From ReFH Design Events to Continuous Hydrology

Research question

Which parts of the ReFH and continuous-hydrology research record are strong enough to publish, which tested branch failed, and which regional-continuous claims remain outside the evidence gate? The report separates parameter transfer, a public-baseline comparison, daily lumped simulation and future hourly continuous evidence because they do not share a target or validation design.

Evidence and method

Three completed experiments enter the numerical ledger. A 640-catchment refit withheld one hydrometric area at a time. A commercially controlled pilot compared aggregate outputs from the public ReFH1 coefficient and equation stack with genuine ReFH2 outputs across 18 complete cases. A daily two-store study applied a combined QMED, L-CV and KGE decision gate to nine stations and retained an eight-station bounded sensitivity check.

Regional continuous Stage A is handled differently. Publication required immutable score artefacts in this implementation worktree and a match to the cited evidence ledger. Those artefacts are absent, so the regional section carries a structured not-yet-demonstrated status and no score distribution.

Baseflow-lag refit

The flow-derived baseflow-lag refit reached leave-one-hydrometric-area-out R² 0.528 [0.451, 0.583] across 640 catchments. The published WHS in-sample R² 0.35 is retained as a reference, giving a difference of 0.178 [0.101, 0.233]. Withholding whole hydrometric areas makes the refit result relevant to regional transfer rather than within-sample fit alone.

The target custody is material: targets are flow-derived and bias-mapped. The refit retained held-area signal, while the published WHS figure is a contextual reference only because it uses a different validation design. This is not a like-for-like contest between identical experiments. A disattenuated statistic is omitted because it is not measured predictive performance.

Commercial evidence custody

Across 18 complete cases, the public ReFH1 stack differed from genuine ReFH2 pilot outputs by 47.6% on Q100 peak; its mean absolute gap was 53.5%. This is a baseline gap, not a ReFH2 beat. It establishes that the open baseline and the licensed reference were materially different in the pilot.

Only aggregate statistics cross the publication boundary. No licensed case output, private row, coefficient set or reconstruction detail is reproduced. The comparison therefore documents baseline adequacy and evidence custody without exposing the commercial reference surface.

Failed daily two-store branch

The tested daily lumped branch cleared its combined QMED/L-CV/KGE gate at two of nine stations. Under the pre-bounded relaxation, performance reached at most four of eight complete stations. Across the objective variants, model L-CV of 0.18–0.21 remained too narrow to establish the required between-basin growth behaviour.

Only that per-basin daily lumped branch is closed. The result does not rule out regional models, persistent hourly state, alternative model classes or continuous simulation generally. Nor are daily held-period maxima presented as instantaneous design-flood evidence.

Stage A publication gate

Regional continuous Stage A is classified as not-yet-demonstrated in this version. The required immutable score artefacts are absent from the implementation worktree, so the evidence gate fails and no Stage A performance metric, percentile or station count is published.

If a future version satisfies the immutable gate, its admissible claim would still be an observed-forcing hourly hindcast, not weather forecasting, Q100 skill or ReFH2 replacement. That future boundary is stated now so later evidence cannot silently broaden the scientific target.

What survived and what failed

The held-area baseflow-lag refit survived as bounded regional parameter evidence. The aggregate commercial-custody comparison survived as a diagnostic of the public ReFH1 starting point. The per-basin daily two-store branch failed its combined gate and remains closed within that exact implementation boundary.

Regional continuous publication remains not-yet-demonstrated because its evidence gate was not met. That is not a negative score for the model class: the necessary immutable proof is missing from the publication workspace and the public claim remains open.

Operational implications

• Keep parameter-transfer results tied to the target construction and geographic holdout used to obtain them.
• Use the ReFH pilot only to establish the open baseline gap and the need for controlled licensed validation.
• Do not spend further effort tuning the closed per-basin daily lumped branch as a standalone growth-curve solution.
• Require immutable, report-addressable artefacts before publishing a regional continuous score or distribution.
• Keep hydrograph response, event replay and flood-frequency targets in separate claim ledgers.

Limitations

• Baseflow-lag targets are flow-derived and bias-mapped rather than direct observations of a unique physical store parameter.
• The WHS reference is in-sample and therefore does not share the held-area validation design of the refit.
• The genuine ReFH2 comparison contains 18 complete pilot cases and publishes aggregate gaps only.
• The daily evidence concerns a small frozen cohort, daily resolution and one per-basin lumped architecture.
• Stage A contributes no public performance result because its immutable evidence gate was not satisfied in this worktree.

What we do not conclude

• We do not present a disattenuated association as measured held-area predictive performance.
• We do not treat the public-stack gap as superiority to genuine ReFH2 or as evidence of product equivalence.
• We do not generalise the daily two-store failure beyond the tested per-basin daily lumped branch.
• We do not convert daily maxima or hydrograph response into Q100 evidence.
• We do not publish regional continuous performance without the immutable artefact and evidence-ledger match.

Disclosure boundary

This report discloses the research questions, broad cohorts, validation designs, aggregate distributions, failed gate, claim statuses, source revisions and scientific boundaries. Its only sources are the frozen baseflow-lag refit, public-stack gap, daily Arm A, flood-objective and regional Stage A evidence records.

It withholds licensed case outputs, private rows, learned coefficients, fitted parameter artefacts, exact feature construction, calibration sequences, runnable recipes and deployable model internals. The public evidence is sufficient to audit each claim without reconstructing the proprietary method or importing unverified Stage A results.