News

Release notes for BVDOutbreakSize. Major versions of the report are kept as GitHub Releases; each push to main also republishes the rendered analysis and the output/ artifacts.

v1.18.0

Changes since v1.17.0

Performance

• The observation models no longer box the locals their closures capture (#656).

Four model bodies assigned a variable inside a branch or a loop and then captured it in a comprehension, which makes Julia hold it in a Core.Box. A boxed local is type-unstable at every use, and Mooncake answers type instability with a dictionary lookup per call site on every gradient evaluation. Assigning each name once leaves the log-density bit-identical and roughly halves the gradient on the joint fit, from about 21 ms to about 10 ms. test/test_boxed_captures.jl now fails if any method under src/models/ carries a box, and the rule is recorded under "Closures in model code" in the contributing guide. Three sites outside src/models/, in forecast replication and plotting, still carry the pattern, and are left alone because nothing there runs on the gradient path.

Report

• The symptom-onset reporting-triangle panels are nowcasts rather than fits (#648).

Each panel drew the expected count for the bar it was compared against, which does not condition on what that snapshot had already printed, so its interval was set by uncertainty in the onsets rather than by how much of the reporting delay had elapsed. Each panel now starts from the counts its own figure printed, adds only the reporting the fitted delay curve puts between that figure and the latest one covering each onset date, and puts the result through the measurement error a digitised bar carries. It is read against that latest reading, so the band and the point it is scored on are the same quantity. The single figure by onset date is unchanged, since it is read for the gap between the modelled onsets and what the figures carry, which is ascertainment.

• Seven references the text quoted by hand are now citations, so they reach the References page (#627).

The onset-to-sample cohort, the epidist model behind it, the Wilson-Hilferty median, the RealStar assay and the three GeneXpert sources were all named in prose or given as a bare DOI, which left the assay sensitivities and the delay prior unsourced on the page.

Data

• Advanced the model cut-off from SitRep 110 (1 September) to SitRep 115 (6 September) (#655).

Confirmed cases reach 6686 and confirmed deaths 3226, with 1563 recovered and 819 in isolation. Every day's net confirmed-case and confirmed-death change matches that report's own printed 24h figure, with no harmonisation anywhere in the run. The treatment-centre flows and the daily suspected case and death series stay frozen, as they have since the reports dropped the tables that carried them.

• The ERVEBO ring and front-line-worker vaccination campaign is now tracked as a candidate signal (#651, #650).

Eight rows backfill every numeric mention from its first appearance in SitRep 097 through SitRep 113, each double-read against the source report. The signal is not fitted, and SitReps 114 and 115 have not yet been read for it.

Infrastructure

• Removed forecast_vs_truth_trajectory, which no code called.

It scored a cumulative trajectory from the exponential-growth model's r, expected_reports_T and k, none of which the renewal model carries.

Dependencies

• Turing 0.48 is allowed alongside 0.45, 0.46 and 0.47, in the package and in the Enzyme test environment (#652, #653).

v1.17.0

Changes since v1.16.0

Fixes

• The Python onset-curve digitiser no longer disagrees with the Julia reference (#594).

Three off-by-one errors in the port's translation of the reference's 1-based ranges had put it 175 cells adrift over SitReps 083, 094-096 and 106-109. Rounding half to even is not translation-invariant, so bar windows rounded in the 0-based frame sat one column from the reference's. Both scripts now reproduce the committed onset CSV byte-identically across all 44 vintages, and no committed row changes.

• The symptom-onset reporting triangle now carries a per-scan level error (#507).

A bar's height is read in pixels and converted with the axis scale that scan calibrated, so the absolute part of the digitisation error belongs to the bar and the multiplicative part is one number for the whole figure. Scoring that second part as independent per-cell noise got the spread right and the shape wrong: independent errors average out across a snapshot's cells, so the net correction each snapshot adds was predicted far too tightly, at 1 of 11 snapshots inside a nominal 50% interval while the 90% interval and the aggregate variance were both at nominal. The modelled level each cell differences now carries its own scan's multiplier, sampled rather than assumed, and the per-cell scale keeps counting and pixel noise alone. On a simulated reporting triangle this moves per-snapshot central coverage from 0.29 to 0.48 against a nominal 0.50, and 90% coverage from 0.69 to 0.89. The outbreak size and the recent reproduction number are unchanged by the term. C_T moves from 13022 to 12566 against a 95% interval about 12000 wide, and R_T from 1.009 to 0.976 against an interval width of 0.85. On the joint fit itself, per-snapshot central coverage moves from 8 of 30 to 17 of 30 against a nominal half. The reporting delay's calendar walk narrows from 0.49 to 0.11, so it had been absorbing per-scan level movement as reporting drift. The same term gives the fit somewhere to put a vintage that reprints at its predecessor's level, which previously could only be fitted by driving the reporting hazard towards zero at the delays that vintage covers.

• Bed forecasts are now scored against the occupancy the situation reports print.

The projection carries the reclassification offset the model absorbs a change of reporting basis with, so it no longer sits above the series it is compared against. The persistence baseline drops a window spanning one of those basis changes, and the symptom-onset stream has a baseline for the first time. For an incident stream the baseline's step pool holds changes in the window total rather than the window's own count (#623, #612).

Report

• The bed-occupancy panels read as occupancy rather than as counts of new events (#628).

All three are census stocks but shared a flag with the per-day flows, so the vintage predictive plots labelled them "Daily count" and "New per vintage". A rising occupancy series under a new-events axis, among true incidence panels, reads as an accumulating total.

• The symptom-onset methods section now derives the digitisation error from how a bar is read rather than quoting a single per-scan percentage, and the reporting-delay summary table and pair plot carry the sampled per-scan level (#507).

The recovered stream's description is also corrected: it still said the observed totals were 12 to 40 over 6-13 June, when the series now runs from 12 on 6 June to 1409 on 31 August over 81 vintages.

• The frozen validation fits no longer include the streams the situation reports have stopped updating (#611).

Suspected cases and suspected deaths stopped reporting early, so their frozen fits scored a validation panel that had already gone quiet, 16 fits a build down from 18. The release-fit registry is hashed into every fit's cache key, so the next docs build refits the whole set once.

Data

• Advanced the model cut-off from SitRep 108 (30 August) to SitRep 109 (31 August).

Confirmed cases reach 6186 and confirmed deaths 3007. Sud-Kivu prints its first bed-capacity figure, 25 beds. The daily suspected case and death series and the seven treatment-centre streams stay frozen, as they have since the reports dropped the tables that carried them.

• Three scan-versus-mirror disagreements in the confirmed case and death history are settled against the situation-report PDFs (#624).

The committed values are right in all three, so no fitted number changes. 2026-08-08 is the mirror filing SitRep 085 under its publication date rather than its reporting date. 2026-08-11 is SitRep 089's printed total of 4566 against its own province rows summing to 4567. 2026-08-25 is a mirror transcription error, 2755 for a figure that reads 2744 in the headline, the total row and the province sum alike. The cross-check script now records each with both values, and fails on any disagreement it does not document or on an entry that stops reproducing. It no longer prints TOML regenerated from the mirror, which covers 86 of the 102 report dates the manifest holds.

• SitRep 090's mixed-direction harmonisation gets no break-date entry (#569).

The gap is 2 cases and 3 deaths, smaller than the SitRep 065 precedent already left off that list.

• SitRep 098 stays out of the digitised onset curve, and the reason is now the measured one (#594).

The render size is not what makes it read high. It is the only vintage embedded losslessly, and the fixed colour masks lose a bar-edge fringe to JPEG blur on every other one.

• The SitRep 102 to 103 date-alignment failure is not a misread axis tick on either vintage (#617).

Both neighbouring pairs land cleanly on shift 0.

• The per-scan digitisation error on the onset curve is a per-vintage level, not a stationary error (#636).

The colour masks are fixed thresholds, so how much of each bar survives them depends on how blurred its edges are, and a smaller render blurs more. Edge softness roughly doubles between SitRep 105 and 106 as the render halves in area, and the digitised total falls 214 on onset dates that can only accrue. The shift does not cancel in the between-vintage increments the reporting-delay hazard is fitted through.

Infrastructure

• The onset-curve digitiser and the file it writes now have tests, including the date-alignment sweep and a parity check holding the Python port to the committed file (#629).

• The release rescore's hard failure is now permanent rather than a stopgap (#588).

It used to fall back to the committed scoring tables when the rescore step failed, but that fallback would publish skill scores and validation figures against whatever truth series those tables were last written from, with nothing on the rendered page saying so. A failed rescore now fails the build outright, which a re-run of the job recovers from; a silent substitution was not detectable from the page at all.

v1.16.0

Changes since v1.15.0

Report

• The cross-release forecast scores, the forecast-versus-observed validation tables and the situation-report data tables stay tables on the page.

A DataFrame is html-showable, so each one went into the page as a raw HTML block, and Documenter compiles a regex from every raw block's own text, which fails once a block passes PCRE's compiled-pattern limit. The scoring tables grow with every release and crossed it after v1.15.0, taking the docs build down with them. The stop-gap printed each table as a fixed-width block of text instead, which built but is not a table. They now go out as markdown tables, which carry no such limit. Float columns are rounded rather than printed to their full binary expansion, and numeric columns are right-aligned.

Data

• Advanced the model cut-off from SitRep 102 (24 August) to SitRep 108 (30 August), six vintages.

Confirmed cases reach 6100 and confirmed deaths 2950. Two health zones enter the case table at SitRep 104, Biena and Manguredjipa in Nord-Kivu, taking the affected count from 58 to 60 of 151. Bas-Uélé prints its first bed-capacity figure at SitRep 106. The daily suspected case and death series and the seven treatment-centre streams stay frozen, as they have since the reports dropped the tables that carried them.

• SitRep 107's isolation occupancy is excluded rather than recorded.

Its page-1 tile reconciles exactly to the province prose, but only because Nord-Kivu is absent from that section altogether, having carried 236 the day before and 282 the day after. Bed capacity for the same date is kept, on the split already established at SitRep 085.

• The onset-curve digitiser reads SitRep 108.

That vintage embeds its chart at a size no neighbour uses, which anti-aliased the axis border and the weekly ticks too light for the mask the digitiser reads dark pixels with. A looser mask now runs only where the strict one finds nothing, so every earlier vintage still digitises to the committed values.

Dependencies

• Turing 0.47 is allowed alongside 0.45 and 0.46, in the package and in the Enzyme test environment.

v1.15.0

Changes since v1.14.0

Fixes

Five defects in the one-week-ahead forecast are corrected here. All five affected the headline forecast in the published report, not only the evaluation, so skill numbers from earlier releases were measured against a forecaster carrying them. That is why 90% coverage sat at 1.00 at nearly every stream and horizon against a nominal 0.90.

• Each observed count stream now projects from its own cumulative trajectory.

cumulative_reports and cumulative_deaths_total were named by the forecast and by the stream registry but defined in no model, so reported cases, suspected deaths and confirmed deaths all fell back to inverting the cumulative total under exponential growth. That inversion collapses towards zero once the fitted growth rate is at or below zero, and on one cached fit gave a median 0.55 confirmed deaths per day against an observed 17.

• The confirmed-death forecast is no longer capped at the cumulative suspected deaths.

The model does impose a thinning, but per day and on the latent pool, before positivity scales it down. The forecast capped a different quantity against a reported headline that froze at 246 on 26 May while confirmed deaths passed 246 on 19 June, clamping the forecast below its own origin and flooring the new count at zero in most draws.

• Each stream is replicated through its own dispersion rather than the population mean.

The per-stream forecaster already used the right one, so the two disagreed for the same model on the same stream.

• The observation replicate accumulates day by day rather than in one draw on the horizon total.

The dispersion is fitted against single-day counts and short vintage increments, so carrying it on the summed mean inflated the overdispersion term by roughly the length of the horizon.

• The reproduction number continues its fitted walk past the cut-off.

The forecast repeated the last sampled innovation as a fixed daily slope for the whole horizon, so the spread in log-R_t grew with the horizon rather than with its square root, putting the four-week reproduction number between 0.245 and 14.9. It now draws fresh weekly innovations at the fitted step scale and interpolates between them, as the walk is built in the first place.

The forecast validation is corrected with them.

• A retrospective harmonisation now comes out of the cumulative truth as well as the new-count truth.

A projection cannot contain an administrative reattachment, so leaving it in the truth scored the forecast against something it could not produce. This is a no-op at the current cut-off and bites whenever a break day falls inside a validation week.

Report

• The joint posterior predictive checks are split by whether a stream is still reported, cumulative then per-vintage within each group, and the vintage axes tick weekly rather than crowding as the outbreak runs on.

• The forecast is validated only against streams the situation reports still update.

Reported cases and suspected deaths stopped on 26 May, so their new-count truth was a guaranteed zero and the figure was scoring a forecast against an unmoving series, which the release scoring already withheld. They are still drawn, in their own figure, as a projection rather than a validation.

• The sensitivity page said the confirmed new-count rows keep any retrospective harmonisation step.

The code has subtracted it since the break-day correction was consolidated.

Data

• Advanced the model cut-off from SitRep 100 (22 August) to SitRep 102 (24 August).

Both are clean days, with the cumulative case and death increments equal to the printed 24h gross. Bas-Uélé prints its first traveller count at SitRep 102.

Infrastructure

• The docs-preview cleanup collects previews that land after the pull request closes.

The close-triggered job assumed the preview existed by then, but the docs build is fanned across runners and publishes hours later, so it found nothing and exited green every time, leaving two hundred preview directories against three open pull requests. A nightly sweep now drops every preview directory with no open pull request behind it, and refuses to run when the pull-request listing fails.

Known issues

• The release comparison mixes two forecast constructions.

Every row in it is reconstructed by running each release tag's own code, so the historical rows keep the defects corrected above and cannot be regenerated without rewriting what those tags would have produced. Read a change in forecast width across this release as a change in the code rather than in the outbreak.

• The automatic version increment is wedged by a stale branch and skips silently on every push, so Project.toml does not advance on its own (#607).

v1.14.0

Changes since v1.13.2

Data

• Advanced the model cut-off from SitRep 089 (11 August) to SitRep 100 (22 August). The reports returned to their full-length format at SitRep 090 after five vintages of the shorter "MVEBDB" brief, restoring the laboratory, point-of-entry and continuity-of-care sections. The confirmed-case and confirmed-death series, recoveries, isolation occupancy, bed capacity and the 24h analysed volume all advance with them, and the symptom-onset curve is digitised for each new vintage that carries a fresh figure. The daily new-suspected-case count stays frozen at 5 August and the treatment-centre patient-movement series at 2 August.

• A sixth province, Bas-Uélé, records its first confirmed case at SitRep 090 (ZS Buta). No isolation, occupancy or laboratory figures are printed for it in any vintage, so it contributes zero to every per-province stream, following the convention already in place for a non-reporting province.

• The 17 August point was recorded from the INRB-UMIE mirror alone while INSP had published nothing beyond SitRep 093, verified against its posts and media endpoints rather than inferred from a failed fetch. Publication has since resumed and every stream is scanned from the reports again. scripts/check_new_sitreps.jl reports how far the mirror leads insp.cd, so a mirror-only point stays visible rather than silently absorbed.

• candidate_signals.csv gains a province column and holds one row per signal per vintage per province, replacing the national figure that was a hand-made sum over whichever provinces happened to print a count. A province the report is silent on now gets no row, which distinguishes it from a printed zero. province = Ensemble marks a national figure the report prints itself. Coverage changes are now readable from the data rather than only from the prose in source_note: SitRep 069, for example, gives Ituri and Nord-Kivu for eds_death_alerts but Ituri alone for eds_investigations_performed.

• Corrected SitRep 089's eds_investigations_performed, whose national value of 107 included the 15 CTE EDS its own note recorded as excluded. The per-province rows carry community-alert EDS only and sum to 92.

Fixes

• The still-expected-deaths density no longer extends below zero. The draws were never negative; the kernel density spread mass past the smallest draw, putting a tail on the impossible side of the bound. The axis now stops at the bound, as the count and CFR panels already did. The same applies to the projected-total panel, which had been drawn crossing left of the observed-deaths rule.

• The forecast reproduction-number density stops at zero, the same bound already applied to the still-expected-deaths and projected-total densities. A posterior sitting close to zero picks up a tail on the impossible side from the kernel.

• The posterior outbreak age on the tree-prior sensitivity figure stops at zero for the same reason. An age in days cannot be negative, and that figure renders in every published build.

• The doubling-time interval is the image of the growth rate's interval rather than the quantiles of its own draws, which bounded nothing once the posterior for the growth rate spanned zero and reported a range of -306 to 301 days. The row now runs from the fastest decline, through the zero-growth pole, to the fastest growth, and is ordered by growth rate rather than by value. The narrative summary carried the same fault and is fixed with it.

• The joint model's score tables show the joint model alone. Every fit was rendered as a row, so each stream's individual fit appeared twice, once in the headline table and again in the section built for it. A frozen fit, which is the joint model refit at an earlier cut-off, was also read as an individual fit rather than a joint one.

• The persistence baseline is withheld where its own window is not covered by the vintage's reporting, rather than scored from a degenerate centre. An uncovered window let the centre saturate at the whole cumulative to the made date, or collapse to a point mass at zero, and either flattered the fits it was compared against. A test now pins that the baseline sees nothing after the date it is made on, and fails under three deliberate leaks.

• The in-report forecast validation subtracts a retrospective harmonisation from its new-count truth, as the release scoring already did. A frozen cut-off sitting before a listed break day and a current cut-off after it put records into the confirmed truth that were never notified that week. The cumulative rows are unchanged, the harmonisation being genuinely part of the reported total.

Infrastructure

• The macOS test cell runs the platform-sensitive items only. It was the one cell skipping nothing, so it re-ran the sampling fits and the quality checks on the slowest runner, never finished, and was cancelled at GitHub's six-hour ceiling on every run. It now completes in about twelve minutes. The job also carries a timeout below that ceiling, so a stall reports as a failure rather than as a cancellation hours later.

• The automatic version increment deletes its branch and warns when GitHub refuses the pull request it opens, rather than leaving the branch behind. A leftover branch tripped the increment's own already-exists guard on every later push, so the first failure was loud and the rest were silent.

• Turing's compat bound admits 0.46.

Documentation

• Rewrote the reduced-data-streams warning in the README and the summary page. The situation reports returned to their full format on 12 August, so the note now covers only what stayed frozen: the daily new-suspected-case count and the treatment-centre patient-movement series.

• Rewrote the abstract around the situation reports and the data they publish rather than a list of streams, in the wording the analysis report itself uses. It now covers the digitised symptom-onset curve and the recoveries, which the earlier text left out, and says that each release projects each DRC stream a week ahead and scores those forecasts against later data and a persistence baseline.

• A newly spotted surveillance signal now goes to an issue rather than a row in data/README.md's scan table. The table is a short index, and rows added to it on a data update had to be reverted; the values still accumulate in candidate_signals.csv on the same pull request either way.

• The docs deploy job serialises on the gh-pages ref it writes to. The workflow's own concurrency group is keyed on the source branch, so two pull requests publishing previews at once raced and the loser's push was rejected.

• Gave one display block each to the tested BVD share and the positivity, the cumulative infections and the completed-detection-delay term, and the daily export intensity and its running sum, finishing the one-quantity-per-block pass. The export-death prevalence gets its own block rather than an inline definition.

• Split the death-pool BVD share and the assay positivity built on it into a display block each, matching the confirmed-case pipeline.

• Cut the symptom-onset reporting-delay methods section and its results text to the length and shape the other observation modules carry. The model definition, the priors and the limitations all remain. The identifiability discussion is condensed, and the prose runs one sentence per line like the rest of the document.

v1.13.2

Changes since v1.13.1

Data

• Advanced the model cut-off from SitRep 082 (4 August) to SitRep 089 (11 August): added SitReps 083-089, advancing the confirmed-case and confirmed-death series and the recovered, isolation-occupancy and 24h laboratory-analysed streams, and digitising the symptom-onset curve vintages restored from SitRep 087 (the 084-086 brief-format reports carry a different, notification-week chart and so contribute no onset snapshot).

• From SitRep 084 (6 August) the INSP switched to a shorter "MVEBDB" brief format that no longer publishes the daily new-suspected-case stream (suspected_daily_history freezes at SitRep 083) or the treatment-centre patient-movement streams (frozen at SitRep 080), so the fit now depends on just the confirmed-case, confirmed-death, recovered and isolation streams rather than the full surveillance suite. The README and report carry a warning to this effect; see the inclusion rules in data/README.md.

v1.13.1

Changes since v1.13.0

Data

• Advanced the model cut-off from SitRep 081 (3 August) to SitRep 082 (4 August): added SitRep 082 and the missing 3 August (SitRep 081) confirmed-case / confirmed-death cumulative points that the previous data update left at 2 August, and fixed the SitRep 081 onset-curve digitisation (the x-axis tick detection had undercounted that figure to −54%; the fix brings it inside the noise band) so both the 3 and 4 August onset snapshots are now included.

v1.13.0

Changes since v1.12.0

Data

• Advanced the model cut-off from SitRep 079 (1 August) to SitRep 081 (3 August): added SitRep 080 (2 August) and SitRep 081 (3 August) (#557).

Model

• Widened the growth-rate r prior (eq 9) to LogNormal(log(log 2 / 11.7), 0.40), a doubling time of 5.3–25.6 d at 95%. The centre still matches the BEAST X reanalysis of 139 BDBV genomes (Mbala-Kingebeni and others, 2026), which reports an Exponential-growth doubling time of 11.7 d (95% HPD 6.8–17.5). That HPD is conditional on a single-rate coalescent. This is the assumption the Mongbwalu field epidemiology contradicts (Kupferschmidt, 2026), the same evidence behind the m change in v1.12.0. An independent reanalysis of the earlier genomes puts the doubling time at 15.2–24.5 d (Cuomo-Dannenburg and Ghafari, 2026), which the old spread largely excluded.

• Corrected the growth-rate citation in the rt_walk_model docstring, which still named a generic 20 d molecular-clock estimate the model no longer uses.

• Fitted the digitised symptom-onset reporting triangle as a stream (onset_reporting_model), scored on its between-vintage increments through a discrete reporting-delay hazard that is nonparametric over the delay and drifts over calendar time. The loader collapses reprinted vintage figures by exact value equality, filters to the manifest cut-off, and builds between-vintage increment cells over a trailing 28-day window. Fitted by bvd_joint alongside every other stream and by a new onsets_only_model single-stream composer.

• Gave the onset stream an explicit ascertainment. The reporting hazard is normalised to reach one, so it carries the delay shape alone, and ascertainment is a separate level anchored on the confirmed pipeline's p_drc · τ_test · positivity averaged over the delay PMF, with a logit-scale offset and a slow walk over onset date. onsets_only_model falls back to a constant 0.15 anchor.

• Added a symptom-onset nowcast and forecast (forecast_onsets), splitting the coming week's new onset reports into the part arising from onsets that have already happened but are not yet reported (onset_reports_backfill) and the part from onsets that have not yet happened (onset_reports_future). cumulative_onsets moved from bvd_joint to the shared _latent submodel so every composer carries it. The reported increment is scored across releases as the onset reports stream.

Fixes

• Widened the onset calendar-walk prior so it can follow reporting drift (#530).

• Bound the observation-scale slack below at one, preventing a fitted scale below the measurement floor from outvoting other streams in the joint fit.

• Fall back to four degrees of freedom on a degenerate Student-t nu argument instead of producing a Cauchy likelihood with no mean or variance.

Documentation and infrastructure

• The analysis prose stated the confirmation-process sensitivity prior as Beta(10, 1.76) where test_sensitivity_model samples Beta(38, 2), and gave the wrong rationale for it. Both the number and the rationale are corrected (#548).

• The analysis prose stated a 15-day TMRCA censoring SD where the model uses 16. Corrected to match the code (#548).

• plot_rt, plot_cumulative_trajectories and the single-stream overlay (plot_stream_trajectories) each drew a 50%-width credible band while an inline comment and docstring described it as part of a 30/60/90% ribbon trio. All three now compute a genuine 30/60/90% ribbon and drop the unused median-line description, so the figure and its documentation agree.

• docs/src/contributing.md documented an abstract-marker mechanism (<!-- ABSTRACT:START/END -->) that does not exist in README.md (the real marker is <!-- SHARED:END -->), eleven integrator, submodel and composer function names that do not exist in src/, two integration constants that do not exist either, and a surveillance dispersion prior of truncated(Normal(0, 1); lower = 0) where the code uses truncated(Normal(0.6, 0.2); lower = 0). Rewrote the repository layout, running-and-testing and model-architecture sections against the current code and Taskfile.yml.

• Reflowed every write-up page, comment and docstring to the repository's prose rules: one sentence per line, shorter sentences, no run-ons, no restated points, no development history, and no code identifiers inside narrative prose.

• Removed emphasis-caps and issue and pull request references from src/, test/, scripts/ and the report pages. That history now lives here instead.

• Added a carve-out to the no-issue-numbers rule in contributing.md. An issue number may stay where it is the provenance record for a still-open decision, rather than an account of what changed.

• Removed predict_committed and its three helpers from src/counterfactual.jl. They called delay_convolution and DEATH_INTEGRAL_ALG, neither of which is defined anywhere in src/, left behind by the rename to convolve_delay in renewal.jl. Nothing referenced them, so removing them changes nothing (#545).

• Removed eight unused one-off scripts. scripts/Project.toml gained a missing Base64 dependency, and scripts/README.md now documents which Julia project each script needs.

• Removed the notes/ folder, including the prose style guide the rules were derived from, which now lives in contributing.md. Its remaining open items are filed as issues #544 to #549.

• Added AGENTS.md, pointing at README.md, contributing.md and scripts/README.md, so a session starts with the rules in context.

v1.12.0

Changes since v1.11.0

Model

• Widened and shifted the prior on the cryptic-phase doubling count m, from truncated(Normal(3, 3); lower = 0) to truncated(Normal(5, 4); lower = 0). Field epidemiology in Mongbwalu traced a sustained transmission chain to a death on 25 January 2026, with 500+ suspected cases between mid-January and mid-May (Kupferschmidt, 2026), and the genetic TMRCA (Mbala-Kingebeni and others, 2026) is a lower bound on the outbreak age consistent with an origin that early. The new centre places the implied prior on the outbreak origin at the end of January (at the central 11.7-day doubling, m = 5 gives a cryptic duration of ~58.5 d, i.e. an origin around 30 January), with the wider SD letting the data and the genetic seeding bound pull it earlier or later (#533). The main fit's m prior is now set directly in exponential_growth_model; the backfill's advancing centre (m_prior_centre) is unchanged and tracked separately by #534.

• Corrected the growth-rate r prior prose (eq 9) to match the code and its actual source: the BEAST X reanalysis of 139 BDBV genomes (Mbala-Kingebeni and others, 2026) reports an Exponential-growth doubling time of 11.7 d (95% HPD 6.8–17.5), so the prior is LogNormal(log(log 2 / 11.7), 0.28). The text had previously described an older, generic 20 d estimate that the model no longer uses.

v1.11.0

Changes since v1.10.0

• Updated data processing and added digitised onset data

• Updated forecast evalaution but this remains highly experimental

• Added break days for confirmed cases and deaths due to a data harmoisation adding backdated data.

v1.10.0

Changes since v1.9.0.

Data

• Advanced the model cut-off from SitRep 055 (8 July) to SitRep 064 (17 July 2026) across four batches: SitReps 056–057 (9–10 July), SitRep 058 (11 July), SitReps 059–061 (12–14 July, from the INSP source), and SitReps 062–064 (15–17 July, stepping over the unpublished 063). as_of_date → 2026-07-17. Every fitted stream (confirmed cases/deaths, daily new-suspects, occupancy, bed capacity, recovered, 24h analysed, Tableau 6 treatment flows) is extended through each batch.

• Captured symptom-onset epidemic curves from the DHIS2 line list published in the analytique SitRep figures (059–062, 064). Digitised via scripts/digitize_onset_curve.jl (Julia) and a byte-identical Python port, and recorded as data/onset_curve_scanned.csv. Fitted as a stream, see Model below.

• Re-ran the confirmed/suspect in-care occupancy split on the resumed Tableau 7 data (SitReps 052–055), extending treatment_confirmed_incare_history and treatment_suspect_incare_history through 8 July (#413, revisits #373).

• Added scripts/check_new_sitreps.jl to detect stale manifests by comparing published INSP SitReps against the latest recorded vintage.

Model

• Made the digitised symptom-onset reporting triangle a fitted stream (onset_reporting_model), scored on its between-vintage increments through a discrete reporting-delay hazard that is nonparametric over the delay and drifts over calendar time. Fitted by bvd_joint alongside every other stream and by a new onsets_only_model single-stream composer. The snapshot figure is a posterior predictive, so its band carries the measurement error the likelihood gives a digitised bar.

• Gave that stream an explicit ascertainment. The reporting hazard is normalised to reach one, so it carries the delay shape alone, and ascertainment is a separate level. It is anchored on the confirmed pipeline's own p_drc · τ_test · positivity averaged over the delay distribution, with a logit-scale offset and a slow walk over onset date. Not yet validated. The ascertainment walk and the reproduction-number walk share the onset axis and are both least constrained late, so over the final fortnight and need reporting either side of this change.

• Added a symptom-onset nowcast and forecast (forecast_onsets), splitting the coming week's new onset reports into the part arising from onsets that have already happened but are not yet reported and the part from onsets that have not yet happened. The latent onset trajectory cumulative_onsets moved from bvd_joint to the shared _latent submodel so every composer carries it. The reported increment is scored across releases as the onset reports stream; the triangle's cumulative level is deliberately not scored, since every vintage rereads the whole figure and its ≈4% per-scan level error revises the total both ways.

• Added a results figure reading the triangle along the onset date: the latest digitised bar for each onset date against the posterior predictive for that bar and against the modelled onsets. The nowcast and forecast figure separates the latent components from the replicate the next vintage will print, since only the replicate is scored.

• De-boxed the anonymous map(do) closures on bvd_joint's default log-density path (confirmed_positivity_link = :composition): extracted the composition-positivity loop to a plain composition_positivity() function and replaced three occupancy/split map(1:n) do t … blocks with elementwise broadcasts. Pure type-stability tidy-up; Mooncake gradient is bit-identical (same hash(g)). Clears the layer-1 prerequisite for the opt-in Enzyme reverse-mode backend — Enzyme cannot construct a shadow for anonymous closures that capture conditionally-scoped variables (they get boxed in Base.RefValue), and after this fix reaches LLVM's nodecayed_phis! pass (next blocker tracked in #445; #446).

Fixes

• Restored the renamed clock-sensitivity chain reference (chn_joint_fast_clock → chn_joint_exp_growth_clock) in analysis diagnostics after a merge inadvertently reverted it to the old, undefined variable. This had broken the Documenter → Render analysis step on main when BVD_RUN_SENSITIVITY=true (#418).

Documentation and infrastructure

• Added data/README.md documenting the two data sources, the direct-INSP fetch workflow via the WordPress REST API, the per-SitRep field checklist, and the drift-check protocol.

• Added Julia (scripts/digitize_onset_curve.jl) and Python (scripts/digitize_onset_curve.py) digitisation scripts for the symptom-onset epidemic curves, producing byte-identical CSVs. Python deps managed via uv (PEP 723).

Dependencies

• Widened the Turing compat from 0.45 to 0.45, 0.46, allowing the upstream package evolution while keeping 0.45 pinned as the working default (#415, #416).

• Updated compat bounds for ADTypes (1.22.2), CairoMakie (0.15.13), Distributions (0.25.129), LogDensityProblems (2.2.0), Integrals (4, 5.4), StatsFuns (2.2.0), SHA (0.7.0), Aqua (0.8.16), TestItemRunner (1.1.5), CensoredDistributions (0.2.22), and TestItems (1.0.0).

v1.9.0

Changes since v1.8.0.

Data

• Updated data including movement flows for those in isolation.

Model

• Updated the molecular-clock time estimate to use the outbreak-specific BEAST X analysis (mbalaplacide2026, 139 BDBV genomes from 16 health zones, ~1.1E-3 subs/site/year). The genetic TMRCA baseline moves from 2026-03-25 (SD 15, fixed 1.2E-3 EBOV rate) to 2026-03-15 (SD 16, 95 HPD 09 Feb–12 Apr) under the Skygrid coalescent prior. Added a tree-prior sensitivity comparing the Exponential growth estimate (2026-03-08, SD 16, 95 HPD 01 Feb–05 Apr).

• Updated the growth-rate prior to the outbreak-specific doubling time from the same analysis: centre moves from 20 d (Cuomo-Dannenburg & Ghafari) to 11.7 d (95 HPD 6.8–17.5, mbalaplacide2026), with the log-SD widened from 0.15 to 0.3 to match the wider credible interval.

Report and forecasts

• Saved the one- to four-week-ahead forecasts as a release asset (forecast.csv), plus the one-week-back validation forecast (forecast_validation.csv), at each results release. Past forecasts were shown in the report but never stored, so they had to be reconstructed by re-running each past release's own code on that release's data; these are published separately as a backfill release. Each release now records the forecast it made, so it can later be scored against what is observed. This underpins the new cross-release forecast scoring (CRPS, log-scale CRPS, coverage, bias, and relative skill against a persistence baseline) shown on the sensitivity page.

Documentation

• Updated all prose references to the genetic bound and growth-rate prior to cite the new virological.org report (v1045, mbalaplacide2026) and the outbreak-specific rate and doubling time.

• Added a tree-prior sensitivity section comparing the Skygrid and Exponential growth TMRCA estimates.

Performance

• Halved the default post-warmup draws in nuts_sample from 2000 (2 chains x 1000) to 1000 total (2 chains x 500), and moved the docs fit registry (build_fit_specs, fit_key, fit_content_hash) to the same 500 x 2 setting. This roughly halves the sampling wall-clock at the cost of some effective sample size.

• Trimmed the default NUTS warmup cap in nuts_sample from min(250, samples ÷ 2) to min(200, samples ÷ 2), so at the new samples = 500 default each fit runs 200 adaptation steps rather than 2. Shortens warmup by a further ~20% per fit.

v1.8.0

Changes since v1.7.0.

Model

• Credited the repeat-control confirmation process in the confirmation sensitivity prior. Rule-out is investigative rather than a single negative PCR, so the effective sensitivity is higher than one assay draw (two controls give about 0.98). The headline test_sensitivity_model prior moves from the single-assay Beta(10, 1.76) (mean 0.85) to Beta(38, 2) (mean 0.95) on the confirmed and confirmed-deaths streams. The outbreak-size estimate is robust because the sensitivity enters the multiplicative ascertainment ridge (p_drc · s_test · τ_test); the ascertainment posterior re-centres (resolves the retesting/rule-out part of #374).

• Grounded the confirmed onset-to-sample delay on the NEJM DRC 2026 cohort (Akilimali et al.) as a standard part of the joint model. The onset-to-report and report-to-receipt legs already convolve to onset-to-sample for confirmed cases. The convolution's mean (the sum of the report and receipt leg means) and median (Wilson-Hilferty of the summed leg variances) are fitted to the reported 7.4 d and 4.8 d as soft Normal observations, with the reported 95% credible intervals as their SDs. This grounds the otherwise-unidentified laboratory-turnaround delay directly from the cohort's own uncertainty and adds no latent parameter. On by default in the joint fit; the single-stream and isolation fits lack the laboratory receipt leg and so do not carry it (resolves #359).

• Scored the confirmed-case laboratory positives as an overdispersed BetaBinomial of the observed analysed denominator instead of a plain Binomial. A plain Binomial on denominators of several hundred specimens gave posterior-predictive intervals far too tight, so the confirmed stream was systematically under-covered: the smooth pooled / composition-linked per-window positivity does not capture the day-to-day laboratory batching and within-window positivity heterogeneity the confirmed counts carry. A single intra-window overdispersion ρ (confirmed_overdispersion_model, weakly-informative Beta(1, 24)) inflates each window's variance to n·p·(1 − p)·(1 + (n − 1)·ρ), identified across the laboratory windows, and recovers the Binomial as ρ → 0. The mean structure and the composition link that identifies the background λ_bg are unchanged.

Documentation and infrastructure

• Excluded the published released_estimates.csv overlay from the fit content hash, so the render jobs reuse the matrix fits instead of refitting every model. The render step rewrites that overlay before rendering, which changed the data-tree digest and busted every fit cache key; the overlay feeds only the estimate-evolution figure and is not a fit input. tree_sha256 and content_hash gained an exclude for non-input data files, and genuine data changes still refit.

v1.7.0

Changes since v1.6.0.

Data

• Added the situation-report Tableau 6 treatment-centre patient-movement flows (CTE/CT/CI) as optional daily streams: admissions, in-care deaths, rule-outs and absconded patients (13–23 June). Each stream is resilient: an empty history is a no-op, so the model degrades to the occupancy backbone where a flow is not reported. Advanced the data through situation report 046 (29 June).

Model

• Reworked the isolation submodel into a treatment-centre flow model that fits the Tableau 6 flows alongside occupancy. The bed length-of-stay is an outcome mixture, with the death and recovery branches weighted by an in-care case-fatality CFR_iso = logistic(logit(CFR) + β_iso) — a reported modifier on the infection case-fatality, identified by the in-care death flow rather than estimated independently. The daily discharge flows are scored as optional negative-binomial streams (resolves #338).

• Added a manual, opt-in occupancy reclassification-break offset. Break days are listed explicitly in [occupancy_break_dates], replacing the removed threshold detector that flagged too many days. A level step is fitted into the modelled occupancy mean at each listed day, so the fit tracks a known between-report measurement-basis discontinuity without bending Rt to chase it. Each step is centred on zero, so the fit partitions it into reporting artifact vs real demand. The 19 June DHIS2 reclassification (occupancy 416 → 361, confirmed by the au-lit start-of-day stock) is listed; the joint fit with the isolation stream had been bending Rt up and down to chase this and the later missing-SitRep steps, which no single-stream fit shows.

Report and forecasts

• Added one-week-ahead forecasts for the treatment-centre admissions, in-care deaths and rule-outs.

• Added a per-stream calibration plot (with the calibration table kept in a collapsible block), and posterior-predictive panels for the four flow streams. The treatment-centre flow methods section was rewritten, and the flow streams added to the data-overview table.

• Added a comparison of the confirmed-case projection against Chamla et al. as a second external comparator, forward-projected from a dedicated frozen fit at their 8 June confirmed-case calibration anchor. This carries the confirmed-case testing history, replacing the poorly-identified 27 May proxy that had effectively no testing data. The 8 June fit also shows as a vintage in the estimate-evolution overlay (resolves #340, #349).

• Refreshed the released-estimate evolution overlay to v1.6.0 and refresh it automatically in continuous integration before each documentation deploy. Dropped the per-release current-model re-fits from the estimate-evolution plot, keeping the matched-McCabe cut-offs and the one-week-back validation fit (resolves #341).

• Tightened the reproduction-number plot y-axis to 1.2 times the 90% upper bound so the credible band is legible (resolves #342).

• Added a posterior correlation heatmap across the key estimates (outbreak size, reproduction number, outbreak age, CFR, ascertainment, background, fraction tested and each stream's expected total) and a pairs plot of the per-stream modelled totals against each other and the observed value, so the size-versus-ascertainment trade-off and per-stream over/undershoot are visible in one place (resolves #346).

Performance

• Halved the NUTS warmup: nuts_sample now defaults n_adapts to min(250, samples ÷ 2) (250 adaptation steps at the standard 1000 draws) instead of Turing's min(1000, samples ÷ 2) (500), cutting the discarded warmup iterations on every report fit. Pass n_adapts explicitly to override.

• Centred the per-stream pooled negative-binomial dispersion (pooled_dispersion_model) instead of drawing it non-centred, and made centred the default. The surveillance streams are data-rich, so the non-centred pooling funnelled as τ → 0; the centred form is an exact reparameterisation that removes the funnel (resolves #352).

• Put the bed-capacity baseline C0 on the log scale (LogNormal(log 450, 0.42) in bed_capacity_model and bed_capacity_walk_model) instead of a truncated normal, so the whole capacity C(t) = C0·exp(walk) is log-consistent with no hard boundary, improving the worst-mixing capacity block (resolves #358).

Fixes

• Relaxed the tau_death test assertion to non-negativity: the joint exposes the realised cut-off death-testing intensity (analysed over suspected, a diagnostic computed independently of the death volume), which is not a probability and can exceed one in a backlog regime.

• Widened the per-stream dispersion pooling-SD prior τ in pooled_dispersion_model from HalfNormal(0.3) to HalfNormal(0.6), resolving a prior-data conflict where the posterior τ sat entirely above the old prior's tail because the stream dispersions genuinely span ~9× (resolves #336).

• Fixed the one-week-ahead forecast to project new counts over the horizon and add them to the cut-off cumulative, instead of scaling the cumulative stock by exp(r·horizon), which made a below-one reproduction number imply an impossible shrinking cumulative in the Chamla comparison (cases_cum, deaths_cum, confirmed_cum, confirmed_deaths_cum and recovered_cum; resolves #351).

Documentation and infrastructure

• Split the report into two literate pages rendered from a shared setup: an analysis page (methods, results, one-week-ahead forecast) and a sensitivity page (forecast validation, per-stream outbreak size, estimate evolution, McCabe and Chamla comparisons, delay and clock sensitivity), so the deploy no longer renders one 4.3k-line file in a single job (resolves #364).

• Fanned the documentation build into a list → fit → render → combine CI grid: one content-addressed, cached NUTS fit per matrix job, the two pages rendered in parallel from the cached chains, and a combine job that deploys and publishes, dropping the critical path from all fits serialised behind one runner to roughly the slowest single fit.

• Pinned the shared workspace-root Manifest.toml as the artifact the docs grid uploads and restores, since docs/Project.toml is a workspace member and resolves the root manifest rather than docs/Manifest.toml, so the fit, render and combine jobs resolve the same package set (resolves #368).

• Set include-matrix: false on the render job's env-cache step so the two render jobs restore the precompiled depot shared by list and fit instead of keying on the matrix page, which missed the shared cache and made each render re-precompile the whole stack (~530 deps) before rendering.

v1.6.0

Changes since v1.5.0.

Model

• The isolation BVD treatment length-of-stay uses the BDBV line-list admission-to-death delay as its prior.

• The reproduction-number random walk starts a month before the first situation report (RT_WALK_LEAD = 28, exposed as the bvd_joint keyword rt_walk_lead), so R_t can move over the weeks of transmission leading up to that report. The walk start is floored at the renewal start, and the plot_rt reconstruction uses the same knot grid.

• Added a supply-limited isolation/treatment-bed stream ("Patients en isolement"), the renewal analogue of the convolution secondary-observation model of EpiNow2. Bed occupancy may be supply-driven (demand can outstrip supply), so the model fits a latent bed demand, the suspect inflow carried through a length-of-stay survival (BVD cases with a sampled treatment stay, non-BVD suspects leaving after a sampled rule-out stay), right-censored at an effective bed capacity ρ·C(t) (a censored negative binomial). The capacity C(t) is a random walk (bed_capacity_walk_model) that tracks the beds being added and can be projected forward, pinned by the implied bed count (reported occupancy / "Taux d'occupation" rate) on the days a rate is published. The stream exposes the bed demand, occupancy, capacity, shortfall and utilisation, and carries its own observation dispersion. Added convolve_survival, the treatment_admission_model observation submodel, the isolation_admission_model, bed_capacity_model and bed_capacity_walk_model priors and the treatment_only_model single-stream composer (resolves #265). This is a single national model, so it cannot represent local bed saturation (Ituri at 93.9% occupancy on 13 June against Sud-Kivu 21.9%); the national shortfall understates the local unmet need.

• Gave the non-BVD isolation rule-out stay its own sampled length-of-stay (ruleout_los) in treatment_admission_model, separate from the report-to-receipt laboratory delay, so the occupancy identifies the rule-out stay on its own clock and the lab-turnaround delay is set by the testing, composition and confirmed-death streams. Exposes isolation_ruleout_los_mean.

• Added a recovered-among-confirmed stream ("cumul guéris"), the secondary-observation incidence analogue: survivors among the modelled daily confirmed cases, scaled by a recovery proportion and convolved with a confirmation-to-recovery delay, with its own observation dispersion. The recovery proportion is grounded on the case-fatality ratio (a recovered case is one that did not die) with a log-odds adjustment for the confirmed population, rather than estimated independently. The confirmed model exposes one daily confirmed-case series (confirmed_daily) that both the recovered stream and the cumulative- confirmed trajectory reuse. Added the recovered_model submodel and the recovery_probability_model prior.

• Added the exports_joint_only_model composer, which fits the Uganda export cases and deaths together over the one travel-gated at-risk prevalence. The single-stream comparison in the walkthrough now shows one joint "exports" fit instead of separate export-case and export-death fits.

• The one-week-ahead forecast also projects the isolation/treatment beds: the bed demand a week ahead (need under unconstrained supply) and the supply-limited occupancy it produces, whose gap is the projected bed shortfall, and the cumulative recovered total, each replicated with its own dispersion. Added plot_forecast_beds, which shows the projected bed need against the supply-limited occupancy and the shortfall in the walkthrough's forecast section.

• The forecast-versus-frozen validation now also scores the isolation beds: the frozen one-week-back fit conditions on the isolation occupancy, and the projected bed occupancy is compared against the beds actually held a week later (forecast_vs_truth gains an isolation argument, and plot_forecast_beds_vs_truth plots the projected occupancy against the observed beds). The bed check is weak at a one-week-back freeze because the reported occupancy rate starts only on 9 June, so the capacity rides its random walk back to the freeze date.

• Replaced the per-vintage step background random effect with a smooth daily lognormal random walk (background_walk_model): a per-day background with no reporting-vintage steps, gated to begin a report-to-receipt lead before the first suspected-case report, shared across the suspected-case and suspected-death streams, with a half-normal baseline and a tight random-walk innovation SD. This also removes the per-vintage step in the modelled cumulative-death trajectory.

• Widened the non-BVD background level prior so the laboratory positivity (210/755 ≈ 0.28 positive) identifies it. The suspect pool is inferred to be a minority BVD, which lowers the cumulative-infection estimate (C_T) with a wider credible interval.

• Gated the laboratory analysed-specimen capacity to the testing onset, so no specimens are modelled as analysed before testing existed.

• Redesigned the death pathway. Suspected deaths carry a death ascertainment p_death (the death analogue of the case ascertainment, with an informative prior centred high) and a non-BVD death background that applies a background CFR (cfr_bg) to the suspected-case background and lags it by the onset-to-death delay, so a background death follows its background case. The death background tracks the identified case background rather than a second free, outbreak-size- degenerate rate, and inherits the case background's smooth gated daily shape, so the modelled cumulative-death trajectory is smooth. Added the death_ascertainment_model and background_cfr_model priors.

• Rebuilt the confirmed-death stream as a laboratory pipeline mirroring the confirmed cases. The death analysed volume scales the modelled case analysed volume at the per-day suspected death-to-case ratio, times a testing-intensity scaling (LogNormal(0, 0.25), centred on one), so death testing follows the laboratory's realised throughput; the death-to-case ratio carries the suspect-pool severity and the suspected-death level. The volume is scored through a death-pool composition positivity p = s·q_death + (1−spec)(1−q_death), with q_death the BVD share of the suspected deaths from the death series' own components. The case volume carries the laboratory capacity onset, so the death volume inherits it and no deaths are confirmed before testing began. The joint exposes the death_ascertainment, background_cfr, death_testing_scaling, tau_death and death_composition deterministics and drops m_death.

Data

• Added the daily "Patients en isolement" occupancy for 1-11 June (SitReps 018-028) as a structured patients_isolated column and the [isolation_history] manifest block. The fitted series begins 1 June where the all-patients column definition is stable; the narrower suspects-only count in SitReps 016-017 is a different quantity and is excluded. Corrected the SitRep 020 note (the PDF headline occupancy is 233, not the 173 the note claimed).

• Added the implied bed-capacity series (occupancy / reported "Taux d'occupation" rate ≈ 400-452 beds, 9-13 June) as the [bed_capacity_history] manifest block, which pins the bed capacity in the supply-limited isolation model.

• Added the cumulative "cumul guéris" recovered-among-confirmed total for 6-11 June (SitReps 023-028) as a structured cumul_recovered column and the [recovered_history] manifest block.

• Added the peer-reviewed McCabe et al. Lancet Infectious Diseases publication (online first 9 June 2026, DOI 10.1016/S1473-3099(26)00299-9) as a third scenario vintage in REPORT_SCENARIOS_CI, with inputs as of 27 May 2026 (1031 DRC cases, 240 deaths, three Uganda imports). Both methods now vary the epidemic doubling time (7/10/14 d); the back-calculation assumes 30% of deaths are attributable to Ebola. The published paper swaps the method numbers relative to the Imperial reports, which the noted convention reconciles. Recorded the matching frozen-data snapshot in data/report-snapshot-27may.toml.

Analysis

• The walkthrough adds posterior-predictive panels for the isolation occupancy and recovered streams, a single-stream "in isolation" fit for the isolation occupancy, and surfaces the isolation length-of-stay and confirmation-to-recovery delays in the observation-delay table and pair plot, and the admission proportion, recovery probability and the two new per-stream dispersions in the surveillance-parameter table.

• Cite EpiNow2 (Abbott et al., 2020) for the convolution-and-scaling secondary-observation analogy, and fix the epinow2 bibliography entry so the documentation build no longer warns about a missing field.

• The McCabe et al. scenario comparison now carries a third vintage, the 27 May 2026 Lancet publication, plotted beside our renewal estimate on 27 May, with a frozen re-fit at the 27 May cut-off added to the frozen-fit outbreak-size table.

• Quantified the per-vintage posterior-predictive checks with a per-stream calibration table (stream_calibration): the mean forecast bias and the empirical 50%/90% interval coverage of each stream's one-step-ahead conditional predictive, so the streams the joint fit reproduces less well can be read off rather than eyeballed. Added the bias_sample scoring helper (resolves #269).

Documentation

• Added a one-page Summary dashboard for readers with limited time: the headline estimates as prose and tables alongside the reproduction number, infections-over-time and modelled-versus-observed reported-case figures. It reuses the artifacts written by the analysis build rather than re-fitting, so it refreshes whenever the data updates.

Outputs

• Added the latent symptom onsets (the "symptomatic cases" outcome) to the shared posterior outputs. posterior_draws.csv gains a cumulative_onsets_T column, the cumulative symptom onsets by the cut-off per draw (the onset analogue of C_T), and a new onsets_over_time.csv records the daily new and cumulative onset trajectory over time with 30/60/90% credible intervals. Exposed through onsets_over_time.

v1.5.0

Changes since v1.4.0.

Model

• Confirmed deaths now carry the report-to-receipt laboratory delay, so the laboratory-confirmed-death series lags the death event rather than tracking it instantaneously and the confirmed case and death streams pay a consistent laboratory delay.

• Added an optional daily new-suspect inflow stream ("nouveaux cas suspects du jour") to the suspected-case likelihood. The post-26 May per-day counts are scored against the modelled daily suspected series at each report day, continuing the suspected signal where the frozen cumulative series stops, on days disjoint from it (#222).

• Added the deaths analogue, an optional daily new suspected-death inflow stream ("cas suspects du jour N (M deces)") to the suspected-death likelihood. The post-26 May per-day counts are scored against the modelled daily suspected-death series at each report day, continuing the suspected-death signal where the frozen cumulative series stops, on days disjoint from it, and a matching "New suspected deaths/day" posterior-predictive panel is added alongside the new-suspects-per-day panel.

• Collapsed the laboratory pipeline onto a single suspected-to-analysed volume, fit to the specimens-analysed series through one report-to-analysed delay. The received stream is still recorded but no longer fitted, and the post-cut-off 24-hour analysed volume is now fit directly.

• Late reporting windows, where the cumulative national analysed denominator stops, are scored in one submodel. A day with a published 24-hour analysed count anchors its positivity as a binomial on that count, and the remaining days are scored against the modelled laboratory volume. These are a reporting-format change rather than data blackouts, so the earlier "dark window" framing is dropped.

• Added the delay-corrected confirmed case-fatality ratio, the Nishiura et al. (2009) real-time correction computed per posterior draw on the modelled confirmed trajectory and sampled confirmation-to-death delay. The denominator shrinks from all confirmed cases to those expected to have had a fatal outcome resolve by the cut-off, debiasing the naive confirmed ratio.

Forecast

• The one-week-ahead forecast and its validation now target the laboratory-confirmed case and confirmed death streams. The suspected reported cases and deaths are no longer published, so they no longer serve as forecast targets or as the last-week-versus-now comparison.

Report

• Added a confirmed case-fatality ratio section, setting the delay-corrected confirmed CFR against the structural infection-based CFR and the naive confirmed ratio, with a comparison table and posterior-density plot.

• The estimate-evolution figure now draws each release as a discrete per-fit estimate with nested 30/60/90% intervals, read from data/released_estimates.csv rather than a hand-maintained literal. Frozen renewal re-fits are restricted to the integral-era release cut-offs, since renewal-era releases already are renewal fits. A new scripts/refresh_releases.jl pulls the per-release estimates from the tagged results releases.

Data

• Captured the daily new-suspect counts (SitReps 021-024, 4-7 June) as a structured new_daily_suspects column in the scanned situation-report CSV and a suspected_daily_history block in the observation manifest.

• Captured the daily new suspected-death counts ("cas suspects du jour N (M deces)", SitReps 024-032, 7-15 June) as a structured new_daily_suspected_deaths column in the scanned situation-report CSV and a suspected_daily_deaths_history block in the observation manifest, the deaths analogue of the daily new-suspect inflow.

• Extended the confirmed case and death series to SitRep 025 (8 June).

• Added the trusted-day 24-hour analysed laboratory counts (1, 4-7 June) as a tests_analysed_daily_history block to anchor late-window positivity.

v1.4.0

The methods switch flagged in v1.3.0: the continuous-time, fixed-growth-rate model is replaced by a discrete-time renewal model that is simpler and avoids the single-stream-versus-joint size tension of issue #212. This is a substantial revision; the changes below are relative to v1.3.0.

Model

• Replaced the integral exponential-growth model with a discrete-time renewal process on a daily grid. Infections follow the renewal equation under a time-varying reproduction number (a weekly log-scale random walk with an intervention ramp), and every observed stream sits downstream of latent onsets through its own sampled, discretised delay.

• The prior is placed on the growth rate (the molecular-clock doubling time) and the first reproduction number is derived forward through Euler–Lotka. The generation interval is a Gamma with shape and scale taken from the cited source and its reported uncertainty.

• The onset-to-event delays are taken from a Bayesian reanalysis of the 2012 Isiro line list on their natural Gamma parameters, with one onset-to- admission delay serving both suspected-case reporting and export detection and onset-to-death the convolution of two atomic components.

• Two-phase seeding: a single import grows through an unobserved cryptic exponential phase to the renewal start, with the outbreak age bounded by the genetic time to the most recent common ancestor.

• Confirmed positivity is tied to the suspect-pool composition through an assay sensitivity and specificity, and exports are travel-gated from infection and scored on their dated detection days.

• The DRC streams are fitted on the incidence scale, as the between-vintage increments across successive situation reports (the first vintage being the cumulative count to that date).

Report

The report was rebuilt around the renewal model; the analyses carried over from v1.3.0 (the one-week-ahead forecast and its validation, the no-onward-transmission counterfactual, the delay and clock-rate sensitivity analyses, and the McCabe et al. comparison) were re-implemented for the new model rather than added here.

• Restructured the methods in generative order (infections, epidemiological processes, observation models, the joint model) with the model maths given explicitly.

• Reworked the figures (reproduction number with credible ribbons and sampled trajectories, cumulative infections, onsets and deaths, outbreak size by data stream, and estimate evolution across releases).

• The McCabe et al. scenario comparison now carries their reported 95% confidence intervals.

• The one-week-ahead forecast and its last-week-versus-now validation now target the laboratory-confirmed cases and confirmed deaths. The suspected reported cases and deaths are no longer reported, so they are dropped as forecast targets.

Data

• Advanced the cut-off to 7 June 2026 (SitRep 024). The laboratory-confirmed streams run to the cut-off while the suspected streams stay frozen at their 26 May values.

v1.3.0

Final release of this model formulation

This is the last planned release of the continuous-time, fixed-growth-rate model. The laboratory and testing observation model has outgrown the available data, and the joint fit now implies a larger outbreak than any single data stream does on its own (issue #212). We are replacing this model with a discrete-time renewal model, which is simpler and avoids these problems, in a follow-up release that will note the methods switch. Treat the estimates here as provisional.

Changes since v1.2.0.

Data

• Advanced the model cut-off to 28 May 2026 and switched the DRC streams to the INSP national cumulative totals read from the situation-report PDFs, rather than the per-zone CSVs whose zone sums drop cases not yet attributed to a zone. The suspected streams are frozen at their 26 May values, after which INSP stopped publishing a national suspected total; the confirmed and laboratory streams run to 28 May.

• Added per-sitrep-vintage confirmed cases, confirmed deaths and laboratory throughput (samples received and analysed) to data/observations.toml, alongside the suspected cases and deaths.

• Extended the observed series through 5 June 2026 to validate the forecast out of sample.

Modelling

• The joint model now fits four DRC streams per sitrep vintage (suspected cases, suspected deaths, laboratory-confirmed cases and laboratory-confirmed deaths) by conditioning on the between-vintage increments, alongside the Uganda exports and export deaths. A single-vintage stream reduces to the cumulative likelihood.

• Confirmed cases are fitted through a laboratory-throughput queue: suspects enter a received backlog after a report-to-receipt delay, a capacity-limited drain sets the samples analysed, and the new positives in each window are a Binomial on the samples newly analysed. Windows with no published analysed count fall back to the queue's expected throughput, so no free per-window denominator is introduced.

• Test positivity is severity-first: early specimens skew toward severe presentations and relax toward the latent case composition as analysed volume accrues.

• Suspected cases and deaths are BVD onset-to-report convolutions plus additive non-BVD background rates; confirmed deaths share the case-lab PCR sensitivity and specificity.

• Split ascertainment into independent DRC and Uganda reporting fractions (DRC centre 0.75), and recentred the growth prior on the molecular-clock 20-day doubling time ((Cuomo-Dannenburg and Ghafari, 2026)).

• Exports and export deaths are timed from infection via an infection→detection delay convolution rather than a rectangular detection window, reducing to the McCabe et al. window as the delay collapses to a point mass.

• The headline estimand is cumulative infections (2^m), with the under-ascertainment multiplier anchored on the laboratory-confirmed cases.

Outputs

• Posterior summary table, a laboratory-pipeline pair plot, and posterior-predictive panels for the confirmed-case and confirmed-death streams in the per-stream-versus-joint grid.

• Recast the forecast around the four trusted quantities (infections, true BVD deaths, confirmed cases, confirmed deaths) over a one-week-ahead and a counterfactual-year horizon, dropping the untrusted suspected and tests-analysed streams.

• Restored the forecast validation as a last-week-vs-now out-of-sample check: fit the joint through 28 May, forecast forward, and score the predicted confirmed cases and deaths against the observed counts.

• Added a conditional one-step-ahead predictive across the sitrep series, each vintage predicting only its new increment.

Documentation

• Surfaced the delay priors as equations, clarified that the latent pool is the true-case count rather than the tested or confirmed count, and added limitations on the constant-growth assumption and on per-sitrep increments mixing incidence with backfill.

Infrastructure

• Added streaming progress to nuts_sample via an optional callback (a dependency-free file stream or TensorBoard), and optional Enzyme reverse-mode AD alongside the default Mooncake backend.

v1.2.0

Modelling

• Improved the comparison to the McCabe et al. report by making sure that 95% credible intervals are being compared and reordering it.

• Added a custom chain rule for SpecialFunctions.gamma_inc. This allows us to differentiate through the analytical solution to the gamma convolution integral.

Data

• Moved the cut-off to 23 May 2026 and switched the DRC source from the WHO AFRO joint sitrep to the situation reports of the Institut National de Santé Publique (INSP), transcribed by INRB-UMIE/Ebola_DRC_2026. The INSP series gives a per-zone, per-sitrep daily vintage trajectory (suspected and confirmed; this analysis uses suspected). Cumulative counts at 23 May: 905 suspected DRC cases, 220 suspected DRC deaths, across the 12 reporting health zones. The 18 May INSP vintage (516 cases, 131 deaths) matches the WHO joint sitrep 01 total exactly.

• Updated Uganda to three travel-related imports with one death, reflecting the third import announced on 23 May 2026 (woman from DRC who travelled Arua to Entebbe to Kampala; tested positive on follow-up). Two further Uganda-confirmed cases announced the same day (a driver and a healthcare worker) are domestic contacts of the first import and are excluded from exported_cases because the model treats Uganda as imports only.

• Added a reported_case_history block in data/observations.toml with eight INSP sitrep vintages (14 May to 23 May 2026), ready for the cumulative-trajectory likelihood once it merges.

Infrastructure

• Moved the submodels out of the analysis file and into the supporting package. Instead we now print these in the analysis.

• Added additional package infrastructure including Aqua.jl and Jet.jl.

• Streamlined the package unit tests.

v1.1.0

Modelling

• Bound the seeding time T from below with a soft prior on the genetic time to the most recent common ancestor (TMRCA), following a suggestion from Neil Ferguson to combine the genetic signal with the other data streams as a seeding bound.

• Switched the export deaths to a daily (time-resolved binned) Poisson process: a continuous survival weight for the no-death stretch before the first dated death, then a per-day Poisson from that day to the cut-off.

• Bound T with export-death timing through that survival weight, and with case-export detection timing through a first-export-detection survival term on the Uganda admission date. Dates supplied in data/observations.toml.

• Death-convolution quadrature adapted to the sampled delay scale.

• Added a clock-rate sensitivity: refit the joint model under the faster 1.9e-3 early-epidemic TMRCA estimate and compare the impact on outbreak size, seeding time and growth rate against the 1.2e-3 baseline.

• Sped up the deaths-among-exports likelihood: precompute the onset-to-death CDF once and reuse it across bin edges (ExportDeathDelay), replacing the per-node nested quadrature.

• Removed hardcoded death and case constants that diverged from the observations in data/observations.toml.

• Added a forecast validation: fit the joint model to the original report's data, project it forward to the current cut-off, and compare the predicted cumulative and new counts per stream against the counts observed since, as a table and a 2×3 coverage plot.

Data

• Updated to the McCabe et al. 20 May 2026 report, comparing both report versions.

• Sourced the genetic TMRCA seeding bound from the BEAST temporal-tree estimate in the 2026-05-21 virological.org update (mean 2026-03-25, 95% HPD 2026-02-20 to 2026-04-20, at the 1.2e-3 EBOV clock rate this analysis assumes).

Infrastructure

• Dropped MCMCChains for FlexiChains and prepared for registry release.

• CI docs preview PR comments and version-bump automation.

Docs

• Added a scope note to the README and analysis report framing the work as an external view built on our understanding of real-time infectious disease dynamics, and inviting feedback, reuse and adaptation.

• Surfaced results from the README and analysis landing page, added stable and dev docs badges.

• Plotting and labelling fixes: surveillance dispersion on the 1/√k scale, predictive histograms labelled as frequency, and coarser (four-weekly) start-date axis ticks so the labels stay readable.

• Reworked the headline summary to report the credible intervals as sentences rather than leading with a median, defined the prior-IQR shift, and explained the reported-case scaling in terms of the DRC reporting fraction with a link to the pair plot.

• Replaced the model-structure diagram with a parameter-to-observation table.

• Culled promotional register in the analysis report.

v1.0.0

First release. A joint Bayesian re-analysis of the McCabe et al. report that fits all data streams together in a single Turing model over the latent cumulative case count.

• Conditions on the exported cases and DRC deaths the report uses, plus reported DRC cases (with an ascertainment component) and deaths among exported cases.

• Adds a no-onward-transmission projected-deaths counterfactual, a one-week-ahead forecast of newly reported cases, deaths and exports, and an onset-to-death delay sensitivity analysis.

• Replaces the deaths-convolution and small-growth-rate exports closed-form approximations with their exact forms.

• Maths-first analysis page with code folded behind dropdowns and a diagram of the model build-up.

• Compares against a joint reimplementation of the report's approach and its original published estimates.