REWT Rivers of England and Wales, Temporally

Temporality: which dates get published

A reconstruction “between Domesday and 1900” has to say which years it actually draws, and why not all of them. This page is that decision and its reasoning.

Why a continuum is not publishable

Flow is modelled on the network. The water balance is resolved over the network itself, so changing which channels exist changes where the water goes — not merely which lines are drawn. An epoch is therefore a complete run of the pipeline, not a query over a built one.

Be precise about what that does and does not promise. The reference year is a parameter, and the pipeline can be run at any year — that capability is real. What does not exist, and should not be implied, is slicing an already-built network to an arbitrary date without rebuilding it. Promising “a model at any date” should mean the first of those, and say so.

The cost per epoch, as measured in the scoping exercise and not yet re-earned here: roughly half an hour of pipeline, 29 MB of vector tiles and an 89 MB dataset. Seven epochs is around 200 MB of tiles in a repository served as static pages. That is the constraint, and it is why the answer is a curated set of dates, not a slider.

The phases of change

The published dates are chosen to sit between phases, so that each is a stable state rather than a moment mid-transformation.

phase roughly what changes
medieval mills, fisheries, monastic drainage 11th–15th c. leats, diversions, embankment; the Broads dug and flooded
late-medieval retreat 14th–15th c. works decay; some channels revert
statutory river improvement begins 1539–1620s the legislative record starts; navigation works
the great drainage schemes 1626–1660s Vermuyden; the Great Level and Hatfield Chase rebuilt
river navigations 1660–1760 locks, cuts, made rivers
the canal age 1760–1830 wholly new channels on new lines
parliamentary enclosure and land drainage 1760–1850 field drainage; warping
Victorian water supply 1850–1900 reservoirs drown valleys
modern 20th c. pumping, flood relief channels

The proposed set

epoch why this date brackets
1086 Domesday — corroborated by independent evidence the project can hold: the coastline, and the mills before almost all drainage
1300 the high medieval maximum, before plague and the late-medieval retreat monastic drainage; the Broads
1540 the Dissolution; the legislative record begins in 1539 after decay, before statutory improvement
1600 the eve of the great drainage the last state before the Level is rebuilt
1700 after Vermuyden and Hatfield, before the canals the great drainage schemes
1830 after canal mania and enclosure drainage, before the railways displace both the canal age
1900 the survey datum — the map the courses are traced from Victorian reservoirs

The middle column is also what a date selector shows beside each stop, taken from the same file rather than retyped — served as epochs.json for anything outside this site.

And one that is not in the list. The modern network is the datum the whole series is worked back from, and it is published as the first edition. It is not an epoch in the sense used here — it draws no historical claim and needs no evidence beyond the survey — but it is the cross-section every other one is measured against, and it is dated by the issue of the Ordnance Survey product it repairs rather than by a year in the past.

If fewer are wanted, 1540 goes first. It is the least distinct: little changed hydrographically between 1500 and 1600 outside the navigations.

1600 must never be dropped. It is the last state before the Fens and Humberhead were rebuilt, and the single most useful cross-section in the series.

Why the datum is not 1348

An earlier reference year of 1348 was not a historical judgement. It was the terminus of the navigation synthesis the work leaned on, which carries no start date on any of its 652 records — so 1348 was the one year at which every course it attests is attested by the date being drawn. The datum was fixed by the weakest constraint in the evidence.

Once courses are traced from dated sources, each carries its own bounds and that constraint disappears. 1348 keeps a role as a continuity date, but stops being the year the project is organised around.

What the map does with this

A stepped selector, not a slider. A continuous control implies a precision the data does not have. A control that moves between published epochs claims exactly what is true: these are the dates that were built, and there is nothing in between.

And flow cannot be interpolated across epochs. Each epoch’s figures are modelled on its own network, so the selector switches datasets rather than animating one. That should be visible in the interface rather than buried in a note: a reader moving between two epochs is looking at two separate models, and the difference between them is a finding, not a transition.

Editions and epochs are different axes

Editions are how complete the evidence is. Epochs are which date is drawn. They are independent, and confusing them means either waiting for the research before publishing any date, or publishing seven dates of a network that is not yet sound.

The first edition publishes one epoch. Later editions add both evidence and dates.

edition what it adds epochs it supports
First the modern survey repaired into a working drainage network. No historical claim. the modern network itself
Second old courses from map evidence: the labelled channels, the drowned valleys. Bounded and countable. 1900, then 1830 and 1700
Third and after documentary and cartographic research, released a course at a time. Never completes. + 1600, then 1300 and 1086

An epoch is published when its evidence supports it and not before. An empty medieval map is worse than no medieval map, because it looks like a finding.

The first edition is not a historical epoch at all, and labelling it as one would be the first misreading to invite. It is the modern survey repaired, so the honest date for it is the survey’s own issue and not a year in the past. 1900 is the earliest date the map record reaches and the first cross-section that has to be reconstructed — which makes it second-edition work, not first.

What varies between epochs, and what must not

Varying: which channels exist; the flow modelled on them; the coastline, for which transcribed Holocene sea-level curves exist; and which attestations are in force.

Not varying: identity. A stretch present in 1600 and in 1700 must carry the same identifier in both. Epochs are cross-sections of one network, not seven networks — anything else makes change between them impossible to measure, which is the main reason to publish a series at all.

This is also what makes incremental release safe, and it has to be designed in at the first edition rather than discovered later:

How a date is recorded

An epoch is a year the project draws. It is not the same thing as what the project knows about a stretch, and conflating the two is how the old 1348 datum came about. Dates on evidence are therefore recorded separately from the dates that get built, in the Linked Places Format when object: a set of timespans, each with an optional start and an optional end, each of those being either in a year or bounded by earliest and latest, with a certainty on the whole.

Both bounds are optional, and that is the point. Almost every source this project uses knows one and not the other:

what a source actually says recorded as
a mill on a map surveyed in 1824 start: {latest: 1824}
a channel labelled on a sheet surveyed in 1899 start: {latest: 1899}, end: {earliest: 1899}
a reservoir completed in 1869 start: {in: 1869} — and the drowned channel’s end: {in: 1869}
a valley drawn as river on the 1st edition and water on the 2nd end: {earliest: 1854, latest: 1894}
a declared year built that is visibly rounded start: {earliest: 1870, latest: 1880}, certainty: less-certain
a waterway already navigable when a survey window opens in 1600 start: {latest: 1600}
navigable at some point, no year given at all start: {latest: 1600} — the same encoding, honestly
an attestation with no start date at all end only, start absent

The last two rows are the argument for the whole scheme. A model that requires a start date forces every source to invent the half it does not have — which is precisely how a datum can come to be fixed by the weakest constraint in the evidence. An absent start says nobody knows when this began, which is true and is a publishable fact. And where a source records “already open at the start of my window” and “known to have been navigable once, no year”, those two carry identical information: a plain year column would make one look like a date and the other like a hole in the data.

It also puts a finding into the data instead of a footnote. The whole point about the pre-Ordnance-Survey mill mapping is that its year is the map’s year and not the mill’s — a terminus ante quem. start: {latest: 1824} says that structurally. A date column says the opposite and needs a paragraph to undo it.

Unknown is not the same as unbounded. A missing start because nobody knows is a different statement from a channel deliberately modelled as having no beginning, and the two must not be allowed to collapse into the same empty field. Where that distinction has to be drawn, it is drawn explicitly.

Two things this must not become. It is not a licence to interpolate: a bound is a bound, and a stretch with start: {latest: 1600} and nothing else must not be drawn at 1540 because the arithmetic permits it. And documentary uncertainty must not be blended with model uncertainty — how confident we are that a channel existed in 1700 is a different quantity from how confident we are in a modelled flow along it, and one “confidence” number over both is worse than neither.

On convergence. The choice is deliberate rather than convenient. Linked Places is the interchange format for historical gazetteer work, and its current draft both relaxes the requirement for a start and carries certainty on the timespan — which is exactly what evidence of this kind needs. The PLATO attestation ontology expresses the same information as a four-date model, startEarliest / startLatest / endEarliest / endLatest, aligned with PeriodO, and adds a precision term and an explicit flag for a bound that is deliberately open rather than unknown. The two are inter-convertible, and this project’s aim is to be expressible in either rather than to invent a third thing. That matters more here than it might elsewhere, because a river reach carrying its own evidence is an attestation, and the modelling problem is one other people are already solving.

How a release is archived

Editions are published as releases, and a release may carry more than one epoch. Each release is deposited to Zenodo automatically, through its link with GitHub, and minted a DOI — so a reader can cite what they used and a later reader can fetch exactly that, after this project’s own site has changed or gone.

Two DOIs, and citing the wrong one loses the thing that makes a citation useful. Zenodo issues a concept DOI that always resolves to the newest version, and a version DOI fixed to a single deposit. A paper about this project cites the concept DOI. A result computed from a particular edition must cite that edition’s version DOI — otherwise the reference silently comes to point at a later network, and a figure that could once be checked no longer can. Since a release may carry several epochs, the version DOI is the only identifier that says which network a number came from.

And an archive of this repository is not an archive of the data. Nothing acquired, derived or built is committed here — the inputs are declared and fetched, and the outputs are reproducible from them — so a deposit made from the repository alone holds the method, the configuration and the hand-made judgements, and none of the network. That is worth archiving and it is not what a reader following a DOI expects to find. The built outputs are deposited with it, or the record says plainly what it contains; the one thing that must not happen is a DOI that looks like a dataset and is not.

One consequence worth accepting

The first edition will be a modern network, and some readers will take it for a historical one whatever the documentation says. The defence is the same as everywhere else here: every stretch carries what it is and what it rests on, and the map draws the difference rather than describing it in a paragraph nobody reads.