SSI Index · Strategic Brief · SB-02
The Verification Standard for Asset-Level Adaptation Intelligence
The SSI Index audit and compliance methodology — its instruments, its registers, the command by which a reader re-derives each one, and the exceptions they recorded in 2026.
“A rule with no check is not a rule.”
The organising principle of the estate’s drafting doctrine, 24 August 2026
Executive summary
The methodology, in eight sentences
The SSI Index operates eight controls across its data, its doctrine and its publications. A conformance register compares every declaration doctrine makes against what the deployment observably does, and carried 116 rows at the period close — 47 conforming, 69 diverging, each diverging row owned. A figures register holds every published number with what it counts and when it was true, and governs which claim type may carry which figure. An epistemic tier register classes every quantitative claim as executable, measured, observed or modelled, and holds that only the executable tier establishes fact. An amendment procedure requires that no doctrine changes without a pinned decision, and that no draft is applied before it is pinned. A publication gate carries eighteen named checks and an execution rule that a gate is ticked only by the person who ran it, with the date and the output. A reference-integrity instrument checks every reference in the estate against two registries in about twelve minutes, and publishes its own measured false-positive rate. A source register holds a declared source for every metric in every jurisdiction. A mosaic rule states what licenses combining sources that disagree on granularity, period or jurisdiction, and requires every crossing to declare itself on the record.
Each control is re-derived from an artefact or a command stated in Annex A. The exceptions each recorded in the period are reported in Section 4 at control, class, count and owner. The counts are stated whether or not they flatter the estate, because a register reporting no exceptions describes an unexercised control, not a clean environment.
Section 5 reduces those controls to six questions a reader can put to any asset-level adaptation claim, each answerable from published material alone, each with a reading attached. That section is the part of this document intended for use on work other than ours.
Introduction
An asset-level claim is now a regulated object, and there is no published standard against which to test one
AN ASSERTION ABOUT the resilience of a single piece of infrastructure used to be a private matter between an owner and an engineer. It is no longer. Under the Critical Entities Resilience Directive, substation operators are critical entities and their preparedness is a matter of public record.1 Under Article 11 of the Taxonomy Regulation, a claim that an investment makes a substantial contribution to climate adaptation is a claim with financial consequences and a disclosure obligation behind it.2 Under Article 5 of the European Climate Law, continuous progress in adaptive capacity is a thing Member States report on.3
Three distinct readers — a regulator, an investor, an evaluator — are therefore now obliged to decide whether a particular asset-level adaptation claim can be relied upon. None of them holds the operator model the claim was built from. Each of them is in the position of being asked to accept a number.
Their position is structurally awkward and worth stating precisely, because it is the condition this methodology is built for. The detailed inventory that governs how a substation behaves under stress — what equipment it holds, how that equipment is secured, what it serves — is held by the network operator and is rarely published. A body with a mandate to ask whether a fleet is prepared therefore holds no more of that inventory than the public does. Assessments of this kind are consequently not independently checkable in the ordinary way: they rest on data supplied by the organisation whose preparedness is being assessed, or on open substitutes whose limitations are rarely characterised.
That leaves verification as the only available lever. If a reader cannot re-obtain the inputs, the next best thing is to be able to interrogate the apparatus: what it declares, where it diverges from its own declaration, how much of its construct had data behind it for the asset in question, and what it has measured itself against. Each of those is answerable from published material if, and only if, the publisher has built the instruments to answer them and is willing to publish what the instruments return.
What is missing is not analysis. It is a standard: a published statement of what an index of this kind checks, against what declaration, at what cadence, who owns an answer when a check fails, and what the checks returned in a given period. Without one, the only available posture toward an asset-level claim is deference or refusal, and both are expensive.
The index this methodology governs is an open one. Thirty-nine jurisdictions, 622,104 substations carried at asset level, each with a resilience score, a band, and two coverage values stating how much of the declared construct stood behind that particular number — published in full, version-controlled, under a licence that permits reuse, and assembled from public sources throughout. No operator model, no licensed inventory, no proprietary feed. An estate built entirely from open sources can be checked by anyone, which is precisely why it is worth building the checks for.
This brief publishes that standard, as operated. It is not a description of good intentions: every instrument in it carries the command by which a reader re-derives its output, and every exception those instruments recorded in the period is reported with an owner against it. The estate’s own position is the worked example, which is the only honest way to publish a verification standard — an instrument nobody has run against live data is a specification, not a standard.
The document is ordered accordingly. Section 1 states the scale and architecture of the apparatus the methodology governs, because a control environment can only be judged against the size of what it controls. Section 2 sets out the controls. Section 3 works three of them through at the detail a reader would need to run them. Section 4 reports what they processed in the period. Section 5 reduces the whole to six questions usable on any asset-level claim, and Section 6 applies the same controls to work not yet done.
Key takeaways
Seven statements, five carrying a tier
- Thirty-nine national fleets and approximately 4.5 million kilometres of line. 622,104 records carry a score; the count re-derives by direct census, sums to the 39 published manifests exactly, and every movement from the previous census reconciles to a dated decision bar one named residual. Measured
- An acquisition layer that is heterogeneous by necessity. Thirty-seven jurisdictions source their substation population from the open map; one federates four national and provincial registers; one takes the regulator’s own grid-installation register. Measured
- Six weighted components, twenty metrics, seventeen operative modifiers, one master equation. Every metric carries a published provenance tier, and only the top tier — a deployed value traced to a named external source at a frozen commit — is cited as verified anywhere in this methodology. T0
- Every control is evidenced by an artefact a reader can obtain, and two of them execute. The conformance register re-derives by a single command, the reference instrument by a three-script chain across the whole estate in about twelve minutes; the two measurement passes behind this brief’s own figures take one command each. The registers and procedures are evidenced by their own documents, which Annex A names beside the commands. T0
- No row is ever closed by editing the declared column. A divergence is recorded against the implementation, never against the declaration — the single rule that keeps a conformance register from becoming a record of adjusted expectations.
- Coverage and resolution travel with every record. Component-level and intra-weighted coverage are present on 622,104 of 622,104, so neither can be read alone; asset-granular claims hold inside six jurisdictions, 255,707 assets, and are stated as such. Measured
- The registers return exceptions, and the period’s are reported at control, class, count and owner. Two of the eighteen publication checks cannot be satisfied by the author alone, and that is where the period’s entries concentrate — which is what a check outside the author’s control is for.
Section 1
The apparatus
A control environment can only be judged against the size of what it controls. This is the size.
Seven layers, and the volume at each.
From acquisition to the assurance layer. Each layer adds a point at which a published figure has to be verified, which is what the instruments in Section 2 are placed against.
Measured T1
Re-derived 8 October 2026 from SSI_FOUNDATION_judgement.yaml, the ingestion census and a direct count across the deployed payload.
The count carries a movement trail
A cohort figure is only as good as the account of how it changed. The published count is censused per jurisdiction, and the movement between any two censuses reconciles jurisdiction by jurisdiction against dated decisions.
The published count is a measured quantity with a movement trail.
Between the census of 25 August and the count of 8 October, every record of the difference reconciles to a dated decision, bar one residual that is named and owned.
Measured T1
Re-derivable: SSI_JURISDICTION_MEASUREMENTS.json against a direct count across the deployed payload; deduplication figures from the figures register entry of 28 August 2026.
| Jurisdiction | 25 Aug | 8 Oct | movement | accounted for by |
|---|---|---|---|---|
| Germany | 168,776 | 108,016 | −60,760 | deduplication, 28 August |
| United States | 97,915 | 73,859 | −24,056 | deduplication, 28 August |
| France | 175,660 | 168,894 | −6,766 | deduplication, 28 August |
| Italy | 47,906 | 41,662 | −6,244 | deduplication, 28 August |
| Spain | 12,621 | 12,438 | −183 | deduplication, 28 August |
| Sweden | 1,192 | 3,774 | +2,582 | restore, 28 August |
| Japan | 7,073 | 6,168 | −905 | not reconciled — open, owner ingest |
| The remaining thirty-two jurisdictions did not move. | ||||
| total | 718,436 | 622,104 | −96,332 | 99.1% documented |
That table is the methodology in one object. The count did not simply fall; it fell by amounts that were decided, dated and recorded, in five jurisdictions, with one restore in a sixth. The residual is 905 records in a single jurisdiction, and rather than being absorbed into a rounding it is carried in the conformance register with an owner. A count without a movement trail is an assertion; a count whose every movement reconciles to a dated decision is a measurement.
One cohort figure rests on a different basis, and the brief says which it uses. An earlier cohort-scale figure of 796,121 is carried in the repository README and in the title of a submitted paper. Its per-country itemisation and the live register are not yet reconciled: the nine jurisdictions it names account for 98% of its total, leaving a residual 8.3× smaller than those remaining jurisdictions carry, and two of the nine itemise fewer records than are published for them — a shape a deduplication, which only subtracts, does not produce. The figures register carries both while that reconciliation is worked, and this brief takes the measured base, 622,104, throughout.
Three properties of that stack are worth drawing out before the instruments are described, because each one is a reason the instruments are necessary rather than ornamental.
The acquisition layer is heterogeneous by necessity
Thirty-seven of the thirty-nine jurisdictions take their substation population from OpenStreetMap through Overpass. Two do not: one federates four federal and provincial registers with a proximity dedupe, and one takes the national regulator’s own grid-installation register. Those two are also the only jurisdictions carrying positional accuracy or survey vintage at all.4
That heterogeneity is not a defect to be removed. It is what an estate assembled from open sources looks like, and the alternative — restricting the cohort to jurisdictions with a single uniform source — would reduce it to a handful. The consequence is that every field means slightly different things in different places, and an instrument that records what each field is declared to mean becomes a precondition of using the estate at all rather than a refinement of it.
The construct is large, and every part of it carries a published evidence tier
Six components carry the composite, weighted: continuity at 0.30, infrastructure exposure at 0.25, saturation at 0.20, voltage quality and economic at 0.10 each, transition at 0.05. Twenty metrics sit beneath them and seventeen operative modifiers act on the result, through a single master equation.5
A distinction this methodology exists to hold
Twenty metrics are declared and weighted. Fifteen carry a number on a record. Six carry provenance traced to a named external source at the pinned commit. Every metric carries a published provenance tier, recorded against it in the register, and the top tier is the only one this brief cites as verified.
Those are three different quantities, and the methodology reports them separately because a single count of twenty would not distinguish them. A metric below the top tier has a definition, an intra-weight and a published value; what the top tier adds is a traced chain from that value back to a named source at a frozen commit — a demanding test, and one most comparable indices do not set themselves. What completing the remaining tiers adds is not data volume — it is the right to cite as verified what is already computed.
The published surface states its own coverage on every record
Each of the 622,104 scored records carries a score, a band, and two coverage values — component-level and intra-weighted. The second is present on 622,104 of 622,104, which means the first cannot be read without it. A consumer of a single substation’s figure can therefore see how much of the declared construct stood behind that particular number without reading any accompanying document.
That surface is 2.35 GB across thirty-nine jurisdiction manifests and forty-two shards, the largest single file at 49.0 MB, and completing every declared metric takes it to 2.64 GB. The remaining work is therefore not a data-volume problem: what completion adds is not storage but the right to cite as verified what is already computed. Annex B states the surface in full.6
Thirty-nine national fleets, censused; the count sums to the published cohort exactly.
Scored records per jurisdiction on a log scale. The two fleets marked in cayenne take their substation population from a national register rather than the open map.
Measured T1
Re-derivable: direct count across the deployed payload, 8 October 2026. Sums to 622,104.
Volume at this scale is easy to overstate, and the apparatus narrows sharply at two of its steps: a climate variable resolved at a quarter of a degree gives 25,165 distinct cells across 622,104 assets, and a simulation run keeps 3,732,624 values from almost 112 billion draws. The index publishes the narrowed figure at each step and states the gross beside it, and each ratio belongs to its own step rather than to a product of them. Annex B tabulates both, with the granularity at which a regulator’s own figure resolves beside them.
The distribution is steep and the steepness matters for how any cohort-level statement should be read. The ten largest fleets carry 85.7% of the scored cohort; the smallest nine carry under one per cent between them. An unweighted average across thirty-nine jurisdictions is therefore a different object from a record-weighted one, and this methodology states which it is using wherever it reports a cohort figure.
Section 2
The methodology
Eight controls. What each checks, against what, how often, who owns an answer when a check does not clear, and how a reader re-derives it.
Eight controls stand between what doctrine declares and what the estate publishes — four registers, two procedures, one instrument that runs, and a rule that governs how sources are assembled. Each compares a published artefact against the declaration it is supposed to implement, and each returns divergences to the declaration as register rows with owners — never as revisions to the declaration itself. Annex A tabulates all eight with the artefact or command that re-derives each, together with the two measurement passes this brief’s own figures come from.
The conformance register
The conformance register compares what doctrine declares against what the deployment observably does, element by element and aspect by aspect. A row is opened when the two disagree; it carries the declared value, the observed value, a severity and an owner drawn from five classes — engine, doctrine, pipeline, ingest, site. It is re-derived on a standing watch, and each watch publishes a dated diff recording rows opened, rows closed and rows that moved severity or owner.7
The register grades what it finds on four severities, which are not interchangeable, and it publishes the label the instrument emits rather than a gentler one. BLOCKING is the strictest: doctrine and the deployment disagree about a published value. MATERIAL means a declared element has no data or no coordinate behind it. DISCLOSED means a known divergence with a stated reason. DEFERRED means a declared position awaiting an agenda item. The instrument exits non-zero while any row stands at the strictest class, by design. A severity that halts the estate’s own build is unusual for an index to carry at all, and carrying it is what makes the other three worth reading.
The distinction between the first two carries most of the weight. A row at the strictest class marks a published value the construct does not yet account for, and the rows are enumerated so a reader can see precisely which values are affected. A MATERIAL row marks a declared element for which the estate does not yet hold data or a coordinate — a disclosed absence rather than a contradiction, and one the published product already carries in its coverage fields. The two are kept apart because conflating them would let a contradiction be reported with the same weight as a disclosed absence.
Ownership is assigned to the layer that can close the row, not to the layer where it surfaced. A row owned by engine requires a change to the scoring code; pipeline, to how data moves and is written; ingest, to what is acquired; site, to what is rendered; and doctrine, to what is declared. The last class is the one that keeps the register honest: a row owned by doctrine is a row where the declaration itself is the thing in question, and routing it there rather than quietly amending the declaration is how the no-edit rule is enforced in practice.8
The rule that makes the register an instrument rather than a record
No row is ever resolved by editing the declared column. A divergence is a defect against the implementation. A register that permits the declaration to be amended toward the observation records only what the deployment already does, which is not a check — it is a transcript. Where a declaration is genuinely wrong, it changes by the amendment procedure below, with a dated decision, and the register row closes as a consequence of that decision rather than in place of it.
The figures register
Every number the estate publishes is entered in a register before it appears in an artefact, with two properties recorded: what it counts, and when it was true. The register then governs which claim type may carry which figure. Claims about the estate’s reach take the coverage figure; claims about what was scored, ranked or classified take the treated figure; a published paper’s own figure is never altered when that paper is quoted or its title cited.9
The discipline exists because an estate that grows produces a sequence of true numbers for the same object, and a reader comparing two artefacts a month apart will otherwise read a contradiction where there is a chronology. Smaller figures in earlier artefacts are not errors; they are the estate as it stood when that artefact was made.
Entry is the rule the register carries today, and the estate has already named the one that follows it. A figure is correct when it is entered, and the artefact carrying it may stay unsigned for days while the estate moves underneath it. The estate states the consequence itself: a figure in a document the flag officer signs should be re-measured at signing time, not quoted from a decision document three days old.10 Measured on 8 October, a decision document’s cost table did not reproduce at three days — and the finding records that the recommendation it carried survived, and only the numbers a subsequent amendment could cite changed.
A sixth rule is therefore named here and not declared: a figure carried in an unsigned or pending artefact is re-derived before that artefact is acted on. It is this brief’s opening principle turned on the register that governs its own numbers. It is proposed rather than pinned, because the amendment procedure forbids declaring a rule in the document that introduces it.
The epistemic tier register
Every quantitative claim carries one of four classes. Tier 0 is executable reality — a figure a reader obtains by running a stated command against a stated artefact. T1, Measured is read directly from a primary source or from the deployed estate. T1.5, Observed is an observed physical or administrative fact. T2, Modelled is methodology, projection or plan. The binding rule is that only Tier 0 establishes fact; everything above it is a claim about the world that carries its own provenance and may be wrong without the apparatus being wrong.
The tier appears as a badge against the claim in the artefact itself rather than in a methodological appendix, so that a reader skimming a figure sees its class at the same moment they see its value.
The amendment procedure
Doctrine changes by a drafted amendment that is explicitly not applied until a decision pins it. An amendment draft states, in its own header, that nothing in it has been written into doctrine held. The effect is that at any moment the estate can distinguish three states cleanly: what is declared, what is proposed, and what is deployed.11
The procedure is what makes the conformance register’s no-edit rule operable. Without a route by which a wrong declaration can be corrected on the record, the pressure to quietly adjust it would be constant.
The publication gate
Eighteen named checks stand between a manuscript and a submission, covering specification, scope, language, external verification, references, form and evidence integrity. The eighteenth, added at the period close, requires a quantified comparison of the work’s own output against a named incumbent method’s published output.
The execution rule is the operative part: a gate is ticked only by the person who ran the check, with the date and the output; an unticked or self-asserted gate blocks submission. The checks are recorded in a ledger in the manuscript’s own folder, written as they are run.
The source register
The construct declares a source for every metric in every jurisdiction. At the end point that is 780 source declarations — thirty-nine jurisdictions by twenty metrics — carrying 12,442,080 published values between them, a mean of 15,951 each. The mean is not what any one declaration carries: a declaration is one jurisdiction by one metric, so it moves that jurisdiction’s whole record count. One wrong source declaration moves every published value of that metric in that jurisdiction — between 43 numbers and 168,894, a spread of nearly four thousand to one, which is why the register records the range beside the mean.
Against those 780 declarations the estate currently names 1,097 distinct upstream stores and records an acquisition state for every one of them: 177 carry a code path — 96 fetched unattended, 81 needing a credential — 54 carry a documented human step, and 866 are named and not yet retrieved.12 Naming a store the estate does not yet hold is what makes the distance countable at all; an index that declared only what it had fetched would report no distance to cross. The primary external stores are ERA5-Land, CERRA, CMIP6 and Copernicus CAMS for the climate layers, and thirty-nine national regulators and statistical offices for the rest. The distance between a store named and a store retrieved is itself a register entry rather than a plan, and it is why a source register is an instrument and not an appendix.
The mosaic rule
The hardest verification question an index of this kind faces is not whether a source is good. It is what licenses combining sources that do not agree — on granularity, on the period they measure, or on the jurisdiction they cover. No regulator publishes continuity of supply at the granularity this index claims. Across every publication located so far the finest unit is roughly sixty substations to a measured value and the coarsest is 6,751 — a hundredfold spread across what is published. Annex B states the narrower pair that the estate has so far documented, and the two are not the same quantity. An index that takes one source per jurisdiction and calls the result uniform has answered the question by not asking it.
Doctrine answers it with a named convention rather than a practice. Convention #22, pinned 28 September 2026, holds that the unit of assembly is a metric: one metric may be built from several sources, and building it may cross granularity, time and jurisdiction. A ruling of 5 October adds the boundary — a mosaic may not cross concept. That ruling is in force and applied; its convention text is drafted to land beside the first and has not yet landed, which is the amendment procedure above behaving as declared rather than an exception to it. A power-weighted interruption index and a customer-weighted one measure different quantities, and substituting one for the other is refused however convenient the granularity.
What the convention requires at the point it matters
Every crossing declares itself on the record — not per jurisdiction, not on a methodology page. The declaration names what was crossed, the donor the value came from, the vintage the source measures rather than the period it is used for, and, for a granularity crossing, the allocator that disaggregated it, by name.
The reason is a verification reason. Without the per-record declaration a reader cannot distinguish a measured national figure from a cross-jurisdiction inference, and the source requirement becomes unenforceable exactly where it bears — on the published value. Assembly widens what may be built; it does not widen what may be asserted. Every source entering a mosaic satisfies the same source standard on its own terms, with any dissent between sources recorded. A mosaic of unverified parts is not a metric.
The combination rule is a ladder rather than an average. Sources enter at whatever granularity they have — national figure, administrative unit, per-operator return, the operator’s own publication — above a definitional layer that is never itself a value and is read first, because two numbers that count different things cannot constrain each other. The finest admissible layer sets the value; every coarser layer is retained as a constraint on it rather than discarded. A number with one provenance is believed or not; a number that reproduces an independently published second is checked. The check is the deliverable, not the number.
No mosaic-assembled value is published at the pinned commit. The metrics the rule is written for sit below the top provenance tier with no source yet declared, so the per-record declaration has nothing yet to declare, and a sweep of the published payload finds no such declaration on any record — which is the conformant state, not an omission. Two things follow. The method was declared before the acquisition rather than after it, which is the order that makes it a standard; and the conformance register carries no row testing Convention #22 — a binding convention the register does not yet test is a declaration without an automated check, and the row it implies is a per-record one: wherever a metric’s value is assembled, the crossing, the donor, the vintage and the allocator are present or the row opens.13
A second route from convention to instrument stands in a different state, and the difference is the point. A parse exists that rebuilds the judgement file’s rules block from the conventions register and exits non-zero when the two have drifted; its prepared output sits beside the file it would replace, and nothing schedules it to run. That is not an oversight. Wiring it before the prepared file is committed would refuse every landing over a sidecar only the signing authority can apply, and both the reason and the sequencing are on the record. A check withheld for a stated reason is a deferral; a check withheld silently is a gap. The severity taxonomy above carries a class for the first, and this brief reports it as one.
The reference-integrity instrument
Every reference in the estate is checked against Crossref and DataCite — by DOI where one exists, by bibliographic search where none does. The instrument covers 874 references across 18 manuscripts and runs in about twelve minutes. It reports four buckets: clean match, no match, year mismatch, and wrong first author.
It also reports its own error rate. Of the no-match bucket, hand-checking a sample established 9 of 11 as false positives (~80%) — the correct reading of that bucket is the citation is too abbreviated to verify automatically, not the paper does not exist. Year mismatches are chiefly an artefact of registries holding later editions; wrong-first-author flags are substantially an artefact of registries matching a book review rather than the book.14
Why an instrument publishes its own false-positive rate
A verification instrument that reports only its findings invites the reader to treat every flag as a defect, which overstates the instrument and understates the work. One that publishes its measured error rate per bucket allows a reader to weight each flag correctly. A calibrated instrument reports its own uncertainty; an asserted one does not. The same principle governs the tier register above.
Section 3
The instruments in operation
Three of the eight, worked through at the detail a reader would need to run them.
3.1 · The conformance register, read
At the period close the register carried 116 rows: 47 conforming and 69 diverging, across four severities and five owner classes. Each diverging row carries what doctrine declares, what the deployment is observed to do, and the party who can close it. Fifteen stand at the strictest class and are under curation, with an owner against each — twelve pipeline, two engine, one site.15
Two properties of those rows belong to the instrument rather than to any particular period. Every row names a measurement with its population attached — reproduces on 0.0% of 14,287 sampled records rather than provenance unclear — so each has a closure condition a reader can test and nobody reads a sampled finding as a complete one. And every row has a party who can close it, none of whom is the drafter: several turn on a doctrine decision rather than a code change, being rows where the deployment states a normalisation doctrine has not yet parameterised.16 Routing a row to the party who can settle it, rather than to whoever found it, is what keeps a register from being read as a work queue.
The register’s examined surface is itself a measured quantity and is reported alongside the conformance count, because the ratio of conforming to evaluated is a function of what the instrument has been pointed at. Rows are opened wherever a declaration exists specific enough to test, so the constraint on the examined surface is the precision of doctrine rather than the convenience of the result.
3.2 · The figures register, worked
The register’s central discipline is that a figure is never a number alone: it is a number, what it counts, and the date on which it was true. 622,104 is the scored cohort today. 718,436 is the same quantity at the census of 25 August. Neither supersedes the other and neither is an error; they are one measurement at two dates, and §1 reconciles the movement between them to a single named residual.17
718,436 is the figure carried in documents filed while it was current, and the register holds it with that date rather than retiring it. The line dimension — approximately 4.5 million kilometres — is a route length, registered as measured at Tier 1 and carried under a rule of its own: it is claimed alongside the substation count and never as one. The register has already corrected it once, from a form that rendered a length as a count, which is the kind of error a figures register exists to catch before a reader does.
| Figure | What it counts | True as of | Carried by |
|---|---|---|---|
| 796,121 | an earlier cohort-scale figure, on a different basis | Jul 2026 | carried while the reconciliation is worked |
| 622,104 | treated cohort, after five deduplications | current | every claim about what was scored, ranked or classified |
| 718,436 | the scored cohort at the previous census | 25 Aug 2026 | the lineage baseline, and documents filed while it was current |
| 725,462 | cohort at an earlier submission round | 27 Jul 2026 | the titles of papers submitted then; never altered when quoted |
| ~4.5 M km | route length of power line | current | claimed alongside the substation count, never as one |
The register’s value is visible in the rows that are no longer current. None is deleted. A reader comparing a document filed in August with one published in October otherwise sees two cohort figures for one estate; the register converts that from an apparent contradiction into a chronology with dates attached, and §1 supplies the movement between them. Where a figure rests on a basis the chronology does not yet reconcile — as the July cohort-scale claim does — the register says so rather than carrying it as current.
The procedural rule matters more than any individual figure: any number appearing in a new artefact is entered in the register first, with what it counts and when it was true. An artefact that introduces an unregistered figure has bypassed the instrument, and the defect is in the artefact.
3.3 · The publication gate
Eighteen checks, and the two a manuscript cannot pass by assertion.
The publication gate as a sequence. G7 requires a reader outside the work; G18 requires another method's published output to measure against. Neither is satisfiable by the author alone, which is what makes them load-bearing.
Schematic
Sixteen of the eighteen checks are satisfiable by the author with sufficient care: a specification page that predates drafting and names one claim and one reader; a transfer-function chain in which every link is held as data or cited to literature; a term register with no unexplained coinage; notation complete in both directions; references verified and read; an abstract within length; arithmetic closure on every table; figures regenerated with the commands recorded.
Two are not. G7 requires an external read by a person who has not seen a draft, and states explicitly that a machine does not satisfy it. G18 requires a quantified comparison against a named incumbent method’s published output, which requires that another method have published output on an overlapping population. Neither can be produced by the author working alone, and both therefore depend on something outside the work being obtained.
That is the design. A check an author can satisfy by trying harder measures diligence. A check an author cannot satisfy without an external object measures whether the external object exists — and it is the second kind that catches what diligence does not. The period’s publication exceptions concentrate at exactly those two checks, which is the result the design predicts.
| Check | Requires | Evidenced by |
|---|---|---|
| G1 · specification | a specification page predating drafting, naming one claim and one reader | the page, with its date |
| G2 · scope | a transfer-function chain in which every link is held as data or cited | the chain, written out |
| G3 · terms | a complete term register; no unexplained coinage | the register |
| G4 · acronyms | nothing under four uses abbreviated | script output |
| G5 · notation | a notation list complete in both directions | the list |
| G6 · form | no field names in prose; no numbered sections | search output |
| G7 · verification | an external read by a person who has not seen a draft — a machine does not satisfy it | the reader’s returned form, named and dated |
| G8 · references resolve | the verification instrument exits clean | the run, with its date |
| G9 · references read | every reference read and logged | the log |
| G10 · citation form | author-date throughout; no preprints in the list | the list |
| G11 · abstract | within length, narrative prose, matching the submission field | both texts |
| G12 · applications | a practical-applications statement within length, no acronyms | the statement |
| G13 · length | a recorded double-spaced page count inside the venue ceiling | the count, recorded |
| G14 · provenance | a complete provenance table; every disagreement disclosed | the table |
| G15 · independence | no composite regressed on its own components | the analysis plan |
| G16 · arithmetic | every table closes | the check output |
| G17 · regeneration | all figures and counts regenerated, commands recorded | the commands |
| G18 · margin | a quantified comparison against a named incumbent method’s published output, on an overlapping population | the comparison, as numbers — or a recorded statement of why none can be run |
Where a comparison cannot be run — because no incumbent publishes output on an overlapping population — G18 is discharged by recording that fact in the submission memo and stating it in the work itself. A manuscript that says plainly why no margin is reported has satisfied the check; one that is silent has not.
The checks are recorded in a ledger held in the manuscript’s own folder, one row per check, each carrying who ran it, the date, and the output or the command. The ledger is written as the checks are run rather than assembled at the end, and the governing rule is that a submission with no ledger has no gates, whatever prose elsewhere in the folder asserts. That rule exists because the failure mode a gate list is most exposed to is not a check that fails but a check that is recorded as passed without having been run — and a ledger written contemporaneously, with an attributed runner and an output, is the only artefact that distinguishes the two after the fact.
3.4 · Three states, kept distinct
The instruments above only work because the estate can tell three things apart at any moment, and name which is which: what is declared, what is proposed, and what is deployed.
The conformance register measures the distance between the first and the third. The amendment procedure governs movement into the first. The deployment is whatever the published artefact actually does, which is a fact rather than a decision. Collapsing any two of the three is the characteristic failure of an internal quality process: a proposal read as a declaration produces a register that measures against something nobody agreed; a deployment read as a declaration produces a register that cannot find anything.
The procedural expression is small and strict. An amendment draft states in its own header that nothing in it has been written into doctrine held. A decision pins it, with a date. The register row then closes as a consequence of that decision. At no point does a row close because a declaration was quietly brought into line with an observation, and the diff history is what makes that auditable after the fact rather than merely asserted.
Section 4
Period findings
What the instruments processed, and what the processing produced.
Across the period the instruments processed the whole estate rather than a sample: 622,104 records per coverage pass in about twenty-three seconds, 874 references across eighteen manuscripts in about twelve minutes, 116 register rows on one command, and all thirty-nine fleets censused for join capability. Annex A names the command behind each.
4.1 · Coverage gained
Mean composite coverage stands at 0.4519 component-level and 0.3101 intra-weighted.
Coverage across all 622,104 scored records, 30 September to 8 October 2026. The three marked points are jurisdictions whose continuity component became linked to their own regulator's published record.
Measured T1
Re-derivable: ssi_tracker_state.py, single pass over the scored cohort, approximately 23 seconds.
That component is continuity of supply.18 Every jurisdiction carries a continuity value; what differs is the class of source behind it. At the period close, three jurisdictions covering 166,324 records carry a value linked to their own regulator’s published record — the strictest class available — and the remaining thirty-six carry a value derived from other sources. The class travels with each record rather than sitting in an aggregate note, so a reader of one substation’s figure can see which applies to it.
Two coverage figures are emitted from the same pass and published together by rule: a component-level mean and an intra-weighted mean. The intra-weighted figure is present on 622,104 of 622,104 records, so the component-level figure cannot be read without it. The gap between them — 0.1419 at the period close — is the quantity a single headline coverage number would conceal.19
A landing — the operation by which a jurisdiction’s component becomes linked to that jurisdiction’s own published record — is itself a controlled procedure, and the control is what the five elements of its record enforce.
| Element | Requirement | What it prevents |
|---|---|---|
| Source identity | the file hashed, with its byte size and sheet named | a later reader obtaining a different version of “the same” publication |
| Join coverage | the proportion of records the join reaches, stated as a number | a partial join reported as a complete one |
| Declared absences | records or sub-metrics the source does not cover, named and excluded | an absence filled by inference and reported as coverage |
| Method departures | any divergence from a prior landing’s method, with the reason | two jurisdictions computed differently without the difference being visible |
| Effect on the record | how many records changed band as a result | a landing reported as an improvement without its magnitude |
The British landing states all five most completely and is worth working through, because each element did work there that a shallower procedure would have skipped.
Source identity. The jurisdiction’s regulator plots the two quantities in two figures of its annual report and prints no table. A chart is not a source. The landing is taken instead from the supplementary datafile published alongside it, with the sheet named and the file hashed — so a reader checking this figure in two years obtains the same object or discovers that they cannot.20
Join coverage. The records carry no operator field at all, so the link is geographic: point-in-polygon against a frozen licence-area layer, itself hashed, carrying fourteen features. 59,130 of 59,744 records assigned, 98.97%, all fourteen areas reached.21
Declared absences and method departures. Six hundred and thirteen records in a jurisdiction the layer does not cover are declared absent rather than approximated from neighbouring areas; one record outside every polygon is recorded as such rather than assigned to the nearest; a sub-metric for which this regulator sets no equivalent ceiling is declared absent rather than carried at a default. A method departure is recorded alongside them: the preceding landing had weighted by record, which across fourteen areas would render two published regulated figures differing by a factor of 2.16 as the same number. Percentiles were taken over units instead, and the departure is recorded with that reason.
Effect on the record. 33,887 of 59,744 records changed resilience band, published with the landing rather than inferred later — a procedure that moves more than half a national fleet between bands is a material event in the published product and is recorded as one.
Declare absent rather than estimate
Across the period this rule is applied in three places where an inference-filled apparatus would have carried a number: 613 records outside a coverage layer; a voltage-quality sub-metric for which one regulator sets no equivalent ceiling; and an unplanned-frequency series one national regulator does not publish at all. In each case the intra-coverage figure for that jurisdiction is lower as a result, visibly, and the alternative would have been a higher number resting on nothing measured.
The per-jurisdiction continuity detail behind these figures — what each regulator publishes, in what form, and the share of the common definition each publication can fill — is summarised at the headline here and not restated per jurisdiction.
4.2 · Resolution, measured
Coverage and resolution are different properties and the distinction is load-bearing for every cohort claim the estate makes. Coverage is how much of the declared construct has data behind it. Resolution is the granularity at which that data distinguishes one asset from another.
A full census over the deployed payload — 39 jurisdictions, 622,104 records, no sampling — establishes the boundary. Six jurisdictions carry per-asset modifier resolution, covering 255,707 assets and 41.1% of the treated cohort. The remainder resolve regionally or from a generator. The census behind these figures is published with its full method in the companion resolution analysis, which this brief cites rather than repeats.22
Two further results from that census bear on how any index of this kind should be read. A distributive-justice modifier carries fewer than one distinct value per twenty substations in thirty jurisdictions, and a single value nationally in one. And the generator’s own dispersion sits inside the band of the genuinely per-asset fleets, while those same fleets’ sourced-but-regional records do not — so distributional plausibility is anti-diagnostic. A value’s spread looking right is not evidence that it was measured.
The inversion result shows why the distinction matters in use. Rank the national fleets by raw compound-hazard exposure, re-rank them after the anti-maladaptation screen, and the order changes by up to 27 ranks — which a reader might take as the screen doing its work. The census shows otherwise. The six jurisdictions carrying per-asset resolution move three ranks or fewer, every one of them, and every double-digit inversion belongs to a jurisdiction whose values resolve regionally or are filled. The movement is an artefact of resolution rather than a finding about preparedness. The commissioned question was answered in the negative and the negative was published.
Two subsidiary results close the loop: a voltage control across the per-asset jurisdictions rules out voltage class as the explanation, dispersion differing by under 0.002 between classes; and the four fleets publishing substations at exactly zero are the four carrying no generated records at all, a consistency check on the disclosure mechanism rather than an anomaly.
The operative consequence is carried in this methodology as a standing constraint: an asset-granular claim about the cohort holds inside the six per-asset jurisdictions and must be stated as such. A claim at 39-jurisdiction scale is a claim about coverage. The two are not interchangeable and this brief does not interchange them.
For a reader applying this standard elsewhere, the generalisable point is the anti-diagnostic one. The usual test applied to an unfamiliar dataset is whether its values are distributed plausibly. On this evidence that test cannot distinguish a measured series from a generated one, because the generator was built to produce a plausible distribution and succeeded. The test that does work is a count of distinct values per jurisdiction against the number of assets, which takes one query and is not reported by anybody as a matter of course.
4.3 · What the registers returned
The instruments returned entries across the period, reported here at control, class, count and owner — the level at which an assurance register reports. The conformance register carries 69 diverging rows across four severities; the publication gate recorded five entries, three of which name the same control; the reference instrument recorded two; and the figures register one.23
Two observations about the shape of that return, both of which are about the instruments rather than about any individual entry.
The entries concentrate where the design predicts. Sixteen of the eighteen publication checks are satisfiable by an author with sufficient care; two are not, because they require something outside the work — a reader who has not seen a draft, and another method’s published output to measure against. Four of the five publication entries fall at those two checks. A check an author can satisfy by trying harder measures diligence; a check requiring an external object measures whether the external object was obtained, and it is the second kind that catches what diligence does not.
The margin control was specified before it was gated. The requirement for a quantified comparison against a named alternative was written into the estate’s publishing doctrine on 26 August 2026, in terms: a contribution must carry “a quantified margin against those alternatives”.24 It existed as a rule and did not exist as a check. It is now the eighteenth gate, with a stated output, under the principle that opens this brief — a rule with no check is not a rule.
4.4 · Disclosure — how the index publishes what it cannot currently measure
The three patterns above share a property worth stating separately, because it is the one a reader of any index should test first: the gaps are carried in the published record itself, per record, rather than in a methodological note.
Three mechanisms do that work: coverage is a field rather than a footnote, resolution is disclosed per record so a regionally resolved value is distinguishable from one measured at the asset, and absences are declared rather than estimated — which lowers the published coverage figure visibly rather than raising it invisibly.
The third of those is the one that costs something. In the period it was applied to 613 records falling outside a coverage layer, to a sub-metric for which one regulator sets no equivalent ceiling, and to a frequency series one national regulator does not publish. Each application produces a lower number for that jurisdiction than an estimate would have, and each is recorded with the reason. An index that fills such gaps by inference reports higher coverage and carries an unbounded error; one that declares them reports lower coverage and carries a bounded one. The second is the only version that can be audited.
The same discipline governs negative results. Where a commissioned analysis returns an answer the estate did not want, the answer is published. A methodology that publishes only its confirmations has no way of demonstrating that it would have published anything else.
Section 5
Applying the standard to any asset-level claim
Six questions, each answerable from published material, each with a reading attached.
A standard that only its author can apply is a description. The eight controls above reduce to six questions a regulator, investor or evaluator can put to any asset-level adaptation claim — including this one — without access to anything the claimant keeps private.25
1 · What does the claim declare, and where is the declaration written?
A claim that cannot point to a published construct — the components, their weights, the normalisation applied to each — cannot be checked against anything. The question is not whether a methodology document exists but whether it is specific enough to generate a testable proposition. Reading: a declaration that cannot be falsified by any observation of the product is not a declaration.
2 · Where is the register of divergences between the declaration and the product?
Every deployed system diverges from its specification. The question is whether the divergences are enumerated, owned and dated. Reading: a published register with exceptions in it is stronger evidence of control than a published methodology with none, because the second is consistent with never having looked.
3 · How much of the declared construct has data behind it, per record?
A composite score is only as strong as the proportion of its definition that was measured for that particular asset. The question is whether coverage travels with the score as a field. Reading: where coverage is reported only as a cohort average, the average conceals its own distribution, and the assets a user most needs to treat carefully are precisely those below it.
4 · At what granularity does each value actually distinguish one asset from another?
Coverage and resolution are different properties and are routinely conflated. A value present on every record may be a national constant repeated. Reading: distributional plausibility is not evidence of measurement; a generated series can sit inside the dispersion band of a measured one. Ask for distinct-value counts per jurisdiction, not for a histogram.
5 · What is published about what the method cannot do?
Absences declared rather than estimated produce lower headline figures and bounded error; the reverse produces higher figures and unbounded error. Reading: a methodology with no published negative results has not demonstrated that it would publish one.
6 · Against what named alternative has the method been measured, and by how much?
A method that has never been run against another method’s published output on an overlapping population has no established margin. Reading: the absence of a margin is not evidence that a method is weak — but it is evidence that its strength is unestablished, and the two should not be read as the same thing. Where no overlapping published output exists, a method that says so plainly has answered the question; one that is silent has not.
Applied to this index
Questions 1 to 5 are answered in Sections 1 to 4, with the commands to re-derive each in Annex A. Question 6 is answered in two parts: for the index itself, the margin is reported per jurisdiction in the companion resolution analysis; for the published contributions, the gate ledger records three entries where no margin was carried, and the check that now requires one. A reader applying this standard to this index should arrive at those entries, because they are the ones we recorded ourselves.
Section 6
Forward application
The same controls, applied to work not yet done.
Each forward entry names the quantified comparison it carries before it is filed.
The publication programme to Q3 2027 under the eighteen-check gate. No entry is scheduled without a stated margin, and two preconditions are recorded as preconditions.
T2 · forward
Tier 2. This is a plan under a control, not a measurement. It carries no evidentiary weight until executed.
Two contributions are specified for the current quarter, and both are selected on the same criterion: the measurement behind each is already complete and Tier-0 re-derivable, so what remains is the writing and the gate rather than the research. A programme that schedules contributions whose evidence does not yet exist is scheduling intentions.
The register data descriptor. Its subject is the substrate itself — the jurisdictions, the records, the line length, the provenance per field, the licence and the reproduction pipeline. Its margin is reported against two named openly published grid datasets, on scale, jurisdictional breadth and per-asset attribute depth, and explicitly not on primacy, which this methodology forbids claiming without a documented search having been run.26 One structural property recommends it as the first entry: a data descriptor’s contribution is the data, so the class of mismatch in which a venue publishes instruments and receives an account of limits cannot arise.
The continuity comparability study. Its subject is what national regulators publish about continuity of supply, measured against one common definition, and the finding that the binding constraint on sub-national comparison is publication form rather than data availability. Its margin is structural rather than appended: the per-jurisdiction fillable share is the quantified comparison, against two named standing benchmark reports in the domain, on an overlapping population. It offers a reproducible comparability measure and the method for computing it, which is what makes the contribution an instrument rather than an observation.
The 2027 entries follow the same discipline. A cascade decision paper reports harm reduced or relocated against the alternative of doing nothing. The seismic screening work returns under the eighteenth gate: its open-register ranking run against a published operator-inventory system on the one region where that system’s output is public, with agreement and divergence reported as numbers. A measured divergence is a result; the entry is scheduled on that basis rather than on an expected direction.
One entry remains deferred as a publication, and the distinction matters. The maladaptation gate’s three conditions — that the substrate contribution exists, that the verification check has been externally satisfied, and that an academic co-author is on the byline — are stated, and the entry does not file until all three are satisfied; the gate is open on all three today. It is carried on the programme as deferred rather than removed, because a programme that silently drops what it cannot yet do is not a programme.
The instrument is not deferred. The gate enters beta in the fourth quarter of 2026 on the commercial side, through the SSI-ENN deployment surface. That trial does not satisfy any of the three publication conditions, and it is not offered as though it did — an operational beta is not an external read of a manuscript, and no figure the trial produces appears in this methodology or in any report of this class, in either direction. What it supplies is the margin. A gate running against live decisions generates the comparison the eighteenth check requires, which is why this entry can now name where its margin comes from rather than leaving the question open.
Where the instrument’s question comes from. A maladaptation gate needs a prior answer to what an index discloses by existing, and the estate measured it on a full census of all 622,104 records: how much of a high-consequence property does the published payload recover that country-level knowledge alone does not? On catchment population — the one candidate of three to survive its own circularity test, named as such because the other two did not — recovery rises from 0.151 to 0.538. Nothing in the index is secret, so the reading is not that it adds hidden information. It is that an index is not a tile in a mosaic — it is the assembled mosaic, and what it spares a reader is the linkage.
The finding that outranks the result constrains the measurement rather than the estate: the recovered property is itself a published field, so a reader looks it up instead of inferring it, and a gain figure means something only against a declared redaction. “With this field withheld, the index recovers 54% of it” is honest; the same figure without the redaction is the index computing what it computes. That is the quantity the gate is built to carry.27
Two preconditions, recorded as preconditions
A deposited register with a resolvable identifier. The data descriptor cannot be filed without one and the comparability study should cite one. It is a precondition on both contributions scheduled for the quarter.
An external verification read. The gate’s own rule states that a machine does not satisfy it. The route of first resort is the academic co-author this methodology requires on contributions of this class, who performs the read — which closes two conditions with one appointment. Where the reader cannot be obtained, the entry does not file.
Conclusions
The standard, offered for use
An index that asks to be relied upon in regulated decisions owes its readers four things: a statement of what it checks, a statement of what the checks returned, a command by which each is re-derived, and an owner against every answer it cannot yet give. This brief is that statement for 2026.
The instruments are published rather than described, and the commands in Annex A run against the published payload, so a reader can re-derive the register, the coverage and the count in this brief without anything the estate holds privately. The conformance register is itself an internal instrument: its returns are reported here at the level an assurance register reports rather than reproduced row by row. What transfers is the method — the six questions in Section 5 require nothing but a counterparty’s own published material, which is the property that makes this a standard rather than an assurance claim.
It is offered for use against any asset-level adaptation claim, the estate’s own included. A reader who runs these instruments against this index and finds something the registers do not carry has found a defect in the methodology.
Three properties of the design are worth restating as the brief closes, because they are the ones that would be hardest to retrofit into an index that had not been built with them.
The declaration is sovereign. Nothing in this apparatus resolves a divergence by moving the declaration toward the observation. Where a declaration is wrong it changes by a dated decision, on the record, and the register row closes in consequence. An index without that rule can report any level of conformance it likes, and the reader has no way to tell which it did.
The gaps travel with the score. Coverage and resolution are per-record fields rather than cohort statistics, so a consumer of one substation’s figure can see how much of the construct stood behind it and at what granularity, without reading anything else. An average conceals its own distribution, and the assets that most need caution are the ones below it.
The exceptions are reported, not filtered. Every instrument above returns findings about this estate, and the period’s are given at the level an assurance register reports — control, class, count, owner. A reader is free to weight them. What they should not have to do is wonder whether there were others, which is why the counts are stated even where the rows themselves are not.
The programme at Section 6 applies the same instruments to work not yet done. The two preconditions recorded there are recorded as preconditions, not as ambitions: where the external verification read cannot be obtained, the entry does not file. That is the only form of commitment a verification standard can honestly make about its own future output.
“No row is ever closed by editing the declared column.”
The conformance register’s governing rule
Annex A
The register of controls, with re-derivation
Every control re-derives from something a reader can obtain: two by running code, six from the document that governs them. The two measurement passes behind this brief’s own figures follow in Table A2.
| Control | Checks | Cadence | Re-derivation |
|---|---|---|---|
| Conformance register | doctrine declared against deployment observed, by element and aspect | standing watch, dated diff per run | python3 SSI_FOUNDATION_conformance.py <site repo> — one command, exits non-zero while any BLOCKING row is open |
| Reference integrity | every reference against Crossref and DataCite | pre-submission, and on sweep | extract3.py → verify.py → verify2.py — ~12 minutes, whole estate |
| Figures register | what each published number counts and when it was true | on every new artefact | the register document itself; entry precedes use |
| Publication gate | eighteen named checks, dated and attributed | per manuscript | the gate ledger in the manuscript folder |
| Epistemic tier register | the class of every quantitative claim — executable, measured, observed, modelled | per claim | the tier badge carried against the claim in the artefact itself |
| Amendment procedure | that no doctrine changes without a pinned decision, and no draft is applied before it is pinned | per amendment | the draft, its header and the dated decision that pins it |
| Source register | a declared source for every metric in every jurisdiction | per declaration | the register, with the stores named and those retrieved |
| Mosaic rule | that an assembled metric declares its crossing, donor, vintage and allocator on the record | per assembled value | Convention #22 in the conventions register, and the ruling of 5 October; no conformance row yet tests either |
| Pass | Emits | Cadence | Re-derivation |
|---|---|---|---|
| Coverage emitter | component-level and intra-weighted composite coverage, together | per scoring pass | ssi_tracker_state.py — single pass, ~23 s over 622,104 records |
| Join census | whether each fleet carries a key at the granularity a publication uses | on survey | scripts/ssi_survey_join_units.py — all 39 fleets, censused |
Scope and figures. 39 jurisdictions; scored cohort 622,104 records, 718,436 at the census of 25 August; approximately 4.5 million kilometres of line, registered as measured at Tier 1.
What the register’s figures depend on. The conformance command takes the published site repository as its argument while the instrument lives in the doctrine repository, so the register’s output is a function of two states and both are pinned here: site repo b2a97845, doctrine repo 282320b. The instrument was strengthened during the period — a revision that reinstated divergences the earlier build could not observe at all — and the 116 rows reported here are the strengthened instrument’s. That revision is prepared and awaiting its commit, so a reader re-running against committed doctrine alone sees the earlier, smaller return. The direction is what matters: every row the revision added is a divergence the register now catches and previously could not.
Tier discipline. Sections 1 to 5 and both annexes are Tier 0 or Tier 1. Section 6 is Tier 2 and is a plan under a control, not a result.
Annex B
The index in numbers
The scale the instruments run against, on one annex.
A control environment can only be judged against the size of what it controls. Every figure here is emitted by ssi_measure_data_quantum.py against the published payload on 8 October 2026, with the delivery split re-counted by file, except where this annex corrects that script and says so.
Three numbers set the scale. A full run draws 111,978,720,000 correlated Monte Carlo samples — 1.34 × 1012 across twelve refreshes a year. At the end point the estate carries 12,442,080 record × metric values, each resting on a source and a pin, re-pinned every run. And the whole of it is 2.35 GB of published, version-controlled data, rising to 2.64 GB when every declared metric carries a number.
| Layer | Quantity | Measured |
|---|---|---|
| The estate | Jurisdictions | 39 |
| Substation records | 622,104 | |
| Distinct field names · fields on a median record | 187 · 74 | |
| The model | Components, weighted · metrics declared | 6 · 20 |
| Metric-correlation constants, as a 20 × 20 Gaussian copula | 18 | |
| Resilience modifiers — 16 multiplicative, 1 additive | 17 | |
| Monte Carlo iterations per substation | 10,000 | |
| The sources | Upstream stores named · with a retrieval held | 1,097 · 231 |
| Source declarations at the end point — 39 × 20 | 780 | |
| Published values across the 780 declarations — mean 15,951, range 43 to 168,894 | 12,442,080 |
Each substation’s seed derives from its own inputs, so a re-run reproduces the published number exactly rather than approximately. Two metrics are deterministic and carry no sigma, which is why eighteen of the twenty enter the simulation.
Where the flow narrows
| Step | Gross | Independent | Ratio |
|---|---|---|---|
| One climate variable over the estate | 622,104 assets | 25,165 cells at 0.25° | 24.7 × |
| One Monte Carlo run | 111,978,720,000 draws | 3,732,624 kept | 30,000 × |
| A regulator’s continuity figure | 622,104 assets | Italy finest at 389 records per value; Spain coarsest admitted at 3,110 | — |
A size claim that only multiplies is not defensible, so the narrowed figure is the one the index publishes at each step. 99.1% of assets share a grid cell with another and the largest cell holds 2,855. Each ratio belongs to its own step and they are not combined — their product would describe no claim anyone makes.
What completing the index costs, and what it delivers
Fifteen of twenty metrics carry a number on a record and six stand at the top provenance tier; 4,901,644 metric values are stored against an end point of 12,442,080, or 39.4%. The metrics block is 8.0% of the record bytes, so completing every declared metric adds 0.29 GB — a factor of 1.12. The remaining work is not a data-volume problem; what it adds is the right to cite as verified what is already computed.
| Measured | |
|---|---|
| Jurisdiction manifests · record shards · legacy root file | 39 · 42 · 1 |
| Published data files, total | 82 |
| Live cohort bytes · with the legacy file | 2.35 GB · 2.37 GB |
| Largest single file — france/ssi-data-substations-11.json | 49.0 MB |
| France shards on one page view · compression over the wire | 13 · 9.1 × |
Where two instruments meet on one file
The estate publishes 82 data files, and two counting rules divide them differently. The measurement script takes one manifest per jurisdiction and assigns every other data file to the shard bucket: 39 and 43. Counted by file, the division is 39 jurisdiction manifests, 42 record shards and one further data file at the repository root — carried at version 4.0.2 where the deployed cohort is 4.2, under a source note describing a jurisdiction other than the one whose data it holds.
Both of those properties are already open in the conformance register, as two of the fifteen rows under curation at its strictest class. The delivery count and the conformance register therefore reach the same object from opposite directions — one counts what is served, the other tests what is declared. A reader holding nothing but the published file list arrives at the two rows the register reaches with doctrine in hand. The split is corrected here and the correction recorded, which is the only reason a reader can see it at all.
Every published byte is a committed byte, because the site is the repository — which is what makes the commands in Annex A runnable by a reader.
Every figure here names its origin, and two come from outside the measurement script. The ~4.5 million kilometres of line is a route length, registered as measured at Tier 1 and governed by a rule of its own: claimed alongside the substation count, never as one. It is not re-derived by the measurement script because route length is not a quantity the published payload carries — a property of the script’s scope rather than of the figure’s standing. And 796,121 is an earlier cohort-scale figure, carried in the repository README and in a submitted paper title, resting on a different counting basis from the live register: the two differ per country in both directions, so a single deduplication does not reconcile them. The figures register carries both while that reconciliation is worked, and this brief takes the measured base, 622,104, throughout.
- Directive (EU) 2022/2557 on the resilience of critical entities. Electricity transmission and distribution operators fall within its scope.
- Regulation (EU) 2020/852, Article 11 — substantial contribution to climate change adaptation.
- Regulation (EU) 2021/1119 (European Climate Law), Article 5 — continuous progress in adaptive capacity, strengthening resilience and reducing vulnerability.
- Ingestion census, 24 August 2026.
- Component weights sum to 1.000: continuity 0.30, infrastructure exposure 0.25, saturation 0.20, voltage quality 0.10, economic 0.10, transition 0.05. The master equation is
R_final = soft_clip_upper( R_base × Π mult ) + Σ ( add − 1 ). - Completing every declared metric adds 0.29 GB, at a measured 38.4 bytes per declared value — a rate that counts the raw input each metric carries beside it. Annex B carries the full surface.
- The diff is taken against a byte copy of the prior register, captured before anything runs on that watch, so a run cannot quietly rewrite its own baseline.
- No diverging row is unassigned: every one of the sixty-nine carries one of the five owner classes.
- The five rules: reach claims take the coverage figure; measurement claims the treated figure; a quoted paper’s own figure is never altered; a new artefact’s figures are entered first; the line dimension is a route length.
- Stated in a finding of 8 October 2026, after a decision document’s cost table was re-measured three days on: the recommendation survived and the figures a subsequent amendment could cite did not.
- The draft also names the decision that would pin it, so the route from proposal to declaration is on the record before it is taken.
- 1,097 stores named against 231 with a retrieval held, per the estate’s own source finding. The 780 end-point declarations are the per-jurisdiction, per-metric slots the construct requires, which is a different count from the distinct upstream stores.
- Convention #22 was added to the conventions register on 28 September 2026. The concept boundary is a decision of 5 October, applied, its convention text drafted and not yet landed. No conformance row yet tests either.
- Of the 874, 302 carry a DOI and are checked against the resolving record; 572 are checked by bibliographic search. A hand-checked sample of eleven from the no-match bucket returned nine real papers the matcher missed.
- The full list at that class, with declared value, observed value and owner, is held in the conformance register and re-derives by the command in Annex A.
- Six of the fifteen turn on a doctrine decision rather than a code change: the deployment states a normalisation doctrine has not yet parameterised, so a declaration or a re-derivation clears them.
- The 796,121 itemisation gives 780,010 to its nine named jurisdictions, leaving 16,111 for the thirty that carry 133,985 published records.
- Continuity carries intra-weights of 0.40 unplanned duration, 0.30 unplanned frequency, 0.15 voltage quality and 0.15 planned duration, and 0.30 of the composite overall.
- The gap is the difference of the measured means, 0.141874, not the difference of the two rounded figures, which would give 0.1418. Both means are emitted from one pass with two accumulators in the existing walk; there is no second pass.
- The 2024–2025 figures underlying the British landing are carried across two published years with rank correlations of 0.8769 and 0.9253 on the two quantities, reported as a stability check rather than as a result.
- The layer is frozen at a stated hash and carries 14 features. Fourteen regulator codes were aliased to its names, each alias recorded with its reason — including a regulator’s own file mislabelling an operator.
- The resolution census is a complete pass over the deployed payload — 39 jurisdictions, 622,104 records, no sampling. Method and per-jurisdiction results are set out in SB-02 Annex A and reproducible via the re-derivation commands documented there.
- The restatement item in the figures register calls for a note rather than a silent edit, on the principle that a filed document is part of the record and is corrected on the record.
- The estate’s publishing doctrine, 26 August 2026, rule R2 — decision papers, not description papers. The rule carried no check at the time it was written, which is the condition the eighteenth gate now corrects.
- The six questions at Section 5 are stated so that each is answerable from published material alone. None requires access to a proprietary model, a licence or a non-disclosure agreement.
- Primacy claims require a documented search under this methodology. Where none has been run, the claim is scale, scope and purpose rather than priority.
- Full census, seeded 70/30 split, country fixed effects as the baseline, missingness indicator, ridge-regularised. Of three candidate properties one was unavailable and one vacuous; only the survivor is cited.
Provenance
SSI Index · Strategic Brief SB-02
issued 2026-10-08 · CC BY-SA 4.0
scope 39 jurisdictions · 622,104 scored records · census
scored cohort 622,104 · 718,436 at the previous census
conformance register 116 rows — 47 conforming, 69 diverging
re-derived 2026-10-08 · site repo b2a97845 · doctrine repo 282320b
instrument strengthened in period, revision awaiting commit — see Annex A