understand-and-prevent/faq.html). Unknown is not negative. An unmatched forecast is a false positive or true negative only when the relevant feed adequately observed that cell and window; post-event information never enters its original snapshot. Separately authorized probability and ranking tracks retain their different objectives and label semantics, with raw identities, checksums, and matching versions preserved. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 10. Model training architecture GeoGridIQ separates local Docker training from production inference. Candidates undergo chronological or event-grouped holdout tests, imbalance and baseline diagnostics, and human-reviewed promotion. The exact artifact, metadata, manifest, checksum, size, and region/horizon identity enter the authorized model store; Railway loads only that identity. Production training, automatic promotion, and cross-scope substitution are not authorized. XGBoost applies to designated regional runs only; its name establishes neither trust nor probability semantics. Fallback scoring is separately governed and is not currently public-serving. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/), [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1), [Saito and Rehmsmeier — Precision-recall is more informative than ROC for imbalanced data](https://doi.org/10.1371/journal.pone.0118432), [Kapoor and Narayanan — Leakage and reproducibility failures in machine-learning science](https://doi.org/10.1016/j.patter.2023.100804), [Taylor et al. — Machine learning evaluation of storm-related transmission outage factors and risk](https://doi.org/10.1016/j.segan.2023.101016), [Taylor et al. — Community power outage prediction modeling for the Eastern United States](https://doi.org/10.1016/j.egyr.2023.10.073), [XGBoost Project — XGBoost documentation](https://xgboost.readthedocs.io/en/stable/). ### Figure — Local training is separate from production inference **Evidence class:** Current verified product state **Purpose:** Show environment separation, human promotion, and exact artifact serving. **Long description:** Local Docker training produces a candidate run and evidence package. Temporal holdout review and human promotion lead to an approved artifact, metadata, manifest, checksum, size, and production lock in the authorized model store. The Railway generation worker loads that exact identity for inference, creates a provenance-bound batch, and persists predictions. Production training and automatic promotion are outside the current verified architecture. - **Development — Local Docker training**: Region/horizon dataset, candidate run, configuration, and evidence package - **Evidence — Held-back evaluation**: Temporal/event separation, imbalance, calibration, segment and baseline review - **Governance — Human promotion**: Explicit approval; no automatic production promotion - **Handoff — Approved model store**: Exact artifact, metadata, manifest, checksum, size, and production lock - **Inference — Railway generation**: Load exact approved identity; do not train or substitute - **Output — Persisted batch**: Complete intended rows with model, source, feature, grid, and validity provenance **Caption:** Controlled local training and human promotion are separate from production inference. Figure sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/), [XGBoost Project — XGBoost documentation](https://xgboost.readthedocs.io/en/stable/). ## 11. Model trust, promotion, integrity, and rollback A candidate becomes eligible only after applicable data, label, feature, grid, leakage, holdout, calibration, segment, serialization, integrity, exact-scope, promotion, runtime-load, current-batch, and validity gates pass. Public states use canonical reason codes. Sources: [GeoGridIQ — Model trust, artifact, batch, and validation contract](/docs/model-validation/), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/), [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1). Trusted and promoted describe one artifact, not a current forecast. Missing integrity, identity mismatch, failed validation, absent promotion, load failure, incompatible features, no batch, or expiry prevents that claim. Rollback retains auditable rows but never substitutes another scope, stale output, or unlabelled fallback. Restore an exact approved identity or show unavailable. Sources: [GeoGridIQ — Model trust, artifact, batch, and validation contract](/docs/model-validation/), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### Figure — Trust continues after promotion **Evidence class:** Current verified product state **Purpose:** Explain why artifact presence alone is insufficient and where fail-closed outcomes arise. **Long description:** The gate sequence is data readiness, leakage audit, training, temporal holdout and calibration, artifact integrity, human promotion, runtime load, source and feature checks, current batch, validity, and monitoring. A failed gate leads to hidden audit evidence or an unavailable public state. Fallback can become public only under a separately implemented and verified public-serving contract; the path never borrows another region or horizon. - **1 — Data and labels**: Readiness, coverage, label quality, and exact spatial/temporal scope - **2 — Leakage audit**: Only issue-time-available evidence; event-grouped temporal evaluation - **3 — Training and holdout**: Candidate metrics, calibration, segments, baselines, and limitations - **4 — Artifact integrity**: Exact family, region, horizon, manifest, metadata, hash, size, and schema - **5 — Human promotion**: Reviewed eligibility for one production scope - **6 — Runtime and sources**: Exact load plus current feature and critical-source checks - **7 — Current batch**: Complete population inside its declared validity window - **8 — Safe outcome**: Trusted current, hidden/audit-only, or unavailable; fallback needs a separate public contract **Caption:** Every required gate must pass before a forecast can be called trusted and current. Figure sources: [GeoGridIQ — Model trust, artifact, batch, and validation contract](/docs/model-validation/), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/), [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1). ## 12. Production inference and prediction generation Generation resolves region, horizon, sources, feature snapshot, freshness, and the exact trusted and promoted artifact; verifies identity, schema, manifest, checksum, and size; scores the complete population; attaches provenance, explanation, and confidence; persists rows; then applies display gates. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/), [GeoGridIQ — Model trust, artifact, batch, and validation contract](/docs/model-validation/). Artifact, load, batch, row, current validity, display, saved-location, and public visibility are distinct states. Hidden rows may persist for audit. Caches change with region, horizon, model, source, feature, grid, or validity; marker counts are not expected outages or total persisted rows. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 13. Anatomy of a GeoGridIQ risk output Read an output as a validity- and provenance-bound bundle. Geography names the scored area, not an asset; horizon and valid times name its window; generated and source times describe batch age and evidence freshness. Mode distinguishes trusted probability, separately authorized ranking, fallback scoring, and unavailable semantics. Fallback scoring can exist under its own contract; public capability remains unavailable without a separately verified public-serving fallback contract. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Probability, authorized rank, risk level, severity, confidence, consequence, drivers, freshness, identity, missingness, and display eligibility differ. Confidence is not an interval, drivers are not causes, freshness is not completeness, and provenance is not current validity. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### Figure — Anatomy of one risk output **Evidence class:** Conceptual illustration **Purpose:** Teach field-by-field interpretation and the boundary of each field. **Long description:** A sanitized output lists geography, horizon, generated and valid times, prediction mode, probability or a contract-defined relative score, risk level, severity, confidence or evidence quality, consequence, associated drivers, source freshness, batch and model identity, display eligibility, and missing or fallback flags. Relative and fallback modes are conceptual unless a separate public-serving contract is authorized. Each row also states what the field does not prove. | Field | What it answers | What it does not prove | | --- | --- | --- | | Region / grid cell | Which bounded area is scored | An exact failed asset, building, address, or customer | | Horizon and valid window | Which forecast contract and future interval apply | Permanent risk or availability at other horizons | | Generated time | When this batch was produced | That every source was fresh | | Prediction mode | Verified public mode or unavailable; other contracts may define relative or fallback semantics | Public authorization or equivalent semantics across modes | | Probability | Likelihood of the declared event when calibrated | Severity, certainty, duration, or customer count | | Relative risk | Ordering among an eligible comparison population | A calibrated chance of outage | | Risk level / severity | Configured category or condition magnitude | A universal standard or probability | | Confidence | Implemented evidence and path quality | A confidence interval or second probability | | Consequence | Potential importance of affected context | Higher physical likelihood by itself | | Top drivers | Inputs associated with this score | Proven physical cause | | Freshness / missing flags | Age, validity, and known evidence gaps | Completeness when only freshness is shown | | Model / batch identity | Which governed path produced the row | Current validity without passing times and gates | | Display eligible | Whether the row may appear in a current UI | Whether it is persisted for audit | **Caption:** A risk result is a provenance- and validity-bound bundle, not one number. Figure sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 14. Probability, confidence, severity, consequence, and risk level Probability estimates an event only when track and calibration support it. Relative risk ranks only on a separately authorized track. Confidence describes evidence quality, not probability; severity is magnitude, consequence is impact, and risk level is a display rule. Sources: [scikit-learn Project — Probability calibration](https://scikit-learn.org/stable/modules/calibration.html), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/), [Public Safety Canada — Risk Management Guide for Critical Infrastructure Sectors](https://www.publicsafety.gc.ca/cnt/rsrcs/pblctns/rsk-mngmnt-gd/index-en.aspx). High probability with low confidence requires caution; moderate probability with high consequence may raise review priority; authorized rank means higher than peers, not a probability; and official alerts remain important. Fallback scoring is not currently public-serving and never inherits model-probability semantics. Authoritative bands never trigger dispatch by themselves. Sources: [Environment and Climate Change Canada — Weather alerts](https://www.canada.ca/en/services/environment/weather/severeweather/weather-alerts.html), [Public Safety Canada — Risk Management Guide for Critical Infrastructure Sectors](https://www.publicsafety.gc.ca/cnt/rsrcs/pblctns/rsk-mngmnt-gd/index-en.aspx), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### Relative Outage Risk — Beta Quebec's six-hour beta is a positive-unlabeled relative-risk ranking. It evaluates 275 authoritative Quebec cells, displays 250 higher-ranked cells, and retains 25 lower-ranked cells for audit; hidden does not mean no risk. It identifies neither an exact asset or customer nor a confirmed outage or calibrated outage probability. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Candidate version 1 (public identity prefix 69bfc11c4416) learns ordering from verified historical outage-positive event groups and unlabeled comparison areas; unlabeled does not mean outage-free. Event grouping, temporal separation, and leave-positive-cell-out evaluation reduce leakage without inventing verified negatives. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). A cell in the 90th percentile ranks higher than 90% of the Quebec cells evaluated in the current batch. It does not mean there is a 90% probability that an outage will occur. The governed bands are Highest at the 90th–100th percentile, Elevated at the 75th–below-90th percentile, Moderate at the 50th–below-75th percentile, and Lower below the 50th percentile. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). The feature contract uses 18 pinned 2020 NRCan land-cover composition fields and four evaluation-time Toronto calendar encodings. Source-safe timing excludes outcome events and later revisions. Live outages, weather, alerts, EONET, saved locations, addresses, and assets are not candidate features, so those live dashboard states do not govern batch availability. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Likely drivers are model-attributed associations, not confirmed causes. Evidence confidence assesses source and contract quality, not probability: HIGH requires CURRENT freshness with no sanctioned missing rows; MODERATE permits sanctioned optional missingness through 15%; LOW permits through 25% or a reviewed last-known-good batch. Unsanctioned missingness, staleness, an incomplete grid, artifact or approval mismatch, critical failure, or fallback makes the product UNAVAILABLE. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Approved event-grouped evaluation reports MRR 0.906, NDCG@10 0.927, and Recall@10 0.991 across 108 groups. Leave-positive-cell-out comparison reports 0.795, 0.797, and 0.844 respectively across 45 groups and 9 held-out cells. These metrics describe early rank, top-ten ordering, and top-ten retrieval—not calibration, Brier score, or outage frequency. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Generation runs every six hours on grid scope ca-qc-authoritative-275-v1 behind exact candidate, artifact, and atomic-pointer gates. CURRENT displays; reviewed last-known-good lasts five minutes; STALE and UNAVAILABLE return no cells. Calibrated probability remains unavailable, saved-location risk remains disabled, and ranking cannot dispatch crews. A future probability product needs its own target, labels, calibration, validation, approval, and release. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### Figure — Probability, confidence, severity, and consequence **Evidence class:** Conceptual illustration **Purpose:** Prevent four commonly conflated concepts from collapsing into one risk word. **Long description:** A semantic comparison defines probability as event likelihood, confidence as implemented evidence or path quality, severity as condition magnitude, and consequence as potential impact. Scenario rows explain high probability with low confidence, moderate probability with high consequence, high relative risk without probability semantics, and low probability beside an official alert. | Concept or scenario | Question answered | Responsible interpretation | Prohibited inference | | --- | --- | --- | --- | | Probability | How likely is the declared event for this area/window? | Use only for a track with matching target and calibration | Confidence, severity, duration, or certainty | | Confidence | How well supported is this path under its implemented contract? | Review freshness, completeness, model state, and timing | A statistical interval or chance the forecast is correct | | Severity | How extreme are the relevant conditions or classification? | Read alongside, not instead of, probability | That an outage will happen | | Consequence | How important could disruption be? | May raise review priority | That physical failure is more likely | | High probability + low confidence | Likelihood estimate with weak evidence support | Review gaps and mode cautiously | Strong certainty | | Moderate probability + high consequence | Moderate likelihood with important impact context | Consider higher review priority | Changed physical probability | | High relative rank | Higher than peers in the eligible population | Use for comparison only | An absolute percentage | | Low probability + severe alert | Different outage and hazard questions | Continue to follow the official warning | That one signal cancels the other | **Caption:** Likelihood, evidence quality, condition magnitude, and impact answer different questions. Figure sources: [scikit-learn Project — Probability calibration](https://scikit-learn.org/stable/modules/calibration.html), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/), [Public Safety Canada — Risk Management Guide for Critical Infrastructure Sectors](https://www.publicsafety.gc.ca/cnt/rsrcs/pblctns/rsk-mngmnt-gd/index-en.aspx). ## 15. Explainability and top drivers An explanation answers why an area, window, and path received its score. Local explanations apply to one row; global explanations summarize an evaluated population. Name SHAP or numeric contributions only when method and denominator are verified. Rule-based fallback drivers are not learned importance, and those rows are not currently public-serving. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Explanations match region, horizon, cell, run, model, and batch; separate missing from zero; limit precision; and protect sensitive features. A driver is association, not cause; causation needs provider, field, or investigative evidence. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/), [Hydro-Québec — Power outage FAQ](https://pannes.hydroquebec.com/poweroutages/understand-and-prevent/faq.html). ### Figure — A driver explains association, not cause **Evidence class:** Conceptual illustration **Purpose:** Illustrate a local explanation without inventing contribution percentages or causal findings. **Long description:** One conceptual cell and window lists forecast wind, antecedent precipitation, and dated vegetation context as direction-only associated driver families, while an unavailable feature is visibly marked missing rather than shown as zero. A final boundary states that physical cause requires later provider or investigative evidence. No percentages are asserted. - **Available — Forecast wind**: Associated with a higher conceptual score for this exact window; direction only - **Available — Antecedent precipitation**: Associated context measured before the issue-time cutoff - **Available — Vegetation screening**: Dated generalized context; not tree height or clearance - **Missing — Optional lightning**: Missing means unavailable, not a measured zero - **Driver ≠ cause — Cause boundary**: Observed cause requires provider or investigative evidence after the event **Caption:** Associated inputs help explain a score; they do not prove outage cause. Figure sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 16. Confidence and uncertainty Confidence is separate from probability and identified by output track because GeoGridIQ has multiple evidence-quality contracts. Factors can include completeness, missing signals, freshness, model or calibration posture, timing, and boundary proximity. Fallback cannot receive the strongest model treatment and is not currently public-serving. A value such as 85 is neither correctness probability nor a confidence interval. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Uncertainty spans source coverage, spatial precision, timing, vegetation age, label sparsity or imbalance, calibration, transfer, drift, artifacts, batches, policy, and field conditions. Gates warn, hide but retain, govern scoring, mark unavailable, or require review. Sources: [Environment and Climate Change Canada — Meteorological Service of Canada free data service](https://www.canada.ca/en/environment-climate-change/services/weather-general-tools-resources/weather-tools-specialized-data/free-service.html), [United States Geological Survey — Landsat Normalized Difference Vegetation Index](https://www.usgs.gov/landsat-missions/landsat-normalized-difference-vegetation-index), [Saito and Rehmsmeier — Precision-recall is more informative than ROC for imbalanced data](https://doi.org/10.1371/journal.pone.0118432), [scikit-learn Project — Probability calibration](https://scikit-learn.org/stable/modules/calibration.html), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 17. Risk thresholds, display eligibility, and alerts Risk bands convert a verified score into a human-readable category. Their thresholds may vary by output mode, region, horizon, and decision use, so public values should come from one authoritative configuration when safe. A threshold does not change a relative rank into a probability, create a universal utility standard, or establish the right operational response for every user. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Display eligibility is a presentation decision layered on persistence. Low-confidence, untrusted, incomplete, invalid, or expired rows may remain stored for audit while being hidden from current views. Hidden is not deleted. A valid current run that explicitly reports no elevated result is different from missing, failed, or stale evidence; the latter states must be shown as unavailable or limited, never as zero. Display limits may also rank a complete population without making hidden rows statistically false. Sources: [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Alerts, where implemented, add separate policy, authorization, account, and delivery gates. A probability or risk band does not automatically dispatch crews, trigger emergency action, or replace official warnings and utility procedures. Expired results lose current display eligibility even if their historical rows remain useful for audit and validation. Sources: [Environment and Climate Change Canada — Weather alerts](https://www.canada.ca/en/services/environment/weather/severeweather/weather-alerts.html), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 18. Fallback, degraded, and fail-closed behaviour A public result requires an exact eligible model or scoring contract, compatible features, healthy required sources, a verified artifact, successful runtime load, and a valid current batch. Model absence, identity or checksum mismatch, incompatible features, stale or missing evidence, unsupported scope, validation or load failure, and missing or expired batches limit what may be claimed. Sources: [GeoGridIQ — Model trust, artifact, batch, and validation contract](/docs/model-validation/), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). An authorized deterministic fallback scores only under its own source contract, never inherits calibrated probability, and is public only with separate verified serving approval. Protected paths may stop and retain earlier rows for audit. With no safe path, the result is unavailable. Sources: [GeoGridIQ — Model trust, artifact, batch, and validation contract](/docs/model-validation/), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). GeoGridIQ never borrows another scope's model, serves expiry as current, or maps missing evidence to zero. Canonical reason codes distinguish current, valid no-elevated results, delayed or unavailable sources, validation, missing or expired batches, and unsupported scopes. A ranking appears only under its own verified public contract. Sources: [GeoGridIQ — Model trust, artifact, batch, and validation contract](/docs/model-validation/). ## 19. Forecast validation and accountability Validation stores a traceable forecast before outcomes, preserving issue and valid times, population, mode, and provenance. Later events match under declared spatial, temporal, label, and coverage rules; inadequate observation remains unknown. Sources: [World Meteorological Organization — Forecast verifications](https://community.wmo.int/site/knowledge-hub/programmes-and-initiatives/wmo-integrated-processing-and-prediction-system-wipps/forecast-verifications), [Kapoor and Narayanan — Leakage and reproducibility failures in machine-learning science](https://doi.org/10.1016/j.patter.2023.100804), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Related: [Model Validation documentation](/docs/model-validation/) Hits match flags; misses are uncovered qualifying outcomes; false positives and true negatives require observed absence. Precision measures matched flags, recall covered outcomes, F1 balances them, ROC-AUC ranks thresholds, PR-AUC emphasizes rare positives, and Brier score or reliability assesses probability calibration. Lead time, burden, and coverage also matter. Sources: [World Meteorological Organization — Forecast verifications](https://community.wmo.int/site/knowledge-hub/programmes-and-initiatives/wmo-integrated-processing-and-prediction-system-wipps/forecast-verifications), [European Centre for Medium-Range Weather Forecasts — Verification metrics guide](https://confluence.ecmwf.int/pages/viewpage.action?pageId=363864334), [European Centre for Medium-Range Weather Forecasts — Reliability diagram](https://charts.ecmwf.int/catalogue/packages/subseasonal/products/s2s-verification-ecmwf-rel), [Saito and Rehmsmeier — Precision-recall is more informative than ROC for imbalanced data](https://doi.org/10.1371/journal.pone.0118432), [scikit-learn Project — Probability calibration](https://scikit-learn.org/stable/modules/calibration.html). Every metric names region, horizon, run, data period, sample and event counts, definition, computation time, promotion state, segments, and limitations. Without a current public-safe package, show conceptual diagrams, not performance numbers. Sources: [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1), [World Meteorological Organization — Forecast verifications](https://community.wmo.int/site/knowledge-hub/programmes-and-initiatives/wmo-integrated-processing-and-prediction-system-wipps/forecast-verifications), [GeoGridIQ — Model trust, artifact, batch, and validation contract](/docs/model-validation/). ### Figure — Validation asks several different questions **Evidence class:** Conceptual illustration **Purpose:** Join outcome matching, error types, calibration, lead time, and segment review without a performance claim. **Long description:** Four conceptual panels are represented as table rows. The first distinguishes hits, false alarms, misses, correct negatives, and unknown observation coverage. The second describes probability reliability without numeric current-model results. The third places issue time, lead time, valid window, and outcome in order. The fourth requires region, horizon, season, hazard, and provider segments. | Panel | Question | Required interpretation | Why one number is insufficient | | --- | --- | --- | --- | | Outcome matrix | Which flags and outcomes matched? | TP, FP, FN, and TN require sufficient provider coverage; otherwise unknown | Error types carry different operational costs | | Reliability / calibration | Do probability bins match observed frequencies? | Bind every diagram to an exact model, dataset, period, and counts | Discrimination does not guarantee calibrated probabilities | | Lead-time timeline | How early was useful evidence issued? | Preserve issue, valid-window, and outcome times | Accuracy without useful lead time may not help operations | | Segment evaluation | Where does performance differ? | Slice by region, horizon, season, hazard, provider, and coverage | Aggregate metrics can hide weak segments and drift | **Caption:** No single validation metric proves operational value. Figure sources: [World Meteorological Organization — Forecast verifications](https://community.wmo.int/site/knowledge-hub/programmes-and-initiatives/wmo-integrated-processing-and-prediction-system-wipps/forecast-verifications), [European Centre for Medium-Range Weather Forecasts — Reliability diagram](https://charts.ecmwf.int/catalogue/packages/subseasonal/products/s2s-verification-ecmwf-rel), [European Centre for Medium-Range Weather Forecasts — Verification metrics guide](https://confluence.ecmwf.int/pages/viewpage.action?pageId=363864334), [Saito and Rehmsmeier — Precision-recall is more informative than ROC for imbalanced data](https://doi.org/10.1371/journal.pone.0118432), [scikit-learn Project — Probability calibration](https://scikit-learn.org/stable/modules/calibration.html), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 20. Human-reviewed operational use When current and trustworthy, GeoGridIQ can support monitoring, storm briefings, area comparison, weather and vegetation review, preparedness, separately verified crew-readiness analysis, explanation, and post-event validation. It supplies reviewable evidence; it does not decide operations. Sources: [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). A responsible user checks source time and missingness, confirms the valid window and mode, separates confidence, probability, drivers, and consequence, and compares official weather and utility information. Procedures, safety rules, field evidence, access, and local authority determine action; recorded decisions and outcomes support evaluation. Sources: [Environment and Climate Change Canada — Weather alerts](https://www.canada.ca/en/services/environment/weather/severeweather/weather-alerts.html), [Hydro-Québec — Power outage FAQ](https://pannes.hydroquebec.com/poweroutages/understand-and-prevent/faq.html), [BC Hydro — How power is restored](https://www.bchydro.com/safety-outages/power-outages/during-an-outage/how-power-is-restored.html), [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1). GeoGridIQ is not an official outage source, alerting system, substitute for ECCC warnings, SCADA, OMS, utility GIS, field inspection, or emergency procedure. It is not autonomous dispatch, an outage guarantee, or an exact-asset detector without a separately validated target. Sources: [Environment and Climate Change Canada — Weather alerts](https://www.canada.ca/en/services/environment/weather/severeweather/weather-alerts.html), [Hydro-Québec — Power outage FAQ](https://pannes.hydroquebec.com/poweroutages/understand-and-prevent/faq.html), [BC Hydro — How power is restored](https://www.bchydro.com/safety-outages/power-outages/during-an-outage/how-power-is-restored.html). ### Figure — Human review connects intelligence to accountable learning **Evidence class:** Conceptual illustration **Purpose:** Place human authority before action and evaluation after the event. **Long description:** Current intelligence flows to a reviewer who checks validity, sources, mode, confidence, drivers, and consequence, compares official information, and applies utility and safety procedures. An authorized action or no-action decision is recorded. Later observed outcomes are matched, and validation informs monitoring and future human-reviewed model work. No autonomous dispatch occurs. - **1 — Current intelligence**: Read the exact region, horizon, mode, sources, valid window, and limitations - **2 — Operator review**: Compare official information; separate probability, confidence, drivers, and consequence - **3 — Authorized decision**: Apply utility procedure, safety authority, access, and field expertise - **4 — Observed outcome**: Record later qualifying events, coverage, action, and context - **5 — Accountability**: Validate, monitor drift, document decisions, and propose future reviewed work **Caption:** GeoGridIQ supports review; authorized people and procedures determine action. Figure sources: [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1), [Hydro-Québec — Power outage FAQ](https://pannes.hydroquebec.com/poweroutages/understand-and-prevent/faq.html), [BC Hydro — How power is restored](https://www.bchydro.com/safety-outages/power-outages/during-an-outage/how-power-is-restored.html), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ## 21. Current regional, provider, and horizon availability Availability is runtime state. The server table reads capability, source-freshness, artifact-safety, registry, feature, and current-prediction records for each region/horizon, returning source, model, batch, availability, reason, and UTC time. Unreadable or inconsistent state fails closed. Sources: [GeoGridIQ — Current regional coverage and capability status](/coverage/), [GeoGridIQ — Model trust, artifact, batch, and validation contract](/docs/model-validation/). Related: [Full coverage and status](/coverage/) Rows remain independent: a trusted artifact without a present, unexpired batch is not a current forecast; unavailable outage data and missing vegetation are not zero; one scope cannot establish another. Artifact-only, validation, expiry, and unsupported remain distinct. Fallback scoring can exist under its own contract; public capability remains unavailable without a separately verified public-serving fallback contract. The table omits private identities and never treats configuration as proof of a current batch. Sources: [GeoGridIQ — Current regional coverage and capability status](/coverage/), [GeoGridIQ — Model trust, artifact, batch, and validation contract](/docs/model-validation/). ### Figure — Current regional and horizon capability **Evidence class:** Current verified product state **Purpose:** Expose current source, model, batch, and public availability without a fabricated risk map. **Long description:** A server-rendered table lists canonical region and horizon with outage, weather, vegetation, model, current-batch, public availability, reason, validity, and exact as-of fields. Artifact-only, expired, unsupported, fallback, missing, suspended, and unavailable states remain distinct. If authoritative state cannot be verified, affected rows fail closed to unavailable. **Capability as of:** 2026-09-08T04:50:43.243576Z Availability is evaluated independently for each exact region and horizon. Unavailable means not verified for current use; it does not mean zero risk. | Region | Horizon | Outage | Weather | Vegetation | Model | Batch | Availability | Reason | Valid window | As of | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | Canada East (Quebec) | 6H | Delayed / warning (as of 2026-09-08T04:30:17+00:00) | Unavailable (as of 2026-09-08T04:49:08.550881+00:00) | Unavailable (as of 2026-09-08T04:49:08.754955+00:00) | Suspended | Not verified | Unavailable | model_retraining_required: The protected model path is suspended while leakage-safe evidence and retraining are reviewed. | Not available | 2026-09-08T04:50:43.243576Z | | Canada East (Quebec) | 24H | Delayed / warning (as of 2026-09-08T04:30:17+00:00) | Unavailable (as of 2026-09-08T04:49:08.550881+00:00) | Unavailable (as of 2026-09-08T04:49:08.754955+00:00) | Suspended | Not verified | Unavailable | model_retraining_required: The protected model path is suspended while leakage-safe evidence and retraining are reviewed. | Not available | 2026-09-08T04:50:43.243576Z | | Canada East (Quebec) | 72H | Delayed / warning (as of 2026-09-08T04:30:17+00:00) | Unavailable (as of 2026-09-08T04:49:08.550881+00:00) | Unavailable (as of 2026-09-08T04:49:08.754955+00:00) | Suspended | Not verified | Unavailable | model_retraining_required: The protected model path is suspended while leakage-safe evidence and retraining are reviewed. | Not available | 2026-09-08T04:50:43.243576Z | | Canada East (Quebec) | 7D | Delayed / warning (as of 2026-09-08T04:30:17+00:00) | Unavailable (as of 2026-09-08T04:49:08.550881+00:00) | Unavailable (as of 2026-09-08T04:49:08.754955+00:00) | Suspended | Not verified | Unavailable | model_retraining_required: The protected model path is suspended while leakage-safe evidence and retraining are reviewed. | Not available | 2026-09-08T04:50:43.243576Z | | Canada West (British Columbia) | 6H | Unavailable (as of 2026-09-05T03:55:55.852599+00:00) | Unavailable (as of 2026-09-08T04:49:09.646790+00:00) | Unavailable (as of 2026-09-08T04:49:09.763300+00:00) | Not configured | Not verified | Unsupported | unsupported_horizon: This region and horizon do not have an approved public prediction track. | Not available | 2026-09-08T04:50:43.243576Z | | Canada West (British Columbia) | 24H | Unavailable (as of 2026-09-05T03:55:55.852599+00:00) | Unavailable (as of 2026-09-08T04:49:09.646790+00:00) | Unavailable (as of 2026-09-08T04:49:09.763300+00:00) | Unavailable | Missing | Unavailable | critical_source_stale: A required source or feature snapshot is stale, missing, failed, or incomplete. | Not available | 2026-09-08T04:50:43.243576Z | | Canada West (British Columbia) | 72H | Unavailable (as of 2026-09-05T03:55:55.852599+00:00) | Unavailable (as of 2026-09-08T04:49:09.646790+00:00) | Unavailable (as of 2026-09-08T04:49:09.763300+00:00) | Not configured | Not verified | Unsupported | unsupported_horizon: This region and horizon do not have an approved public prediction track. | Not available | 2026-09-08T04:50:43.243576Z | | Canada West (British Columbia) | 7D | Unavailable (as of 2026-09-05T03:55:55.852599+00:00) | Unavailable (as of 2026-09-08T04:49:09.646790+00:00) | Unavailable (as of 2026-09-08T04:49:09.763300+00:00) | Not configured | Not verified | Unsupported | unsupported_horizon: This region and horizon do not have an approved public prediction track. | Not available | 2026-09-08T04:50:43.243576Z | **Caption:** Availability is evaluated independently for every exact region and horizon at the displayed UTC audit time. Figure sources: [GeoGridIQ — Current regional coverage and capability status](/coverage/), [GeoGridIQ — Model trust, artifact, batch, and validation contract](/docs/model-validation/). ## 22. Limitations and responsible interpretation Data limitations. Outage feeds can be incomplete, delayed, revised, generalized, or unavailable. Weather forecasts carry spatial and temporal uncertainty; vegetation can be old, cloud-affected, coarse, or distant from line clearance; infrastructure may be generalized or withheld. Dated runtime status governs current use. Sources: [Environment and Climate Change Canada — Meteorological Service of Canada free data service](https://www.canada.ca/en/environment-climate-change/services/weather-general-tools-resources/weather-tools-specialized-data/free-service.html), [Hydro-Québec — Power outage FAQ](https://pannes.hydroquebec.com/poweroutages/understand-and-prevent/faq.html), [United States Geological Survey — Landsat Normalized Difference Vegetation Index](https://www.usgs.gov/landsat-missions/landsat-normalized-difference-vegetation-index), [Public Safety Canada — Current critical-infrastructure overview](https://www.publicsafety.gc.ca/cnt/ntnl-scrt/crtcl-nfrstrctr/index-en.aspx), [GeoGridIQ — Limitations documentation](/docs/limitations/). Label and statistical limitations. Outages are rare, negative labels can be uncertain, severe events may be underrepresented, and history reflects older infrastructure and practice. Metrics depend on target, horizon, geography, threshold, matching rule, and evaluable population; aggregates can hide weak segments. Sources: [Saito and Rehmsmeier — Precision-recall is more informative than ROC for imbalanced data](https://doi.org/10.1371/journal.pone.0118432), [Taylor et al. — Machine learning evaluation of storm-related transmission outage factors and risk](https://doi.org/10.1016/j.segan.2023.101016), [Taylor et al. — Community power outage prediction modeling for the Eastern United States](https://doi.org/10.1016/j.egyr.2023.10.073), [Fatima et al. — Machine learning for power outage prediction during hurricanes — an extensive review](https://doi.org/10.1016/j.engappai.2024.108056). Spatial limitations. Regional and grid outputs do not identify exact failures. Precision varies, proximity does not prove interaction, and boundaries need authoritative support. Maps can overstate accuracy; valid geometry proves structure, not source authority, joins, buffers, or interpretation. Sources: [PostGIS Project — ST_IsValid](https://postgis.net/docs/manual-3.6/en/ST_IsValid.html), [PostGIS Project — PostGIS 3.6 spatial queries](https://postgis.net/docs/manual-3.6/en/using_postgis_query.html), [MapLibre — MapLibre GL JS documentation](https://maplibre.org/maplibre-gl-js/docs/). Model limitations. Associations are not causes. Drift, provider or infrastructure change, and degraded calibration can reduce performance; explanations can be incomplete. Models may not transfer across scope, and a loadable artifact can remain operationally unsuitable or lack a current batch. Sources: [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1), [Kapoor and Narayanan — Leakage and reproducibility failures in machine-learning science](https://doi.org/10.1016/j.patter.2023.100804), [scikit-learn Project — Probability calibration](https://scikit-learn.org/stable/modules/calibration.html). Operational limitations. False positives and misses occur. Risk intelligence does not set restoration priority, dispatch crews, or determine safe access; procedures and field evidence remain authoritative. GeoGridIQ cannot guarantee prevention, savings, or duration reduction. Verified, timestamped state reports current scope and batch limits. Sources: [BC Hydro — How power is restored](https://www.bchydro.com/safety-outages/power-outages/during-an-outage/how-power-is-restored.html), [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1), [GeoGridIQ — Current regional coverage and capability status](/coverage/). Related: [Limitations documentation](/docs/limitations/) ## 23. Security, privacy, licensing, and governance Public-source evidence is separated from account and utility data. Documentation covers least privilege, attribution and licensing, generalized sensitive infrastructure, public-safe evidence, audit trails, and human release approval. Incident and rollback ownership is public at an appropriate level; service and account boundaries govern access. Sources: [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1), [Public Safety Canada — Risk Management Guide for Critical Infrastructure Sectors](https://www.publicsafety.gc.ca/cnt/rsrcs/pblctns/rsk-mngmnt-gd/index-en.aspx), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). The page excludes secrets, protected endpoints, attack-enabling schema, protected-asset coordinates, addresses, crew locations, account data, private keys, topology, and identifiable logs. Saved-location and tenant records stay access-controlled; public maps generalize sensitive context. Sources: [Public Safety Canada — Current critical-infrastructure overview](https://www.publicsafety.gc.ca/cnt/ntnl-scrt/crtcl-nfrstrctr/index-en.aspx), [Public Safety Canada — Risk Management Guide for Critical Infrastructure Sectors](https://www.publicsafety.gc.ca/cnt/rsrcs/pblctns/rsk-mngmnt-gd/index-en.aspx), [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1). Licensing and attribution are reviewed with technical fitness; viewable does not mean reusable. Security is ongoing risk management, not a zero-risk claim, and controls are described without increasing attack surface. Sources: [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1). ## 24. Reproducibility, versioning, and change control This methodology records version, publication, modification, fact-review and next-review dates, owner, and change history. Score-meaning changes receive dated versions; editorial, source, feature, label, model, threshold, and runtime-state changes remain distinguishable. Sources: [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). A reproducible result links source identities, spatial, grid, feature and label contracts, dataset fingerprint, training configuration, artifact checksum, evidence package, approval, batch, valid window, outcome matching, and validation. Reviewers can reconstruct what was known, transformed, approved, and shown. Sources: [National Institute of Standards and Technology — AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1), [Kapoor and Narayanan — Leakage and reproducibility failures in machine-learning science](https://doi.org/10.1016/j.patter.2023.100804), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). Runtime status has its own timestamp and may change without a methodology release. Versioned methodology explains meaning; dynamic state explains availability. HTML and machine-readable Markdown share the same structured content. Sources: [GeoGridIQ — Current regional coverage and capability status](/coverage/), [GeoGridIQ — Prediction implementation, contracts, and regression tests](/docs/model-validation/). ### Public methodology change history - **2.2.0 (2026-08-04)** — Expanded Quebec 6H beta interpretation, evidence confidence, ranking evaluation, cadence, and future probability boundaries. - **2.1.0 (2026-08-02)** — Added the approved Quebec 6H relative-risk beta contract and probability-separation boundaries. - **2.0.0 (2026-07-30)** — Full explainable methodology overhaul with dynamic current-state disclosure. - **1.0.0 (2026-06-07)** — Initial public methodology documentation. ## 25. Worked interpretation examples Every value in these examples is a Conceptual illustration. None is a current forecast, model-performance result, protected asset location, or historical reconstruction. The examples demonstrate interpretation boundaries and the different actions that evidence quality and consequence can support. ### Example A — Elevated probability with strong evidence At 12:00 UTC, illustrative cell A-17 receives a trusted 24-hour probability of 0.68 for the 12:00-to-12:00 valid window. Required sources pass their freshness contracts, evidence quality is high under the implemented confidence rules, and associated inputs include forecast gusts, recent antecedent precipitation, and dated vegetation-screening context. The right interpretation is review this cell and window alongside official information. It does not mean a particular asset will fail or that an outage is certain. The issue time, valid window, prediction mode, current source states, and conceptual nature of every value travel with the example. ### Example B — Elevated score with low confidence Illustrative cell B-04 has a separately authorized relative-risk rank, but the outage feed is delayed and a normally expected vegetation layer is missing. The interface labels the incomplete path and may hide it, retain it for audit, or require review. Fallback scoring may exist under its own contract, but public capability remains unavailable unless a separately implemented public-serving fallback contract is verified. A high score supported by weak evidence is not equivalent to a trusted high-confidence forecast, missing is not zero, and a rank is not an absolute percentage. ### Example C — Moderate physical likelihood with high consequence Illustrative cell C-22 has a calibrated probability of 0.34 and current evidence, while generalized public context indicates an important community function inside the cell. Consequence may raise review priority even though the model probability remains 0.34. The exposure does not retroactively make the physical outage more likely and does not reveal a protected facility's exact location. An operator could review official sources and preparedness procedures without claiming that the cell's critical context caused the forecast. ### Example D — Fallback, missing batch, and valid zero In one conceptual row, no exact trusted artifact is eligible but an authorized deterministic contract can still generate or score under its own rules. Public capability remains unavailable unless a separately implemented public-serving fallback contract is verified, and the signal never inherits machine-learning probability wording. In a second row, an artifact exists but the batch is expired, so status is unavailable. In a third, a valid current run finds no elevated result. These states must not share a generic zero; separate reason, mode, source, batch, and as-of fields expose the difference. ## 26. Frequently asked questions The visible FAQ answers common interpretation and accountability questions in the same language used by the main methodology. Each answer opens with a direct response and is emitted into FAQ schema from this one visible source; structured data never contains hidden, expanded, or differently worded answers. ### What does GeoGridIQ predict? GeoGridIQ estimates a declared outage-risk outcome for a stated geography and future valid window only under a verified public-serving contract. Calibrated probability appears only when that track supports it. The separate Quebec six-hour beta ranks 275 cells and shows 250; it does not estimate outage chance. Fallback scoring never implies public availability without its own verified serving contract. GeoGridIQ neither predicts weather nor reports an outage in progress. ### Does GeoGridIQ predict the exact line, pole, address, or customer that will lose power? No, GeoGridIQ does not claim exact line, pole, address, or customer failure unless a separately validated asset-level target is explicitly identified. Current regional and grid-cell outputs apply only to their named analysis unit. Map centroids and cell shading must not be read as surveyed failure locations. ### What data does GeoGridIQ use? GeoGridIQ uses only the source and feature families authorized for the exact region, horizon, and output track. Depending on the active manifest, those may include verified outage records, forecast weather and alerts, vegetation screening, geographic and terrain context, approved infrastructure exposure, and dated history. Every active source is bounded by provenance, timestamp, coverage, and quality rules. ### What is the difference between a weather forecast and an outage-risk forecast? A weather forecast estimates future atmospheric or hazard conditions, while an outage-risk forecast estimates a defined loss-of-service outcome using weather plus other eligible evidence and learned or deterministic rules. An ECCC alert remains authoritative hazard context. It is not automatically an outage label or proof that service will be interrupted. ### What is the difference between probability and confidence? Probability estimates the likelihood of the declared outage event, while confidence describes the implemented quality of the evidence and prediction path. Confidence may reflect freshness, completeness, model status, calibration context, spatial precision, and timing. It is not a statistical confidence interval or the chance that the probability is correct. ### What is relative outage risk, and is it a probability? Relative outage risk is a ranking or score that compares eligible areas, and it is not automatically a probability. It can show where evidence appears more concerning relative to peers without saying that an outage has an X-percent chance. Probability wording is reserved for a track whose target and calibration support it. ### What does a top driver mean? A top driver is an input or feature family most associated with a specific score under the implemented explanation method. It can help a reviewer understand why that row differs from others. It does not establish the physical cause of a future or observed outage, and a percentage is shown only when its exact contribution semantics are verified. ### Does critical-infrastructure exposure make an outage more likely? No, critical-infrastructure exposure does not by itself make an outage physically more likely. It describes potential consequence if disruption occurs and can raise review priority. Only a separately verified model feature and supported interpretation could affect probability, without disclosing protected coordinates or topology. ### How does GeoGridIQ prevent future-data leakage? GeoGridIQ prevents future-data leakage by freezing features at a declared issue-time cutoff and allowing only evidence genuinely available by then. Outcome events, later revisions, final observations, and holdout-period information remain outside the original snapshot. Temporal splits, availability timestamps, feature contracts, and snapshot identities make this rule auditable. ### How are outage predictions validated? GeoGridIQ validates stored forecasts against later qualifying observed outcomes using declared spatial, temporal, provider-coverage, and label rules. Evaluation considers discrimination, precision and recall, calibration where probabilities are claimed, lead time, alert burden, coverage, and segment-level behaviour. Published metrics must identify the exact region, horizon, run, dataset, period, counts, and limitations. ### What counts as a false positive or a missed outage? A false positive is a flagged unit without a matched qualifying outcome only when provider coverage is sufficient, while a missed outage is a qualifying observed event not covered by the declared forecast. If an outage feed is missing, partial, or failed, an unmatched row may remain unknown. Evaluation never silently converts unknown coverage to a negative label. ### What happens when a trusted model is unavailable? GeoGridIQ marks the public result unavailable unless a separately implemented public-serving fallback contract is verified. Fallback generation or scoring may exist under its own source and validity contract, but the current public live resolver does not treat those rows as a current public track. GeoGridIQ never substitutes another region or horizon's model, labels fallback as a trusted machine-learning probability, or serves an expired forecast as current. Public-safe reason codes explain the outcome. ### Why might a prediction be hidden? A prediction may be hidden because it fails current display gates for trust, validity, freshness, completeness, confidence, or another documented presentation rule. Hidden does not mean deleted or absent: a row can remain persisted for audit while not being suitable for current display. Display eligibility is separate from persistence. ### How often are predictions and source signals updated? Predictions and source signals update according to their own provider, ingestion, generation, and validity contracts, not one universal schedule. The interface shows exact source, generated, and as-of timestamps where safe. A recent fetch time does not prove that the provider observation itself is current. ### Why can an unavailable source not be shown as zero? An unavailable source cannot be shown as zero because no successful observation established that zero qualifying records exist. Missing, stale, partial, and failed are states of knowledge; valid zero is an observed result from a successful current source contract. Conflating them would create false reassurance and corrupt evaluation. ### Can NDVI identify an individual dangerous tree? No, NDVI cannot identify an individual dangerous tree. It measures vegetation greenness or condition at the sensor's spatial and temporal resolution, not tree height, branch condition, species, conductor clearance, or corridor encroachment. Field inspection and utility vegetation-management expertise remain necessary. ### Is GeoGridIQ an official outage map or emergency alerting system? No, GeoGridIQ is neither an official utility outage map nor an emergency alerting system. Users should consult utilities for current outage status and ECCC or the appropriate authority for warnings. GeoGridIQ provides supplementary risk intelligence for review, not replacement operational truth. ### Which regions and forecast horizons are currently available? Current regions and horizons are exactly those marked available in the server-generated capability table with a valid as-of timestamp. Each region and horizon row is evaluated independently from authoritative source, registry, artifact, and current-batch records. An artifact without a current batch, or one available horizon in a region, does not establish broader availability. ### Does GeoGridIQ guarantee that an outage will or will not occur? No, GeoGridIQ never guarantees that an outage will or will not occur. Forecasts are uncertain, false positives and misses occur, and source or model limitations can change interpretation. Users must read probability or rank together with confidence, mode, freshness, consequence, and the valid window. ### How can operators use the information responsibly? Operators can use GeoGridIQ responsibly as one reviewable input for monitoring, briefings, preparedness, inspection attention, and post-event learning. They should verify source health and window, distinguish probability from confidence and consequence, consult official information, apply utility procedures and field expertise, and record actions and outcomes. The system does not autonomously dispatch crews or override safety authority. ## 27. Sources, related documentation, and next steps The source register identifies publisher, source type, title, date or version, review date, and relevance, while adjacent citations show which claims each source supports. Internal reading paths connect Data Sources, Model Validation, Limitations, Outage Prediction, the outage-prediction and Forecast Accountability articles, vegetation and Critical Infrastructure methodology, status-safe GeoSQL material, current coverage, and the Knowledge Hub. Sources: [GeoGridIQ — Data Sources documentation](/docs/data-sources/), [GeoGridIQ — Current regional coverage and capability status](/coverage/). The primary next step is to review current GeoGridIQ data and prediction availability. The secondary next step is to explore the outage-intelligence dashboard. These calls to action support verification and bounded use; they never promise an exact failure, prevention, savings, or guaranteed forecasting. Account and dashboard access continue to follow the product's existing authentication and authorization flow. Sources: [GeoGridIQ — Current regional coverage and capability status](/coverage/). Related: [Review current availability](/coverage/) Related: [Explore the dashboard](/dashboard/map/) Related: [Browse the Knowledge Hub](/knowledge-hub/) ### Reviewed source register - **GGI_MODEL_STATUS — GeoGridIQ**. [Model trust, artifact, batch, and validation contract](/docs/model-validation/). reproducible internal evidence; reviewed 2026-07-26; relevance/limit: The public validation contract defines eligibility gates; current availability remains server-generated.. - **GGI_VEGETATION_PIPELINE — GeoGridIQ**. [Vegetation provider and data-freshness implementation](/sources/ndvi-data/). reproducible internal evidence; reviewed 2026-07-26; relevance/limit: Provider adapters and configuration do not by themselves prove a current usable observation.. - **POSTGIS_DATA_MODEL — PostGIS Project**. [PostGIS 3.6 spatial data management](https://postgis.net/docs/manual-3.6/en/using_postgis_dbmanagement.html). primary technical documentation; reviewed 2026-07-29; date/version 2026-06-08. - **POSTGIS_QUERIES — PostGIS Project**. [PostGIS 3.6 spatial queries](https://postgis.net/docs/manual-3.6/en/using_postgis_query.html). primary technical documentation; reviewed 2026-07-29; date/version 2026-06-08; relevance/limit: An indexed or successful query is not proof that its source, CRS, join, or interpretation is correct.. - **POSTGIS_VALIDITY — PostGIS Project**. [ST_IsValid](https://postgis.net/docs/manual-3.6/en/ST_IsValid.html). primary function documentation; reviewed 2026-07-29; date/version 2026-06-08; relevance/limit: Structural validity does not establish semantic correctness or operational fitness.. - **POSTGIS_TRANSFORM — PostGIS Project**. [ST_Transform](https://postgis.net/docs/manual-3.6/en/ST_Transform.html). primary function documentation; reviewed 2026-07-29; date/version 2026-06-08. - **GGI_GEOSQL_STATUS — GeoGridIQ**. [GeoSQL pilot status and governance audit](/blog/how-geogridiq-uses-geosql-map-in-the-loop/). reproducible internal evidence; reviewed 2026-07-29; date/version 2026-07-29; mutable source; relevance/limit: The article documents an isolated, governed pilot boundary and does not establish production writes or production outage modeling.. - **F1 — Environment and Climate Change Canada**. [Weather alerts](https://www.canada.ca/en/services/environment/weather/severeweather/weather-alerts.html). primary operational documentation; reviewed 2026-07-26; date/version 2026-02-20; mutable source. - **F3 — Environment and Climate Change Canada**. [Meteorological Service of Canada free data service](https://www.canada.ca/en/environment-climate-change/services/weather-general-tools-resources/weather-tools-specialized-data/free-service.html). primary technical documentation; reviewed 2026-07-26; mutable source. - **F4 — Environment and Climate Change Canada**. [GDPS documentation](https://eccc-msc.github.io/open-data/msc-data/nwp_gdps/readme_gdps-datamart_en/). primary technical documentation; reviewed 2026-07-26; mutable source. - **F5 — Environment and Climate Change Canada**. [RDPS documentation](https://eccc-msc.github.io/open-data/msc-data/nwp_rdps/readme_rdps-datamart_en/). primary technical documentation; reviewed 2026-07-26; mutable source. - **F6 — Environment and Climate Change Canada**. [HRDPS documentation](https://eccc-msc.github.io/open-data/msc-data/nwp_hrdps/readme_hrdps-datamart_en/). primary technical documentation; reviewed 2026-07-26; mutable source. - **F7 — Environment and Climate Change Canada**. [GEPS documentation](https://eccc-msc.github.io/open-data/msc-data/nwp_geps/readme_geps-datamart_en/). primary technical documentation; reviewed 2026-07-26; mutable source. - **O1 — Hydro-Québec**. [Power outage FAQ](https://pannes.hydroquebec.com/poweroutages/understand-and-prevent/faq.html). primary operational guidance; reviewed 2026-07-26; mutable source. - **O2 — Hydro-Québec**. [Understand and prevent outages](https://pannes.hydroquebec.com/poweroutages/understand-and-prevent/). primary operational guidance; reviewed 2026-07-26; mutable source. - **O3 — BC Hydro**. [How power is restored](https://www.bchydro.com/safety-outages/power-outages/during-an-outage/how-power-is-restored.html). primary operational guidance; reviewed 2026-07-26; mutable source; relevance/limit: BC Hydro example; not a universal restoration order.. - **V1 — United States Geological Survey**. [Landsat Normalized Difference Vegetation Index](https://www.usgs.gov/landsat-missions/landsat-normalized-difference-vegetation-index). primary technical documentation; reviewed 2026-07-26; mutable source. - **V2 — United States Geological Survey**. [NDVI — foundation for remote-sensing phenology](https://www.usgs.gov/special-topics/remote-sensing-phenology/science/ndvi-foundation-remote-sensing-phenology). authoritative technical synthesis; reviewed 2026-07-26; date/version 2018-11-27. - **V5 — Matikainen et al.**. [Remote sensing methods for power-line corridor surveys](https://doi.org/10.1016/j.isprsjprs.2016.04.011). peer-reviewed review; reviewed 2026-07-26; date/version 2016-09. - **G1 — BC Hydro**. [Tree-management program](https://www.bchydro.com/safety-outages/trees-power-lines/pruning-removing-trees.m.html). primary operational guidance; reviewed 2026-07-26; mutable source. - **G2 — North American Electric Reliability Corporation**. [FAC-003-5 Transmission Vegetation Management](https://www.nerc.com/pa/Stand/Reliability%20Standards/FAC-003-5.pdf). primary standard; reviewed 2026-07-26; date/version 2024-04-01 effective; relevance/limit: Applicable bulk-transmission scope; not a universal distribution standard.. - **I1 — Public Safety Canada**. [National Strategy for Critical Infrastructure](https://www.publicsafety.gc.ca/cnt/rsrcs/pblctns/srtg-crtcl-nfrstrctr/index-en.aspx). primary official policy; reviewed 2026-07-26; date/version 2009. - **I2 — Public Safety Canada**. [Current critical-infrastructure overview](https://www.publicsafety.gc.ca/cnt/ntnl-scrt/crtcl-nfrstrctr/index-en.aspx). primary official policy; reviewed 2026-07-26; mutable source. - **I3 — Public Safety Canada**. [Risk Management Guide for Critical Infrastructure Sectors](https://www.publicsafety.gc.ca/cnt/rsrcs/pblctns/rsk-mngmnt-gd/index-en.aspx). primary official guidance; reviewed 2026-07-26; date/version 2010-07-01. - **A1 — National Institute of Standards and Technology**. [AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1). primary standards guidance; reviewed 2026-07-26; date/version 2023-01-26. - **A2 — World Meteorological Organization**. [Forecast verifications](https://community.wmo.int/site/knowledge-hub/programmes-and-initiatives/wmo-integrated-processing-and-prediction-system-wipps/forecast-verifications). primary intergovernmental guidance; reviewed 2026-07-26; mutable source. - **A3 — European Centre for Medium-Range Weather Forecasts**. [Reliability diagram](https://charts.ecmwf.int/catalogue/packages/subseasonal/products/s2s-verification-ecmwf-rel). primary operational guidance; reviewed 2026-07-26; mutable source. - **A4 — European Centre for Medium-Range Weather Forecasts**. [Verification metrics guide](https://confluence.ecmwf.int/pages/viewpage.action?pageId=363864334). primary technical guidance; reviewed 2026-07-26; mutable source. - **A5 — Saito and Rehmsmeier**. [Precision-recall is more informative than ROC for imbalanced data](https://doi.org/10.1371/journal.pone.0118432). peer-reviewed primary methodological; reviewed 2026-07-26; date/version 2015-03. - **A6 — Kapoor and Narayanan**. [Leakage and reproducibility failures in machine-learning science](https://doi.org/10.1016/j.patter.2023.100804). peer-reviewed systematic study; reviewed 2026-07-26; date/version 2023-08-04. - **C1 — Government of Canada-led assessment**. [Canada's Changing Climate Report, Chapter 4](https://changingclimate.ca/CCCR2019/chapter/4-0/). authoritative government synthesis; reviewed 2026-07-26; date/version 2019. - **C2 — Government of Canada-led assessment**. [Canada in a Changing Climate — National Issues, Chapter 2](https://changingclimate.ca/national-issues/chapter/2-0/). authoritative government synthesis; reviewed 2026-07-26; date/version 2021. - **M1 — Taylor et al.**. [Machine learning evaluation of storm-related transmission outage factors and risk](https://doi.org/10.1016/j.segan.2023.101016). peer-reviewed primary; reviewed 2026-07-26; date/version 2023-06; relevance/limit: Northeastern United States study; not Canadian product validation.. - **M2 — Taylor et al.**. [Community power outage prediction modeling for the Eastern United States](https://doi.org/10.1016/j.egyr.2023.10.073). peer-reviewed primary; reviewed 2026-07-26; date/version 2023-11; relevance/limit: Not GeoGridIQ validation.. - **M3 — Fatima et al.**. [Machine learning for power outage prediction during hurricanes — an extensive review](https://doi.org/10.1016/j.engappai.2024.108056). peer-reviewed review; reviewed 2026-07-26; date/version 2024-07. - **GGI_IMPLEMENTATION — GeoGridIQ**. [Prediction implementation, contracts, and regression tests](/docs/model-validation/). reproducible internal evidence; reviewed 2026-07-30; date/version Repository state reviewed 2026-07-30; mutable source; relevance/limit: Code and tests define the public semantics; this record is not a current availability claim.. - **GGI_CAPABILITY — GeoGridIQ**. [Current regional coverage and capability status](/coverage/). current verified product state; reviewed 2026-07-30; date/version Server-generated at request time; mutable source; relevance/limit: Current status must pass exact region, horizon, artifact, batch, validity, and freshness gates.. - **GGI_DATA_SOURCES — GeoGridIQ**. [Data Sources documentation](/docs/data-sources/). product documentation; reviewed 2026-07-30; date/version 2026-07-30 review; mutable source. - **GGI_LIMITATIONS — GeoGridIQ**. [Limitations documentation](/docs/limitations/). product documentation; reviewed 2026-07-30; date/version 2026-07-30 review; mutable source. - **XGBOOST_DOCS — XGBoost Project**. [XGBoost documentation](https://xgboost.readthedocs.io/en/stable/). primary technical documentation; reviewed 2026-07-30; mutable source; relevance/limit: Library documentation does not establish GeoGridIQ model trust or performance.. - **SKLEARN_CALIBRATION — scikit-learn Project**. [Probability calibration](https://scikit-learn.org/stable/modules/calibration.html). primary technical documentation; reviewed 2026-07-30; mutable source; relevance/limit: Method reference only; no current GeoGridIQ calibration metric is inferred from it.. - **OGC_SFA — Open Geospatial Consortium**. [Simple Feature Access standard](https://www.ogc.org/standards/sfa/). primary open standard; reviewed 2026-07-30; mutable source. - **EPSG_URI — International Association of Oil & Gas Producers**. [EPSG OGC URI definitions](https://epsg.org/ogc_uri.html). primary registry documentation; reviewed 2026-07-30; mutable source. - **MAPLIBRE_DOCS — MapLibre**. [MapLibre GL JS documentation](https://maplibre.org/maplibre-gl-js/docs/). primary technical documentation; reviewed 2026-07-30; mutable source; relevance/limit: Rendering technology does not establish source accuracy or geographic authority.. ### Related documentation - [Data Sources](/docs/data-sources/) — Source categories, coverage, and public evidence context. - [Model Validation](/docs/model-validation/) — Forecast accountability, trust, and evaluation concepts. - [Limitations](/docs/limitations/) — Decision-support boundaries and current-use cautions. - [Outage Prediction](/outage-prediction/) — How risk outputs appear in the product. - [How outage prediction works](/blog/how-outage-prediction-works/) — A shorter educational introduction to target, signals, and trust gates. - [Forecast accountability](/blog/forecast-accountability-outage-prediction/) — Why stored forecasts and later outcomes must be compared honestly. - [NDVI and vegetation data](/sources/ndvi-data/) — Remote-sensing semantics and limitations. - [Critical Infrastructure](/critical-infrastructure/) — Consequence-aware generalized exposure context. - [GeoSQL map-in-the-loop](/blog/how-geogridiq-uses-geosql-map-in-the-loop/) — The isolated spatial QA pilot and its governance boundary. - [Coverage and availability](/coverage/) — Current public-safe regional and provider status. - [Knowledge Hub](/knowledge-hub/) — All GeoGridIQ research and public documentation.