Changelog

The land-parcel estate is now disclosed whole: ~579.8M cadastral parcels across 72 jurisdictions, listed per jurisdiction on /data and /coverage. Both pages now carry a Land parcels register generated at render time from the live parcel rights register: jurisdiction, estimated row count, licence verbatim, commercial status and source, never a hand-kept list. The register discloses tiers rather than hiding them: 41 jurisdictions are cleared and served (~458.4M parcels); 1 is derived-only (derivatives serve, raw geometry never does); 30 internal-research jurisdictions (~69.2M) are listed with their tier stated and no data access, so "not available" is never mistaken for "not held". Fabric, never ownership: no served jurisdiction carries proprietor data. Figures are planner estimates as of this entry; the registers themselves re-render hourly from live state.

Point and viewport parcel lookups: GET /v2/spatial/parcels. A lat/lon or bounding box returns the cadastral parcels there, with the publisher's fields verbatim plus per-row jurisdiction, licence and commercial status resolved live from the parcel rights register on every request. Rights fail closed: a non-cleared jurisdiction asked for by name is refused explicitly with its tier named (403 rights_hold), and an unfiltered window that crosses rights-held parcels discloses those jurisdictions in a withheld block instead of silently omitting them. Callable by agents through kortex_search_spatial the moment it shipped, because the spatial tool derives its layer list from the live router.

One front story: site screening for large grid connections; load, generation and storage. The site now leads with a single application. The former Applications menu is a Site screening menu (the live screen, worked examples, pilots and pricing, ownership intelligence); the other application surfaces remain, reachable from all applications, presented as what they are: the same evidence layer pointed at counterparties and portfolios rather than sites. The screen itself is unchanged; the scope statement is broader and truer: any large grid connection, not only load. And the builder is now stated plainly on the homepage and /trust: Kortex is built by Orkora, a funded floating-nuclear developer, to screen sites for its own gigawatt-scale plants; the same evidence layer is what we sell.

Pricing is published. The pilot page now states the ladder in the open: a free 14-day evaluation key issued instantly, a 30-day pilot at £48,000 credited in full against a first annual contract, and an annual desk from £150,000. Priced against the alternative; a senior analyst, six weeks of joining datasets per screen, and the risk of a wrong join surviving into a committee paper.

Company figures on /trust brought current. The funding statement previously read "$8m institutional round, July 2026"; the group has raised over $10m in institutional investment and the page now says so.

The derived grid topology reaches the governed surface: 9.8 million connectivity rows, plus five more multi-publisher layers, now served with honest provenance. Sixteen previously unreachable tables joined the governed catalogue after a loader-by-loader rights trace. The largest is the derived connectivity vehicle (grid nodes and line terminations, built from OpenInfraMap and OpenStreetMap, both ODbL); alongside it the DNO network-line layer, water assets, telecom sites, the data-centre layer and the ERCOT station crosswalk. Where a table draws on several publishers it is registered as a multi-source vehicle naming its constituents, with per-row source and licence columns doing the disclosure; no multi-publisher table is credited to a single name. Exemption-aware serving coverage moves to 99 percent; the six tables still withheld await genuine rights determinations, not wiring.

NGED's embedded-generation register now carries counterparty identity, direct from the operator, monthly. The National Grid Electricity Distribution slice of the embedded capacity register is now loaded from NGED's own monthly publication rather than the national combined register, which was six months behind and strips the fields that matter: every one of the 7,211 rows now carries the customer name and site NGED itself publishes, plus meter identifiers, native grid coordinates, and MVA capacity figures kept verbatim beside MW; the other five operators' rows are untouched and keep their own source and licence per row. For a screen this means an embedded generator connects to a named counterparty, not an anonymous point: generation joined toward the corporate ownership layer.

Three hazard-context viewports join the spatial surface: recent seismicity, tropical-cyclone exposure, and ice-mass trends. /v2/spatial/seismic-exposure aggregates the recent USGS event record per mapped infrastructure point and says plainly what it is not: a rolling ~six-month window of activity, never a hazard model, so a quiet season at a fault line reads low by design. /v2/spatial/storm-exposure carries NOAA IBTrACS track history per asset: storm counts, strongest wind, highest category, seasons affected. /v2/spatial/ice-mass-trends serves NASA GRACE-FO mascon trends at basin grain (~300 km), for regional water and cryosphere context, never site screening. All three are the estate-wide siblings of hazard context the 59-site deployment vehicle already carried per site; each row states its source, evidence class and grain.

Entity profiles gain a macro-context section: the rates environment beside the modelled revenue, never blended with it. Every corporate profile now carries the entity's home-country policy rate, consumer-price inflation and 10-year yield (BIS, ECB and FRED series; every row states its series id and as-of date, so staleness is disclosed rather than hidden), plus United States corporate credit spreads (BBB and high-yield option-adjusted) as the market reference. The section is context only: it is never an input to the modelled fleet-revenue estimate, and the two are presented side by side precisely so they cannot be confused.

Methodology 1.31: HVDC converter stations no longer count as AC connection evidence in the siting screen. The voltage-integrity check has always exempted converter stations from the plausibility ceiling, and rightly so; the Western HVDC Link genuinely operates at 600 kV. But exemption left those stations inside the substation count, the best-voltage figure and the counterparty election, so Great Britain reported a best connection voltage of 600 kV at a DC corridor endpoint no load can connect to: the correct number for the wrong concept. Converter stations are now identified by the same predicate the ceiling check exempts, counted separately, disclosed per zone as converters_excluded, and excluded from all three figures. Great Britain now reads best 400 kV from 1,026 genuine AC connection points, with 19 converter stations disclosed alongside. No confidence cost is charged for the exclusion: it is a class distinction on sound data, not a doubt about it.

Australia's National Electricity Market joins the estate: 5-minute dispatch ground truth for all five regions. Four new datasets from AEMO's public NEMWEB archive, August 2025 to present, verbatim: regional energy and frequency-control prices (aemo_dispatch_prices), binding and violated network constraint solutions with shadow prices (aemo_binding_constraints, 2.1M rows across 2,251 distinct constraints), regional dispatch summaries including operational demand (aemo_region_dispatch), and interconnector flows against their dispatch limits for all six interconnectors (aemo_interconnector_flows). All four serve through the governed table catalogue and MCP; rights are cleared for any purpose with attribution under AEMO's Copyright Permissions, carried on every row. Australia is the first market outside Europe and North America with constraint-grain coverage here.

Study 15: the Australian corridors, tested on fully observed data. The corridor doctrine's fourth market: an interconnector at its dispatch limit coincides with inter-regional price separation at Spearman +0.492 median across six corridors, +0.811 on Victoria–New South Wales; pre-registered, passed, with the misses reported on the validation register. The one apparent directional anomaly (Basslink) was investigated the same day and resolved as a labeling artifact of its frequency-control-coupled dispatch envelope; the study text carries the resolution and the withdrawn conjecture.

Mapped voltages are now checked against an independent source, not only against a ceiling. Since 1.28 a mapped substation voltage was tested against its jurisdiction's AC transmission ceiling, and where no ceiling was registered the check abstained rather than assumed. That catches a value which cannot be true. It does not catch a value that is wrong but in range, which is most of the ways a mapped voltage can be wrong. We now compare every mapped substation against the nearest bus in an independently built European transmission model (PyPSA-Eur, open-licensed) and record a contradiction where a reference within 250 metres matches no voltage level recorded at that site. Across the mapped estate, 4,946 substations corroborate and 460 are contradicted at close range. The two signals are graded, not merged. An implausible voltage cannot be true and is excluded from every figure. A contradicted voltage may still be correct and is merely unconfirmed, so it is excluded from the best-voltage figure and the counterparty election, which are the decisive uses, but stays in the substation count, is disclosed separately on the answer as claims_contradicted, and costs less confidence than an impossibility. Treating the two identically would tell you we had disproved a value when we had only failed to confirm it. A reference without coordinates cannot raise a contradiction, however authoritative it is. A transmission owner's own site index is keyed by name, and town names repeat across voltage levels: a 66 kV distribution substation and a 400 kV grid site can share one name. A name match is evidence about a name, not about a place, so it raises a worklist item and never a contradiction. No publisher value is overwritten. Contradictions are correction worklists for the upstream mapping; the mapped value stays as the publisher recorded it and the disagreement is carried alongside it.

Our correction was itself uncontrolled. An outside reviewer found it, and here is the controlled rerun. On 9 August we withdrew the flagship worked example, reran it under methodology 1.29 and published the comparison as evidence that this system fixes its own defects. The comparison was not like-for-like. Candidate B, Leeds, was screened at 150 MW in the withdrawn v1.7 receipt and at 300 MW in the replacement, while the other four candidates held at 300; the public copy nonetheless said the identical parameters were rerun. It was material rather than cosmetic: Leeds' connection-grade evidence is 202.6 MW available at Kirkstall B, a 132 kV grid supply point, which is short of a 300 MW target and comfortably clear of a 150 MW one, so the changed number could have been carrying the verdict. Two things were wrong and only one was the number. The receipt could not have shown it. A receipt froze the request-level target only, so a screen where one candidate carried its own target recorded parameters identical to one where it did not: both said 300 MW, and the difference lived where nobody would check. Receipts now freeze the per-site targets as well, so two materially different requests can no longer produce the same frozen record. The scenario is also declared in one place now and the rerun refuses to publish a comparison whose frozen inputs do not match that declaration, which is the check the phrase "identical parameters" never had. The controlled rerun, at Leeds' original 150 MW, holds it at investigate, and the only verdict that moves between v1.7 and v1.29 is Dublin, which clears to proceed. The conclusion we published survived; the description of what we ran did not, and a claim about our own evidence should not have been the part that went unchecked.

We published a test we cannot run, and why we cannot run it. Validation study 14 showed that a queue appraisal separates built projects from withdrawn ones, but that county-grain context adds almost nothing, because a county is far too coarse a unit for a siting question. The natural next step is node grain: do the binding constraints and the nodal congestion at the connection substation predict the outcome? Kortex holds seven ISO constraint records and a nodal price record, and the answer is that this test is not currently possible on three independent counts, each measured. The constraint records all begin on 1 January 2025 while the last dated queue outcome is 30 December 2024, so the features and the outcomes never overlap; only 9.1% of ERCOT's resolved projects can be placed on a mapped constraint station, and 11.4% even if every station ever seen binding were mapped, because the queue names connection points as bus numbers against a different naming system; and every ERCOT queue project is held at county centroid, so no spatial join escapes the county grain. The nodal price leg fails for a fourth reason worth stating: the 44 projects that resolve inside the price window are all completions, so there is nothing to discriminate. The three pass conditions are declared in the feasibility gate before the data is looked at, and it is re-run when a new vintage lands. Publishing a negative feasibility result is deliberate: the alternative was to run the join on whatever data happened to align and present the number.

Our flagship worked example failed our own newer rules. We have withdrawn it, rerun it, and published the difference. The worked example on this site was computed under methodology v1.7 in July. Since v1.27 every piece of evidence carries a classification: which direction it points (import demand or generation export), which voltage grain it lives at, and which connection tiers it may therefore decide rather than merely inform. Under those rules two of the v1.7 findings should never have fired. Dublin was moved to investigate by a 38 kV station group tested against a 300 MW target, and Teesside by a 66 kV supply point against the same target; no point at either grain could carry that load. Leeds was moved partly by the Embedded Capacity Register, which is predominantly a generation-direction register, and export queue capacity does not consume a demand target's import headroom. The five coordinates were rerun under v1.29 (corrected 10 August 2026: this said "identical parameters", and Candidate B's target had in fact moved from 150 MW to 300 MW; see the entry of 10 August): Dublin clears to proceed, and the other four hold their verdicts on evidence at the applicable grid-supply-point and 132 kV+ grain, with Teesside's constraint restated at a 132 kV GSP reading the same 60.4 MW. The newer rules changed the reasons more than they changed the answers. The v1.7 receipt is retained, marked retracted with its reason and linked to its replacement in both directions; nothing was deleted.

Every receipt now freezes the rules it was decided under, not just a version number (methodology 1.29). Until now a receipt pinned the answer and a methodology version label, but never the methodology text, so an export rendered months later showed the current rulebook above findings computed under older rules. That was true of every receipt we have ever archived, not only the worked example. From v1.29 each receipt carries its complete rulebook, including the screening thresholds and the evidence-applicability registry rows, with a SHA-256 over it: response_sha256 pins what was concluded and methodology_sha256 pins the rules it was concluded under, so both verify from the artifact alone. The applicability registry is included deliberately, because since v1.27 the question of which evidence may decide is answered from a table rather than from code. Receipts archived before this change state on their face that their rulebook was not pinned; they are never back-filled with rules they were never computed under. Supersession now reads in both directions, and retraction is a recorded state with a reason that appears above the title on any exported memo.

Tens of thousands of low-voltage assets were being served as transmission infrastructure. They are corrected, and the check that catches them now runs nightly. OpenStreetMap expresses voltage in volts. One of our two voltage parsers converted volts to kilovolts only when the value exceeded 1000, so anything below that was served verbatim as kilovolts: Paris Metro supply at 400 volts became 400 kV, London Underground and tram rectifiers at 750 volts became 750 kV, a university laboratory switchroom at 415 volts became 415 kV. 67,160 mapped lines and 773 substations carried the defect, 58,645 and 566 of them respectively reading at transmission tier. It stayed invisible because the wrong numbers are plausible ones: 400 kV is the commonest transmission voltage in Europe. The parser is corrected, every affected row is preserved in quarantine with its pre-correction value, and the plausibility regime introduced in v1.28 has been extended from GB transmission lines to substations across every jurisdiction where we can state a voltage ceiling with a basis. Zone screens now exclude claims that exceed their jurisdiction's ceiling from substation counts, best-voltage figures and the dominant-counterparty election, disclose how many were excluded, and reduce the evidence confidence accordingly. HVDC converter stations are explicitly exempt, because links such as the Western HVDC Link at 600 kV genuinely exceed the AC ceiling and a blunt check would have deleted real assets. Where we have not established a ceiling for a jurisdiction, nothing is flagged: not checked and not plausible must never read the same.

The mapped grid inventory now has to pass plausibility checks, and the claims that fail them say so (methodology 1.28). Community-mapped topology is the best open picture of the physical network, but it can be wrong in ways a count never catches: a distribution licensee mapped at 400 kV, a 515 kV circuit on a system whose ceiling is 400, the Irish distribution operator appearing on GB assets. The inventory is now swept against country-scoped invariants of how each jurisdiction's network is actually structured, and in the US it is corroborated against the EIA transmission inventory as an independent federal lineage: at substation grain the two lineages agree on 92.9% of matched connectivity, and the 1.2% of true conflicts sit on a correction worklist rather than in the product. 923 GB claims currently fail an invariant. None of them has been removed: the power-lines map layer serves the failed invariant on the affected feature, and a site screen's connection context states how many mapped claims within 25 km are under review. A flag is a correction worklist for the upstream mapping, never a deletion; and a claim that passes every check is still a mapping, not an operator record.

The land-fabric check reaches ten jurisdictions and 279.3 million parcels (methodology 1.26). Yesterday the site screen could answer "what parcels does this boundary sit on" in four places. It now answers in ten: England & Wales and Scotland, plus open national cadastres for France (94.7M), Italy (86.8M), Spain (52.0M), Czechia (22.0M), Belgium (9.1M), the Netherlands (8.4M) and Luxembourg (0.7M), and sub-national fabric for British Columbia (2.5M) and Utah (1.6M). Most of Europe is open for the same underlying reason: cadastral parcels are a high-value dataset under the EU Open Data Directive, so member states are obliged to publish them freely and openly licensed.

Two postures are carried rather than flattened, because they are not the same kind of permission. Spain is derived-only: the Dirección General del Catastro permits commercial use of transformed products and expressly forbids redistributing the original, so the screen's parcel counts and area shares are servable while raw Spanish parcel geometry never is — and the evidence block says so on every Spanish answer rather than leaving it to a footnote. And the United States differs in kind from Europe: European cadastres publish geometry with ownership held separately, but US state layers routinely bundle owner names, mailing addresses and sale prices into the parcel record. Those layers are therefore ingested under a strict field allowlist that drops owner identity at read, so it never reaches the database or a backup; a guard refuses to run if a field mapping matches an owner or mailing pattern. Utah's fabric is held without a single owner name.

Absence still means absence. Where a boundary extends beyond a cadastre, that is frequently the deliberately non-cadastred public domain — roads, watercourses and rights of way in France and Belgium alike — so the screen does not call the shortfall a mapping gap, and never a suggestion that the land is unowned. Coverage carve-outs are disclosed rather than implied away: Trento and Bolzano run their own cadastres and are not in the Italian product; the Basque foral territories and Navarra are not in the Spanish one; and federal land and First Nations reserves are only thinly represented in British Columbia. In every jurisdiction this is fabric grain only — none of these products carries proprietor data and none is inferred.

Land fabric goes multi-jurisdiction: three national cadastres, 104.7 million parcels (methodology 1.25). Until today the site screen could answer "what parcels does this boundary sit on" only in England and Wales, through the HM Land Registry INSPIRE index. It now answers in three more jurisdictions from their open national cadastres: Scotland (Registers of Scotland, 1,564,345 parcels across 33 historic counties, OGL v3), France (Etalab PCI vecteur, 94,731,804 parcels across 34,916 communes and all 101 départements including Corsica and the overseas départements, Licence Ouverte 2.0) and the Netherlands (Kadaster Kadastrale Kaart, 8,439,086 parcels across 1,110 kadastrale gemeenten, CC BY 4.0). A site is in one country, so the two layers never both answer, and the evidence names which one did along with its attribution. One distinction is kept deliberately rather than smoothed away: HM Land Registry records registered freehold, a tenure concept, while a continental cadastre records parcels and no tenure at all — so counts from the cadastres report as cadastral parcels and never as registered freehold, and the field names differ for the same reason. In every jurisdiction this is fabric grain only: none of these products carries proprietor data, and none is inferred. A related honesty fix: where a boundary extends beyond the cadastre, that is often the deliberately non-cadastred public domain — in France roads, watercourses and rights of way — so the screen no longer calls the shortfall a mapping gap, and never a suggestion that the land is unowned.

A capacity source we expected to add, and could not. We went looking for Dutch grid capacity to sit alongside the six GB distribution networks and France's Caparéseau, on the strength of a station-level capacity map published through the national geo-portal with a lookahead of three, five and ten years. It no longer exists: the dataset was withdrawn from production in June 2025 at the publisher's request, and the endpoint returns a 404. The national capacity map that replaced it is an interactive application with no public data interface, and the two largest regional operators put their equivalent datasets behind an email request form. So there is no Dutch capacity layer in Kortex, and the coverage register says so rather than implying the gap is temporary. Recording it here because the mistake is instructive and ours: the source was ranked on its published description instead of on a request to the endpoint, which is exactly the check that would have caught it a day earlier.

One source cleared, one honestly not. The PyPSA-Eur prebuilt European transmission network clears: its Zenodo record declares ODbL 1.0, the same licence as our OpenStreetMap estate, so it carries the same posture (commercial use permitted, attribution, share-alike on any redistributed derivative database; we use it as an internal modelling substrate). Open-Meteo, reviewed the same sitting, does not clear: the captured terms restrict the free API to non-commercial purposes regardless of the CC-BY data licence, so the ERA5-derived series stay gated in commercial-safe mode until we either subscribe commercially or re-source the identical data directly from the Copernicus Climate Data Store. The licensing register's status badge links the terms in question, as every unsettled badge now does.

Screen receipts are now private to your organisation by default. A screen receipt archives your own candidate shortlist or counterparty list alongside the frozen answer; until today every receipt was a capability URL that anyone holding the id could retrieve. An external vendor-risk review named this the finding a bank blocks on, and it is fixed at the storage layer: receipts computed with your key are bound to your organisation, retrieval and export require your credential, and an unauthorised request gets the same 404 as a nonexistent id. Sharing is an explicit act: POST /v2/receipts/{id}/share makes the URL forwardable (the deliverable is still designed to be forwarded, once you decide that), unshare reverses it, and your organisation can delete any of its receipts outright. Receipts of public-data insights carry no customer data and remain shareable links, as do all receipts issued before today. The trust page now documents the full model: what is stored, who can see it, retention and deletion.

Fibre evidence gets a graded voice (methodology 1.24). The digital-infrastructure block in the site screen previously said "context only" about everything; the purchased worldwide OpenInfraMap extract (13,657 mapped telecom routes), PeeringDB's peering fabric and the operational-datacentre set deserve better than a footnote. The block now states a presence assessment the receipt can cite: "strong mapped presence" requires two independent signals in range (mapped route within 2 km, peering facility within 25 km, operational datacentre within 25 km), single signals report as such, and the upgrade is deliberately asymmetric — presence corroborates but never gates a verdict (findings are the exception ledger, and fibre presence is a virtue, not an exception), while community-mapped absence remains what it always was, a mapping gap that fires nothing, never "no fibre". Route diversity, carrier counts and latency remain unmodelled and unclaimed.

GB published grid capacity is complete, and France joins whole-country (methodology 1.23). Two rights reviews settled today unlock data that was already ingested and gated. National Grid Electricity Distribution's network capacity map (1,180 Primary and BSP sites across all four licence areas, CONNECTED-lane MW) clears under the portal's WPD Open Data Licence, making Great Britain all six DNO groups. And Caparéseau, the S3REnR reception-capacity map RTE publishes with Enedis and 24 regional operators, clears under its stated terms (the portal's data is public and reusable with source citation, attribution "Source - capareseau.fr (RTE)"): France becomes the capacity vehicle's first whole-country layer, 3,119 postes sources with per-substation queue MW and connection levy, generation side only, figures served verbatim as the indicative published capacities they are. Determinations and captured terms are in the rights record; the site screen's published-capacity check now states both coverages in its methodology.

Coverage asymmetry in one grid, and Kortex in two pages. The coverage register now opens with a capability-by-geography matrix: ten screen-decisive capabilities across Great Britain, Ireland, the EU, the United States and the rest of the world, each cell stating full, partial with the boundary named, or none. The asymmetry was always disclosed, but per-check and in footnotes; now it is legible in one glance, and the detailed fitness table remains the governing record beneath it. Alongside it, /executive is a two-page summary written for the person who approves the engagement rather than the analyst who evaluates it: what is being bought, why it is different, the evidence, who stands behind it, and what Kortex is not. It prints clean.

A public status page at /status, with recorded rather than asserted uptime. A probe hits GET /health through the public edge every minute and every check lands in a permanent record; the page shows current component state, uptime over 24h/7d/30d/90d windows, a daily history strip, and an incident log. Two honesty rules are structural: the page states the date recording began (today) and never claims availability for any period before it, and a failed probe is a red minute rather than a smoothed one. Machine-readable at /status.json. This is the availability evidence that makes our stated no-formal-SLA position reviewable instead of just tolerable; a measured business-hours SLA draft now exists for first annual contracts.

A published data-protection statement at /data-protection. The page is the Article 13/14 transparency notice for the platform: which records in the product are personal data (register-published company officers and persons with significant control; ICIJ officer records held internally and never served), the legitimate-interests basis for them, the controller-processor split on customer candidate data, individual rights and the route to the ICO, and the complete subprocessor list by name: Hetzner Online GmbH (hosting and backups, Germany) and Resend, Inc. (transactional email, United States). Nothing on the page is a claim of absence; it says what exists and on what basis.

The trust page stops overstating two things and starts stating a third. The data-protection section previously said the product holds no personal data; public-register records such as company officers and persons with significant control are personal data in law even in business context, so the page now says exactly that, with the lawful basis and a data-protection statement available on request. The customer-data line that said we do not store customer datasets is replaced by the receipts model above, which is what actually happens. And a new section states plainly what was previously implicit: the answer path is deterministic, with no LLM and no learned scoring anywhere in it, which for buyers subject to model-risk frameworks typically makes Kortex a data tool rather than a model requiring validation.

The licensing register now reads as the institutional document it is. An external vendor-risk review caught internal register state leaking onto the public page: four sources' notes said terms were "to be confirmed by" a named individual, two source rows linked to a placeholder instead of a homepage, and three ODbL sources carried OpenStreetMap attribution boilerplate that belongs only to OSM-derived layers. The notes now state each posture without naming staff; the postures themselves are unchanged, and the four sources under review still fail closed in commercial-safe mode. Every row links to a real homepage, and each ODbL source now credits its own database: ICIJ Offshore Leaks, Microsoft Global ML Building Footprints and OpenFlights carry their own attribution, with OpenStreetMap credited where the data genuinely derives from OSM.

The resolved legal entity now leads every fossil-revenue row (methodology 1.3). An external review caught the screen headlining a legacy source label: a row read "E.ON" while its resolved operating entity is Uniper Kraftwerke GmbH, a divestiture the graph had already researched plant-by-plant. Row order now states the identity hierarchy: resolved entity first, source-attested label after it as evidence. Two new machine-readable warnings make the divergences explicit; label_diverges_from_resolved_entity when a label matches neither the entity nor its ultimate parent (subsidiary granularity such as NextEra resolving to Florida Power & Light does not warn), and fleet_split_across_rows_same_ultimate_parent when one corporate family spans several rows: the E.ON-labelled and Uniper-labelled rows both roll up to Uniper SE, and concentration reads understated unless rolled up. Every row now also carries the ultimate parent's legal name for exactly that roll-up.

Every public count now renders from the canonical registry, including the agent capsule. The llms.txt capsule and one coverage-page line were the last hand-maintained counts on the public surface, and both had rotted (572 tables quoted against 631 live). Both now render from the same introspection pipeline as every other page, stamped with their render date; where machine surfaces use wider table definitions (views included), the capsule says so instead of leaving the discrepancy to be discovered.

The US congestion record is complete: ISO-NE joins as the seventh organized market (methodology 1.22). 118,000 binding-constraint intervals from ISO Express (day-ahead hourly and real-time five-minute final reports, 2025 onward, refreshed daily) close the screen's last stated US market gap. Constraint name, contingency and marginal value verbatim; marginal values carry both signs in ISO-NE's convention, and the real-time final report lags the market day by about three days, stated on the evidence block. Every candidate site in a US organized market now gets a congestion answer from that market's own published record: ERCOT at monitored-element grain, the other six as system context that never moves a verdict. The remaining disclosed gap is real but different in kind: non-ISO utility territories (the Southeast and much of the West) publish no comparable binding-constraint record. Source under rights review: served in evaluation mode, withheld from commercial-safe responses until cleared.

NYISO joins the US congestion record (methodology 1.21). The sixth US market in the estate: 388,000 limiting-constraint intervals from NYISO's public MIS archive (day-ahead hourly and real-time dispatch, 2025 onward, refreshed daily) give New York candidate sites the same system-grain congestion context the other interior ISOs carry; limiting facility, point identifier, contingency and constraint cost verbatim, costs kept in NYISO's own mixed-sign convention. NYISO publishes constraint names without coordinates, so this is context that never moves a verdict. One disclosure worth reading: the screen resolves a site's ISO from interconnection-queue majority within 50 km, so Manhattan itself resolves PJM on the northern-New-Jersey queue mass; Albany, Syracuse and Buffalo resolve NYISO. ISO-NE remains the stated coverage gap. Source under rights review: served in evaluation mode, withheld from commercial-safe responses until cleared.

PJM joins the US congestion record (methodology 1.20). The largest US market, and the one Northern Virginia sits in, stops being the screen's own disclosed gap: 370,000 real-time binding-constraint intervals from PJM's Marginal Value feed (five-minute grain, 2025 onward, refreshed daily) now give PJM-footprint candidate sites the same system-grain congestion context MISO, SPP and CAISO carry: how congested the ISO's market actually ran, verbatim from its own reports, with shadow prices kept in PJM's own sign convention. PJM publishes constraint names without coordinates, so this is ISO-system context that never moves a verdict; site-grain congestion evidence remains an ERCOT capability, and footprints without an ingested feed (ISO-NE, NYISO) keep saying so. Source under rights review: served in evaluation mode, withheld from commercial-safe responses until cleared.

Observed connection durations are now cohorted by capacity band and era (methodology 1.19). The path-to-power ledger's duration evidence no longer pools all history into one distribution: a coverage audit of the GB Embedded Capacity Register showed recent-era sub-11kV connections completing in single-digit median months while 33kV-class 10 MW+ connections run around four years, so one pooled median misled precisely the buyers the number matters most to. Durations are now reported for the cohort comparable to the submitted site: capacity band chosen by the target load, energisation window defaulting to the last 36 months, local operator first; when a cohort is thin the widening (to GB, then to all register years) is stated on the block, operators whose registers publish no acceptance dates are named rather than silently absorbed, and every block carries its observation window. Still register history, never a forecast: the distribution says how fast comparable connections actually moved, not when a submitted site would connect.

Trial workspaces open on a worked example, and shortlists can be created in the app. A fresh trial workspace no longer starts empty: it opens on a seeded shortlist of three GB candidate sites, already screened across every check with the receipt attached and a daily monitor running, so the evaluate-screen-monitor loop is visible before the first click; the example is yours to extend or archive. New shortlists can now be created directly in the workspace as well as over the API, and the trial arc gained a third email: two days before expiry, a summary of what the workspace holds, answered by a member of the team to extend or convert; saved work is never deleted at expiry.

US grid congestion joins the candidate-site screen (methodology 1.18). Site screens in the four ingested ISO footprints now carry binding-constraint ground truth: ERCOT sites get site-grain evidence (SCED constraints on monitored elements whose stations map within 40 km, with binding-interval counts, shadow prices and a material finding when an element sits in the top decile of ERCOT constraint persistence); MISO, SPP and CAISO publish constraint names without coordinates, so those footprints get system-grain context that never moves a verdict; other US footprints disclose the coverage gap. Built on 3 million constraint intervals already flowing daily from the four ISOs' own reports; a shadow price is market ground truth, never a deliverability statement, and the evidence says so.

The instant evaluation key is now advertised where machines actually look. Anonymous API responses carry an X-Kortex-Trial header and a Link to the machine manifest; 401 responses explain the one-call mint in their error detail; robots.txt, the OpenAPI description and the agents page all point to POST /v2/keys/evaluation. Nothing new is gated: the same instant 14-day key, now discoverable from every entry point instead of only the contact form.

Outbound email delivers again. Evaluation keys, workspace sign-in links and alert mail now send from a verified domain; deliveries that had queued during the provider switch have been flushed.

AIS terms review concluded. The aisstream.io review that had been blocked at the network level completed today: the service publishes no terms of service at all; its documentation states no use restriction, and the operator openly welcomes commercial projects on its free API. The determination is recorded with archived captures: absent an explicit outbound data licence, anonymous surfaces keep serving derived aggregates only (counts and density, never a vessel identifier) and raw per-vessel layers stay key-gated. The vessel feed moves from "under review, fails closed" to a settled derived-only posture; nothing about what is served changed, only the confidence with which it is served.

Aircraft return to the live layer, on cleaner terms. Global aircraft activity is back on /globe: the live-activity mode now shows around 5,000 airborne aircraft alongside the vessel picture, with a worldwide density surface refreshed every few minutes. The feed is the adsb.lol community network, replacing the OpenSky feed paused on rights review the day before: feeders there contribute under CC0 and the operator publishes the API as an open drop-in service. Served as derived aggregates by construction, exactly like vessels: counts and one-degree density cells only, never an aircraft hex code, callsign or registration, in any access mode. Data from adsb.lol.

Every evaluation key now opens a private trial workspace. The instantly-issued evaluation key no longer stops at the API: each mint now creates its own private workspace on /app, reached by a sign-in link returned with the key. Save a shortlist as an investigation, attach boundary and engineering evidence, run the screen against it and set a daily monitor; every run mints the same receipts and verdict history a pilot workspace gets. Isolation is by construction: each trial lives in its own organization, the same tenancy boundary that separates customers, so no trial can see another. The workspace follows the key's 14 days; saved work is retained at expiry, not deleted, and continues into a pilot on request.

Scotland stops being the thin file. Two coverage gaps named by independent reviews closed the same day. First: SSEN Distribution's published network capacity joins the screen (998 substations across SEPD and SHEPD, including the north of Scotland, Orkney and Shetland): demand headroom in MVA, generation headroom in MW, RAG status and the named constraint, verbatim in the operator's own units under CC-BY-4.0. A Highlands candidate now reads its nearest published point (witness: a primary substation at 1.3 km showing 0.0 MVA, RAG Red) instead of a coverage apology. Second: Scottish SSSI boundaries from NatureScot (1,422 sites, dissolved from 15,877 notified-feature polygons, OGL) join the statutory-designations check with jurisdiction resolved against ONS constituent-country boundaries; a site inside a Scottish SSSI now gets the finding with the right authority named (NatureScot, not Natural England). Stated plainly on every Scottish result: coverage is SSSI only; National Scenic Areas, Scottish national parks and green belts are not yet ingested, and a clear result does not clear them. Methodology 1.17.

Findable, and recognisably Orkora. Kortex now carries the Orkora mark: the favicon and touch icons match orkora.com, and structured data states the relationship plainly (Kortex is a product of Orkora Ltd). The sitemap grew from 11 to 35 pages so every insight page, methodology note and data surface is indexable, joining the existing llms.txt and agent manifest that make the platform legible to AI assistants and crawlers.

Evaluation keys, issued instantly. The contact form now mints evaluation API keys on the spot: 14 days, faceted and insight scope, rate-limited, shown once with a quickstart. No wait, no exchange of emails; the key arrives the moment the interest exists. Pilot and general enquiries still land directly with the Kortex team.

The globe goes live. /globe gained a live-activity mode: a real-time stream of global vessel activity, updated every ten seconds from the AIS feed; the count of vessels moving right now and a worldwide density surface, rendered as it changes. Served as derived aggregates by construction: counts and one-degree density cells only, never a vessel identity, name or track; the stream closes when you leave the tab. Raw position layers remain key-gated pending provider licence review.

ADS-B collection paused. The OpenSky Network licence review concluded that commercial and operational API use requires a written licence; collection stopped the same day and nothing aircraft-derived is served anywhere. The collected archive is retained unserved. Aircraft data returns only under a written OpenSky licence or a replacement source whose terms clear review.

US constraint ground truth, complete across four ISOs. ERCOT joins MISO, CAISO and SPP: 1.8 million SCED binding transmission constraints with shadow prices at 5-minute grain, station-named and kV-tagged, plus the injection geography beside them; hourly wind and solar actuals by ERCOT geographical region and load by weather zone, 2025 to present, all refreshed daily. SPP's 214,000 binding-constraint rows landed on the same rail. A curated crosswalk geolocates 233 ERCOT station codes against the mapped substation layer, built fail-closed: a code that cannot be corroborated stays unmapped rather than guessed.

ERCOT cleared. Terms of Use reviewed and archived at review date: ERCOT's public raw data may be used, reproduced and redistributed in compilations, charts and analyses, and the API terms impose access limits that Kortex's fetch-once-store-locally design respects by construction. ERCOT-derived evidence now serves in commercial-safe mode.

Validation study 13: the model meets Texas, and fails exactly the right way. The US mirror of the German and GB corridor studies is published on /validation, falsification first: the pre-registered branch-grain test came out inverted (−0.39), the post-hoc interface-grain result is +0.65 with the mechanism named (ERCOT's own corridor interfaces bind first and unload the branches behind them), an independent recompute verified every statistic, and the out-of-sample confirmation criterion is frozen before the data that will judge it arrives. Third grid, same doctrine: corridor claims serve, branch claims do not.

US constraint ground truth, second ISO. CAISO binding transmission-interface constraints from OASIS, 2025 to present: 78,000 rows across the real-time (15-minute) and day-ahead markets, with interface, direction and constraint cause verbatim. Scope stated plainly: interface and named-path grain; CAISO does not publish per-branch shadow prices publicly and none are held or implied. Joins MISO's 879,000 branch-constraint rows; ERCOT's leg is built and waiting on API access.

Consistency, again. The pilot page's verify-us section counted nine validation studies while the header counted twelve; both now read from the generated register. The land-fabric check now names Scotland explicitly: HMLR INSPIRE covers England & Wales only, Scotland is Registers of Scotland (not ingested), and a Scottish site's absence of parcel fabric is a coverage gap, never a clear site.

The screen's coverage, stated where buyers read. The siting and pilot pages now say plainly what the candidate-site screen is for: the evidence layer that decides which sites deserve a POI-grade engineering study, a utility enquiry or a formal application; it never replaces one. The check list on /siting caught up with the week: boundaries as inputs, four queue geographies, statutory designations, land fabric, digital infrastructure and observed connection durations.

Planning friction, priced into the screen. England candidate sites are now tested against the statutory designations that decide planning outcomes (methodology 1.16): Green Belt, SSSI, AONB (National Landscapes) and National Parks, 4,362 boundaries from planning.data.gov.uk. A boundary site reports its area overlap per designation; a point site reports containment and the nearest designation within 5 km; any overlap is a material finding with the constraint named. Stated on the evidence: an overlap prices friction, it is not a refusal and not planning advice; a clear result covers statutory designations only; outside England, absence is a coverage edge, never a clear site. Witness: the Teesside test envelope turns out to hold 88.6 ha of Teesmouth and Cleveland Coast SSSI (19.9% of the site); yesterday the screen could not say that.

German sites read the federal register. Queue evidence now covers a fourth geography: candidate sites in Germany read the Marktstammdatenregister at unit grain within 50 km (methodology 1.15): registered-planned units and MW, operating units and MW, permanent retirements, register vintage, with a material finding when 1,000+ MW of registered planned capacity competes nearby. Stated on the evidence: the register is not a connection queue; "In Planung" is registered intent with no acceptance date or queue position, weaker than a contracted position, and units below 100 kW are out of scope at this grain. Witness: a Frankfurt candidate reads 228 registered planned units (1,180 MW) and 3,257 operating units (6.5 GW) within 50 km. Quelle: Marktstammdatenregister der Bundesnetzagentur (dl-de/by-2-0).

How long connections actually took, beside every candidate. The path-to-power ledger (methodology 1.14) now carries observed connection durations: the distribution of acceptance-to-connection times for already-connected projects on the local operator's register, as p25 / median / p75 with cohort size and register vintage. Operator cohort at 20+ observed pairs, GB-wide fallback disclosed. This is register history transformed, never a forecast: embedded grain (1 MW+ distribution connections, predominantly generation), it says how fast the operator's register has moved, not when your site would connect. The registers that carry too few dates (TEC, Irish, US queue vintages) show nothing rather than an interpolation. Witness: the same 300 MW target reads a 10.6-month median history at one GB operator and 34.4 months at another; that difference is now stated evidence.

Digital infrastructure joins the site screen. The candidate-site screen (methodology 1.13) now reports, per site: the nearest community-mapped telecom cable route with the 10 km route count (OpenStreetMap/OpenInfraMap), the nearest PeeringDB facility with network counts at 50 km, and the nearest operational datacentre. Distances measure from the asserted boundary edge where one is submitted. The honest limits, stated on every response: this is context that never moves a verdict; a mapped route nearby is not capacity, dark-fibre availability, route diversity or a wayleave; mapped-cable absence is a mapping gap, never "no fibre"; and submarine-cable and IXP registers are rights-held and deliberately not consulted. Witness: the Teesside test envelope sits 9.6 km from a mapped BT route with a 10-network peering facility at 48.8 km.

The brand colour no longer whispers a verdict. Until today the interface accent and the proceed verdict shared one teal, so every branded surface (buttons, links, focus rings) carried a faint "proceed". The accent is now cobalt and the semantic palette is sealed: teal means proceed or cleared, amber means investigate or under review, red means reject, grey means insufficient evidence, and nothing else borrows them. Alongside it: verdict chips are now machine-readable (data-verdict attributes, not colour classes), receipts and ledgers use tabular numerals with slashed zeros so columns of figures align and 0 is never O, and every page honours reduced-motion preferences. All colour pairings clear WCAG 2.2 contrast minimums.

First US constraint ground truth. MISO's daily binding transmission constraints, 2025 to present: 879,000 rows across the real-time (5-minute, preliminary shadow prices) and day-ahead (hourly) markets, verbatim at report grain with flowgate, branch and contingency names. The US counterpart to German redispatch and GB SO-flagged balancing actions; one transformer constraint bound for 20,787 intervals since January 2025.

The land under the site. The candidate-site screen (methodology 1.12) now reads the registered-freehold parcel fabric of England & Wales: 24.3 million HMLR INSPIRE index polygons across all 318 local authorities. A boundary site reports how many parcels it spans, the share of the site inside the index, and the largest single parcel's share, with a material finding when the fabric implies a real assembly (25+ parcels) or a material unindexed share; point sites get containment context only. The honest limits, stated on every response: INSPIRE parcel identifiers have no public mapping to title numbers or proprietors, so this is assembly complexity, never ownership; ownership evidence stays with the diligence screen's corporate land-control check. Witness: a 446 ha Teesside test envelope spans 50 parcels, largest 17.6% of the site. Source: HM Land Registry INSPIRE index polygons, Crown copyright, reproduced with the permission of HM Land Registry; polygons Crown copyright and database rights Ordnance Survey 100026316.

The homepage now stays on the siting journey. Platform and research material moved off the front page to where it belongs: the three worked graph questions to /applications, the platform-depth numbers and the graph-up-close record to /data, and the is/is-not positioning table to /pilot. The coverage summary was a duplicate of /coverage and is gone. Nothing was removed from the site; the homepage is about a third shorter and everything it dropped is one link away.

The evidence store, in the browser. /app gained an Evidence tab: upload a site outline or GIS layer (GeoJSON), a utility letter, connection study or consultant report (pdf, xlsx, docx and more, 25 MB per file), or a structured extract (JSON), with per-organisation quota, byte-identical dedupe and verbatim download. Shortlists gained an add-site form to match: a point, a stored outline used as the site boundary, or a pasted polygon; structured evidence can be attached to ride beside the site's path-to-power ledger. The guarantees are the API's, restated where you act on them: everything is classed customer_reported, nothing alters a verdict, boundaries resolve when the site is added so removing evidence later never changes what a monitor re-screens, and receipts carry a boundary's sha256 and area, never the outline.

First US hosting-capacity entry. Feeder-level DER hosting capacity for Pepco (Washington DC/Maryland), Delmarva Power and Atlantic City Electric: 1,558 feeders with allowable/maximum PV, large-generation hosting and the pending DER pipeline, in the operator's own kW, verbatim. US sites in the path-to-power ledger now show the strongest feeders within 5 km as generation-hosting context; it is never demand headroom, and no demand finding fires from it.

The path-to-power ledger. The candidate-site screen (methodology 1.9) now composes, per site, every candidate point of interconnection: connection-grade published capacity points within 30 km and EHV topology within 25 km, each assessed against the stated load target in the operator's own units, with queue pressure as competition context, a one-line conclusion, and the cheapest next fact that could change the verdict ("operator feasibility enquiry at the two unresolved candidates" vs "stop; the published record is exhausted here"). A composition of published evidence, never a power-flow result or connection offer, and it says so on every response.

Your studies, beside our record. Workspace sites can now bridge structured evidence objects (operator responses, consultant study extracts) into the path-to-power ledger (methodology 1.10): resolved at item write time so run snapshots stay immutable, served as customer_reported and unverified rows beside the published record. They never change assessments, findings or verdicts; when supplied material already answers the recommended enquiry, the ledger says so instead of recommending you buy the same fact twice.

Land-control joins widened. Overseas-proprietor (OCOD) matching now runs punctuation-normalised across all four proprietor slots with GB legal-form variants, alongside yesterday's four-slot CCOD normalisation.

Who holds the ground: corporate land control productised. The diligence screen's site-control check (methodology 1.5) now counts HMLR titles across all four proprietor slots (co-owned titles included), reports tenure and locality mix with stated price-paid where present, and rolls up landholdings across the counterparty's corporate group; land held in SPVs surfaces against the parent, with a material finding when the counterparty itself holds nothing directly. Entity profiles gained a land-control section: England & Wales titles by group company. Witness: one European utility's UK group resolves to 10,500+ titles across 11 companies. Counts, never areas or valuations; raw registers stay policy-held.

netztransparenz cleared. WebAPI terms confirmed; citation "Quelle - netztransparenz.de" carried on /sources, /licensing and on every row. German redispatch evidence now serves in commercial-safe mode.

A site is no longer just a coordinate. The candidate-site screen (methodology 1.8) accepts a GeoJSON polygon or multipolygon boundary per site; a multipolygon is a land assembly. Distances and radius checks measure from the asserted boundary edge; containment resolves at a derived representative point; flood evidence becomes an area intersection: the share of the site in EA Zones 2 and 3 and the net area outside Zone 3, at screening grain. Boundaries are customer-asserted and verified against nothing (not cadastre, not title, not planning); invalid geometry is rejected, never repaired; assemblies are capped at 100 km2. Receipts stay clean and private: the outline is never stored or echoed; the receipt carries its sha256, area and derived point, so it remains verifiable against the original file without disclosing it.

Customer evidence ingestion. The workspace gained a tenant-private evidence store: customer polygons and GIS layers (geometry), utility correspondence, connection studies and consultant reports (documents), and structured data (JSON), stored verbatim under POST /v2/workspace/evidence with per-organisation quota and content dedupe. Everything is classed customer_reported by construction; nothing here enters the public graph or alters a verdict. A geometry object becomes a site boundary by referencing it as boundary_evidence_id on an investigation item; resolution happens at write time, so run snapshots stay self-contained and deleting evidence later can never silently change what a monitor re-screens. Documents are stored and returned, never parsed into checks; extraction is a separate, future methodology decision.

The live shortlist, on the front page. The homepage now states the standing capability directly under the two-of-five proof: save a shortlist as an investigation, monitor it daily, weekly or monthly, and every verdict change arrives with the driving check named and the old and new receipts chained. The pilot page states the same thing: a pilot ends with a live workspace, not a one-off report. The capability shipped 15 July; today it became visible without signing in.

Validation register made consistent. Studies 10–12 were announced here on 15 July but /validation still carried nine studies and the site counters said nine. The three studies are now published in full (method, results, misses: the JAO PTDF agreement, the German redispatch corridor test with its failed branch-grain index, and the GB constraint-energy test including the disclosed load-weight defect), and every study counter on the site is generated from the validation register itself.

Public registers separated from internal operations notes. The licensing matrix carried internal annotations (named staff, fetch infrastructure, internal file paths) in several source notes. Public notes now state the procurement-relevant fact (cleared and when, or review pending); operational detail moved to an internal ledger. No rights status changed. Also fixed: a stale worked-example limitation claimed GB published headroom covered one operator; the current coverage (five GB distribution operators, plus IE, PT and FR registers) is now stated, and the homepage now distinguishes zone-grain queue history (US-only today) from the candidate-site screen's radius-grain use of GB and Irish connection registers.

Investigations: saved shortlists that watch the graph. The screens gained persistence: POST /v2/workspace/investigations saves a candidate shortlist (sites or counterparties) as a named investigation; each run re-screens it against the current graph, mints a receipt chained to the previous run (supersedes, visible on every public receipt), and returns a deterministic evidence diff: which verdicts changed, which check drove the change, which evidence appeared or vanished. Verdicts and evidence, never a score; attribution abstains when more than one check could explain a flip. Run snapshots are immutable; editing a shortlist never rewrites history. Tenant-private by construction: org-scoped rows in a schema the public catalogue cannot see. Keyed access; runs bill exactly like the underlying screen. Alongside it: receipt exports gained ?compare= (the workbook adds a Changes sheet diffing two frozen receipts); GET /v2/workspace/changes serves the organization's verdict-change feed; and annotations let a customer confirm, challenge or contextualise any finding; customer assertions are stored verbatim, classed customer_reported, and never alter a verdict. Monitors complete the loop: put an investigation on a daily, weekly or monthly cadence and the scheduler re-screens it, chains the receipt, and posts any verdict change to the feed with the driving check named. Scheduled runs bill like manual runs; a paused monitor never fires.

The workspace app. Investigations now have a browser home: /app lists your organization's saved shortlists, shows every run with its verdicts, what changed and the receipt (with one-click XLSX and PDF export), runs a screen on demand, toggles monitoring, and carries the changes feed. Notes and challenges attach to any finding; they are private to your organization and never alter a verdict. Sign-in is by single-use link: no passwords stored, sessions are revocable, and access is provisioned per organization by the Kortex team; request access via the pilot form.

The site now leads with the decision. The homepage opens with the buyer question (which sites deserve a connection study), the strongest proof sits directly under it: two of five site verdicts changed when the public connection registers were joined, receipts linked. Platform statistics moved to a depth section lower on the page; the hero carries three proofs (named sources, published validation studies with misses, frozen receipts). Navigation regrouped into Applications · Insights · Platform · Evidence · Developers with one primary action, Screen a shortlist; every page remains reachable, insight pages each end in a next step. Nothing was removed; the hierarchy changed.

Corridor congestion, two countries. /loop-flows now carries the corridors operators pay to relieve: modelled hourly-2025 north→south transfer for Germany (latitude 51.2°N) and Great Britain (the B6 Scotland–England boundary), each validated against what the operators actually paid to manage congestion (German redispatch: Spearman +0.57; GB SO-flagged Balancing Mechanism constraint energy: +0.82, 100% top-decile hit). GET /v2/insights/congestion-corridor serves both with per-corridor validation receipts. Corridor grain only: branch, overload and N-1 grains failed validation and are deliberately not served.

GB Balancing Mechanism ground truth: the BMU register (~3,000 units with lead parties), 10.3M bid-offer acceptance segments with the system operator's constraint flag intact, and 11.2M settlement-derived accepted-energy rows (BOAV), 2025 to present. Also loaded under the house protocol: the German federal asset register (MaStR, 164k units ≥ 100 kW with Handelsregister identifiers) and German TSO redispatch measures (84k, 2021 to present).

Diligence screen, methodology 1.4. Two evidence blocks join the exception ledger: German registered asset base (operating/planned/retired capacity with the operator's Handelsregister court and number) and GB balancing-market presence (BM units, SO-flagged acceptance history and settlement-derived constraint energy in GWh). Honest zeros at every name join, stated as such.

Validation studies 10–12 published. Raw model sensitivities vs the official JAO flow-based market-coupling PTDFs (median r +0.99 across 11 borders); modelled German congestion vs observed TSO redispatch (corridor grain passes, every branch-grain index fails and stays unserved); GB corridor vs SO-flagged constraint energy (+0.82, including a load-weight defect found by a physics-level check, fixed and disclosed).

Elexon licence cleared. BMRS open data licence confirmed; attribution "Contains BMRS data © Elexon Limited copyright and database right 2026" carried on /sources, /licensing and on every row. netztransparenz and JAO remain validation-only: published statistics are facts about the Kortex model, no third-party rows are served.

Navigation on phones. The menu groups were desktop hover dropdowns, which touch devices cannot open; small screens now get a proper menu button and full-screen panel carrying every destination, built from the same structure as the desktop menus so the two cannot drift apart. A full-body pass followed: the worked-example exception ledgers and the coverage register compressed to unreadable columns on narrow screens and now swipe at readable width; every page verified free of horizontal overflow at phone width. And one footer everywhere: previously only two pages carried the site footer while eight showed a statistics bar floating over the content on phones; every page now ends with the same footer, and the statistics bar sits in the page flow on small screens instead of covering it.

Copy accuracy. Candidate-site screens return evidence-based verdicts, not scores; the landing page said "scored" in two places and now says what the product does. Em dashes removed from page copy across 11 templates (house style).

Counterparty diligence screen live. POST /v2/insights/diligence/entities: entities in (LEI, Companies House number or exact name), exception ledger out with four verdicts (proceed / investigate / reject / insufficient evidence), hard gates for sanctions and dead registries, abstention over guessing. Worked example two published with its frozen receipt, plus a ledger export.

Loop flows: which borders physically carry any CWE zone-to-zone exchange, from a calibrated DC power-flow model (median out-of-sample r +0.88 vs observed 2025 flows; median r +0.99 vs the official JAO market PTDFs). First metric in the modelled evidence class, and the register's red badge to match.

NESO TEC register under twice-weekly cron with the transmission-queue evidence block in the diligence screen; HMLR CCOD/OCOD corporate land titles (4.5M) behind the site-control check; host-side graph analytics shipped: ownership-network centrality (341k entities), measured weather-sensitivity edges, grid criticality (28.8k articulation substations).

Contact and evaluation keys. A short form read directly by the Kortex team; evaluation API keys were approved by hand at launch.

The governed public data estate. Five registry-driven surfaces now expose the whole servable estate; governed tables (/v2/tables), time series, country indicators, asset profiles and entity profiles; with source, licence and commercial status resolved live on every response. The graph is no longer only behind pre-built screens.

New pages: Coverage (fitness-by-domain register, live inventory), Methodology (one entry per material metric: formula, evidence class, failure conditions, known misses), Who controls what (tiered ownership map with evidence-kind edges and FERC EQR wholesale counterparties), Demand signals (counted, never forecast), Insights index with maturity and evidence badges, and year selectors on Grid day + Clean hours (2025 data live).

Tier-1/2 ingest waves closed: EPA CEMS observed capacity factors, USWTDB/USPVDB, SEC XBRL + ISIN-LEI, hosting capacity, FEMA NRI, GEBCO, direct ISO queues, HydroRIVERS, sanctions primaries now incl. the EU consolidated list (49k entities), USGS MCS 2026, Eurostat NUTS-3, EU ETS compliance, GHSL population (WorldPop cross-checked), OE-417 disturbances, JRC surface water at 41k dams, OWID/IEA EV adoption, FERC EQR wholesale aggregates (as-filed unit errors flagged, never deleted).

Validation studies 8 and 9 published. US cooling-water links vs utilities' own EIA-860 filings; 35% of judgeable links contradicted and quarantined. Modelled carbon exposure vs the EU ETS compliance record; 26/27 countries in the expected band; the published miss (Finland, 1.93×) exposed that generic capacity factors overestimate reserve-heavy fleets.

Default-deny API scopes, commercial_safe execution mode (fails closed with the blocking sources listed), retrievable answer receipts (GET /v2/receipts/{id}), and the full TeleGeography cable map served keyed-only pending rights.

Four new dataset layers on the Global Viewer. Water infrastructure (1.4M reservoirs, wells, towers, treatment and pumping assets) and telecom sites (729k masts + 26k carrier/interconnection buildings) are free to explore. US interconnection queues (25.9k geocoded projects with status, MW and study stage; locations are county centroids, disclosed per row) and engineering-grade grid detail (559k transformers, junctions, switches, compensators) are included with an API key. All four serve from cleared sources (ODbL, CC BY 4.0, CC BY-SA 4.0).

Every map layer now hands off to the Workbench; a wb chip on each layer opens a runnable starter Cypher query over the same data, so a visual finding becomes a queryable workflow in one click. Asset drawers gained twelve more traversable node types (water, grid-detail and telecom labels).

Viewer load time cut. Anonymous layer responses are now micro-cached for five minutes (the data refreshes hourly at most), spatial queries serve concurrently instead of queueing behind the heaviest one, and sessions on slow connections or low-memory devices automatically fetch lighter payloads.

Data centres joined the graph as a first-class label; 5,888 facilities (PeeringDB CC BY-SA + curated hyperscaler campuses, geocode precision disclosed per row) with nearest-substation links at site precision, on the map and in the API.

Viewer: basemap style modes (dark / light / colour), real-sun solar lighting toggle, and shareable layer-state URLs. Clicking any asset opens a full-detail drawer with provenance and, for keyed users, one-hop connection analytics that deep-link into the workbench.

Data rights enforced on all public surfaces. Live vessel (AIS) and aircraft (ADS-B) layers on the Global Viewer are now key-gated pending a provider-rights audit. The nuclear-reactor, cable-landing and volcano map layers were withdrawn until their sources clear review. /licensing now opens with a binary production-safe count and states each enforcement action in the source notes.

Corporate-identity audit closed: all 116 open findings resolved with primary-source evidence (EIA-860 plant records, GLEIF, operator publications). Where an earlier verdict was wrong, the new evidence superseded it and the old edge was quarantined with its reason; corrections carry provenance; nothing is silently deleted.

Divestiture watch is live. A weekly screen looks for group-vs-operator contradictions and GLEIF drift across the corporate layer. Its first baseline run caught the Exelon → Constellation nuclear-fleet spin-off and two wrong-identity stitches.

GLEIF golden-copy refresh applied 102,306 property updates including 16,458 legal-name changes, refreshed 5,574 parent relationships, and added 90,300 corporate entities. Runs monthly. Non-Latin entity names gained GLEIF transliterations so search finds them in either script.

Global Viewer shipped at /globe; a WebGL globe over the spatial API: generators, transmission, pipelines, telecoms, water, minerals, conflict and market-zone surfaces, with live layers for keyed users. Free to explore without an account.

Workbench shipped at /workbench; a Cypher studio with a schema-aware sidebar, an interactive graph explorer, and CSV/JSON export, sharing the same key as the Globe.

Lifecycle vocabulary canonicalized at ingest. Generator lifecycle statuses now map to a documented 12-value vocabulary as data is written, with the source's raw value preserved alongside. All 28 example queries on the schema page were live-tested against the running graph.

Identity join upgraded to capacity-weighted majority (methodology v1.1). Every ownership flip the change produced was reviewed before shipping; one regression was caught and fixed pre-release.

Provenance envelope on every insight response: as-of timestamp, the sources touched with their licences, methodology, confidence and stated limitations. The answer and its receipt travel together.

Licence matrix published at /licensing. Every source in the register carries an explicit commercial status; unknowns render as "under review", never as an implied "cleared".

Validation programme published at /validation; six studies against independent references: GLEIF resolution, LBNL queue reproduction, SEC XBRL revenue, EIA-860 capacity (with honest coverage gaps: US wind 16%, solar 22%), spatial containment across 1.54M claims, and Exhibit 21 subsidiary chains. Misses are published alongside passes.

Procurement trust page at /trust, including an explicit no-SLA / no-SOC2-yet block; claims a buyer can check, not marketing.

Public demo scope: the three homepage proof queries now run for anonymous visitors, rate-limited and metered, so the flagship answers are checkable without signing up.

Reactor coordinates repaired. A scraper fallback had given 109 reactors bogus clustered coordinates; 163 of 164 were re-sourced from independent references with country-level verification, and the containment study now validates the row at 100%.

Corporate stitch audit: 82 operators adjudicated impact-first; about 28 GW of capacity remapped to the right owners, ~899 edges verified, ~782 quarantined with reasons.

API elevation programme completed: keyset cursor pagination on unbounded collections, uniform sort and filtering, typed graph traversals, an MCP server exposing five tools under the same billing and scopes, and a live pricing manifest at /.well-known/kortex.json.