• Meltzer et al., US Centers for Disease Control and Prevention. Exponential growth without further intervention
• 550,000 cases by 20 January 2015, 1.4 million corrected for under-reporting
• About 21,000 had occurred by mid January

• Camacho, Funk and colleagues. Estimated \(R\), new infections per infected person, by district in Sierra Leone and forecast weekly with a semi-mechanistic model
• Funk et al. scored the forecasts in 2019. Calibrated one and two weeks ahead
• Three weeks or more ahead they overestimated, by more the further out they looked

• Case investigation forms for 4,507 probable and confirmed cases by 14 September 2014
• Incubation period, serial interval and case fatality from the forms, by ad hoc methods. Park et al. showed in 2024 that these lead to biased estimates
• Their forecast. More than 20,000 cases by 2 November if control did not change. 14,383 had been reported by 11 November
• The models on the following slides lean on delay estimates like these

• 12 January 2020. We started on 2019-nCoV, with sparse, delayed case reports from China and detected cases elsewhere
• 3 February. \(R_0\) in Wuhan from the size and duration of the outbreak under a range of serial intervals
• 28 February. Whether isolating cases and their contacts could control new outbreaks, by simulation
• By March we were estimating \(R_t\) for China, Italy, the UK and a growing list of other countries

• \(R_t\) daily for several thousand national and subnational locations, on a public dashboard from spring 2020 to March 2022
• EpiNow2. A renewal process, where today’s infections are \(R_t\) times a weighted sum of recent infections, with delays to report
• Weekly to SPI-M-O, the UK government’s modelling advisers. Other governments used the dashboard directly or through their scientists
• Software widely used by researchers and public health agencies worldwide
• In 2022 I wrote: “we sometimes published nonsensical estimates”

• Variants. Alpha in late 2020, Delta in 2021. The routine model could not estimate their transmissibility on the timeline policy needed
• Chains. So we went back to using the output of one model as the input to other, simple models
• The data changed under us. Tests to hospitalisations in 2021, tests unreliable and sparse by 2022, the Office for National Statistics survey and sequences arriving
• Omicron. Most of the specialised models of 2021 could not take its immune escape and shorter generation time. Chains again

• A global outbreak. A public health emergency of international concern on 23 July 2022
• More than 16,000 cases in 75 countries by then
• Endo et al. A heavy-tailed sexual partnership distribution explained the sustained growth among men who have sex with men
• Why it turned over. Depletion of the high-activity core. The counter-argument was behaviour change. Both sides agreed heavy-tailed networks played a part
• Reuse. Existing COVID-19 models lacked the flexibility to adapt. New, relatively simple, single-source models were built instead

• Nowcasting. Estimating what has already happened but has not yet been reported. Cases by symptom onset arrive late, so the most recent days look like a fall
• Overton, Abbott and colleagues nowcast the England outbreak for UKHSA
• UKHSA could not install other nowcasting methods, so the models were built on standard regression packages

• WHO alerted on 5 May to a high-mortality illness in Mongbwalu, Ituri. The Democratic Republic of the Congo declared its 17th Ebola outbreak on 15 May. A PHEIC on 17 May
• 6,757 confirmed cases and 3,267 deaths in DRC by 7 September. 20 cases and 2 deaths in Uganda
• Six provinces and 61 health zones. The largest Ebola disease outbreak recorded in DRC
• The data arrive as a daily PDF situation report, in French

• McCabe et al., Imperial, 18 May. Outbreak size from the exports to Uganda and from the deaths. 400 to 900 cases in their 20 May update
• We replicated it on 19 May, in a few hours
• Then a joint renewal model. Suspected, confirmed, deaths, exports and bed occupancy from one infection process
• Each stream alone gives a different size. The joint estimate sits inside their spread

• The streams in the situation reports. Suspected and confirmed cases and deaths, laboratory, isolation, recoveries, a digitised onset curve, Uganda exports and a genomic bound
• Updated up to daily. 145 releases since 20 May, each with the draws
• A one-week-ahead forecast scored against the data that came later
• No other regularly updated public model of this outbreak that I know of

• The question first. The estimand decides which streams matter
• Catalogue the data. What each stream measures, and how late it arrives
• One process for infections and an observation model per stream. Delays sit inside the joint model
• Fit, check against what was known independently, report with a date. Then again tomorrow

• Methods built for one outbreak are not reused for the next
• Methods not flexible enough to take the data sources that exist
• The data change under the model. Streams start, stop and get redefined. Daily suspects stopped on 6 August this year
• Evaluation. Knowing which of it worked

A language model with tools. It reads files, writes code, runs it, and loops until a check passes

• I was travelling for summer schools, teaching and workshops, so made a very focused go of working through agents only
• All code and most text I have produced in the last six months has been agent written, pushed from a bot account, a second GitHub login
• Agents write the tasks, from a spec or from a conversation with me
• I am reviewing at different levels. Sometimes all the code, sometimes only the outcome
• My agents also work with Sebastian Funk’s, who is working in a similar way

• Two approaches on epiaware.org. Composed distributions, an event tree of delays. Composable Turing models, transmission, latent process and observation as swappable parts
• What kind of infectious disease modelling language works well for agents?
• epidist. A meta-analysis model for published delay estimates, from scratch in about a week
• EpiDelays. A UHasselt package rewired onto the primarycensored likelihood in two days
• primarycensored. The generalised gamma integral solved analytically by an agent. Three earlier ones we solved by hand

• The situation reports from the Institut National de Santé Publique (INSP) are French PDFs. Agents read every stream from them
• Where a public transcription has the same series, the two are cross-checked
• Recent reports get two blind reads
• Symptom onsets are reported only as a figure. An agent-written digitiser recovers the counts from pixel colour and axis tick geometry
• These are cross-checked again by several agents, and by digitisers written in both Python and Julia
• Onsets also move between reports because reporting is late. That nowcasting problem is in the model

• Model development through issues and pull requests. 333 of 425 pull requests from the bot account
• Model structure and approach set by me, at a high level
• Automated performance work by agents, reviewed by agents. 19 perf commits, all from bot identities. One review found nothing worth shipping. One halved the gradient cost
• Agents managed two versions of the model in tandem. The original integral model and its renewal process successor, kept in step until the renewal model merged on 9 June 2026

• Issues with the McCabe et al. Imperial College report were addressed, and every stream they had used was fitted jointly
• Going wide. One posterior over the sitrep streams, the Uganda exports and a genomic bound
• Checks on every rebuild, following the Bayesian workflow. Predictive checks per stream, forecasts scored against later data, the McCabe replication
• That much debugging would not have happened by hand. Agents did not choose the model. Infrastructure and review kept them in check

• SitRep 081. JPEG compression left the axis ticks one pixel tall. About 10 of 19 ticks found, and a total 54% under the printed one
• SitRep 087. The y-axis grid moved from steps of 20 to 25. Caught only because it broke the reporting-triangle invariant on dates that should have been stable
• SitRep 098. A losslessly encoded image read about 7% high against both neighbours and was excluded
• Every fix must reproduce every previously committed block unchanged

• Forecasts. Bugs keep turning up. Five defects in the one-week-ahead forecast, all in the published report, fixed in August. Four weeks ahead, \(R_t\) ran from 0.245 to 14.9
• Automatic differentiation. Enzyme cannot differentiate the joint model, open since July. Mooncake, the backend we use, paid for type-unstable closures until a fix in September halved the gradient
• Observation models. Agents proposed models that did not match how the response works. Bed capacity as a walk that could only rise. One length of stay for patients who died and patients discharged
• Found by looking at plots, by profiling, and by domain knowledge

• For the Bundibugyo model this was ad hoc. Agents proposed patches to the model as issues, one branch each. Every release scored the forecast against later data, and I reviewed the patches against domain knowledge
• Where there is a known target, the systematic version is a tree search, or an inference step over models. An MCMC proposal in model space
• An agent proposes candidate models, one loop each. Fit and score, keep the best, branch again
• Open questions. How to enforce domain knowledge, and how to pick the evaluation target

• Martinson et al. 2026. A large language model (LLM) guided tree search that writes forecasting code. An ensemble went to the CDC FluSight, COVID-19 and respiratory syncytial virus (RSV) hubs
• Most candidates adapt or recombine models already in the hubs. 142 prompts in three families
• Prospective, 2025/26 season. Matched or beat the CDC ensembles, and ranked first of 43 eligible FluSight, 12 COVID-19 and 4 RSV submissions
• Bracher and Funk 2026. The 11% retrospective gain claimed earlier came from data revisions leaking into the evaluation
• 207,500 candidates scored. My guess is the gap in compute and people against other teams is huge
• The hubs, the data, the target and the review were still built by people

• BVDOutbreakSize. The live DRC outbreak model
• How I am LLM
• A workflow for infectious disease modelling
• epiaware.org/approaches
• The prompts that wrote these decks