• \(R_t\), the reproduction number over time. How many people each case goes on to infect
• Weekly \(R_t\) estimates and reports to SPI-M-O, the Scientific Pandemic Influenza Group on Modelling, Operational sub-group
• SPI-M-O reported to SAGE, the Scientific Advisory Group for Emergencies
• The public saw a consensus statement

We don’t know whether any of this gave any policy maker any useful information that helped them make better decisions…

Abbott and Funk, 2022

• Many groups. Each used its own data sources and its own model
• Evidence from several models was considered, discussed and combined
• A range went out. The detail behind each estimate did not

If results are presented individually this can get lost and it incentivises being first out (and maybe therefore cutting corners).

Abbott, 2023

• epiforecasts.io/covid, April 2020 to March 2022
• Daily \(R_t\) and nowcasts for several thousand locations. A nowcast estimates what has happened but has not yet been reported
• The front page was a map. Five words, increasing to decreasing, and no numbers
• Just over 500,000 unique users and 1.2 million page views by March 2022

• Orange for estimates from partial data, adjusted for right truncation, the recent cases not yet reported
• Each figure caption carried the same sentence. “These should be considered indicative only”
• The limitations sat on a methods page that 20,000 of 1.2 million page views reached

• CSV files on GitHub, comma separated values, used directly by other governments or by scientists who set them beside other evidence
• The estimates travelled. The limits travelled less well
• Checking several thousand locations a day was beyond us
• Could AI do that checking?

Of course this can cause issues if the limitations of the method are poorly communicated.

Abbott and Funk, 2022

Because we did not have the capacity to manually inspect the data and estimates on a daily basis we sometimes published nonsensical estimates.

Abbott and Funk, 2022

• Wisconsin reported on a Sunday for the first time. The model read the catch-up as an outbreak
• What I wrote down afterwards. Be as open as possible. State the limitations first. No point estimates
• I left the press to people whose speciality it is. No training, no time

• Each team submits weekly. The hub combines them into an ensemble
• The European hub ensemble beat 83% of teams’ case forecasts and 91% of death forecasts on relative WIS, the weighted interval score

However, they can be very hard to learn from, and though we have really tried to learn more about how to forecast it has been difficult.

Abbott, 2022

• Why. A forecast can only be read against the outcome and a score. Agreeing the score is agreeing what a good forecast is
• A hub is an agreement. Same targets, same dates, same quantiles, same score. WIS is the hubs’ headline measure
• Shared practice. scoringutils implements the scores. Pairwise comparison gives a relative skill that teams who miss weeks can still read
• UKHSA, the UK Health Security Agency, asked us for guidance on evaluation. Practice still differs between agencies

• Who runs ours. The US Centers for Disease Control and Prevention, UKHSA, the Robert Koch Institute and the World Health Organization
• Users run the defaults, so the defaults are the advice
• The rest is support. We meet analysts from health agencies and answer their questions
• In some cases the package is too complex for anyone else to use

• EpiNow2 has been on CRAN since 1 September 2020
• Every argument a user leaves alone is a decision they inherited from us
• The generation time and the delays. The prior on \(R_t\). The sampler settings
• The code on the right is the vignette. It is what gets copied

• epinowcast. Researchers and public health analysts. 31 seminars from May 2023 to May 2026
• A forum, a monthly meeting with a rotating chair, release notes that name everyone who reported a bug or joined a discussion
• In 2023 I wrote: “I am also interested in why efforts in this direction keep failing”

Trying to build a community of practice as community >> methods or models alone.

Abbott, 2023

• The live report on the Bundibugyo virus disease outbreak puts its limitations before the methods, in bullets, most consequential first
• Drafted with agents. The report, and the limitations list with it
• The feedback is that a list of limitations first reads as off-putting
• How should we warn?

Every estimate is a model-based extrapolation under strong assumptions, not a measurement.

Abbott, Sherratt, Brand and Funk, 2026

• A bot account. seabbs-bot, from this morning. Agent work goes out from it
• A disclosure line in the Bundibugyo repository’s README and on the limitations slide of the WHO call about it
• A prompts page with the brief and the steers
• A bio drafted by the bot. My JuliaCon bio came from seabbs-bot, “who has a very suspiciously high opinion of me”

• The JuliaCon 2026 site. One brief, then 130 steers by 14 August 2026
• Mid-build the prompts page counted 53. The number more than doubled before the talks
• This site has one too. The brief, the steers and the notes these decks were rewritten from

• Fast, and the result looks the part
• I am dyslexic. Drafting is the part I am glad to hand over
• More of this got built than I would have built by hand

• A layer between what I want to say and what gets said
• The words arrive already formed, so I have to push to get mine back
• Every steer on the prompts page is me pushing

• 114 claims pulled from one deck by a second agent, each checked against git and GitHub’s application programming interface (API)
• 67 needed changing
• “530 pull requests” was GitHub’s latest issue number. The count was
  353. A pull request is a change proposed for review
• 712 of 1,371 merged bot pull requests carried no review, so “reviewed by me” became “about half”

• A GitHub App. It reviews each pull request I or my bot open, and again when either of us asks
• 213 pull requests reviewed since 17 August 2026. Each review ends “Not a human review”
• I fix what it gets right, say which findings I reject and why, and check each against the source myself
• It has caught real bugs. It has also been wrong from its own truncated greps

• The live Bundibugyo virus disease model. Four named authors. Most of its 497 commits are from bot identities
• One answer to how we warn. Say who drafted it and who answers for it. The report does, at the top of its README

The model code and analysis were drafted by a language model, then reviewed and revised under human oversight. The named authors are responsible for that oversight.

Abbott, Sherratt, Brand and Funk, 2026

$ git log --format='%an' | sort | uniq -c | sort -rn
 238 Sam Abbott (bot)
 144 seabbs-bot
  50 dependabot[bot]
  40 Sam Abbott
  11 Sam
   7 Sebastian Funk - robot edition
   3 Sebastian Funk
   3 Claude
   1 Samuel Brand

• Sebastian Funk’s bot opened the issue on 20 July with a diagnosis, and came back the next day with a fitted prototype
• My bot closed it on 4 August with the commit that fixed it
• No person wrote in the thread. The issue before it went the same way
• sbfnk-bot has opened 11 issues on EpiNow2 this year. One was a Jacobian error in the \(R_t\) prior. It was right

• 91 issues opened by seabbs-bot on repositories outside my own organisations
• Upstream bug reports land well. Mooncake, DocumenterVitepress and AirspeedVelocity each closed theirs
• Review requests on colleagues’ repositories land less well. This one got “Done”, from me, two days later
• I do not like sending them. I do not love receiving them either

• EpiAware, our Julia packages. This slide is from the JuliaCon 2026 talk about them
• Nobody from outside the team has committed since February 2026, so the other party on an issue is usually one of my robots
• Answering a robot is dispiriting. Do 🤖 shut down collaboration?
• 🌍 We want contributors who are not Julia developers, and not in well resourced settings

• Two large language models, Claude Opus and Claude Sonnet, with no tools. “What is EpiNow2 and who maintains it?”
• Both said I maintain it. I created and developed it
• Opus named three contributors who are not in the author list, and a function removed in 1.9.0
• Sonnet claimed less and got less wrong

• “Who is Sam Abbott, the infectious disease modeller?”
• Right. The PhD, the supervisors, my ORCID researcher identifier and the packages
• Wrong. “Helped run the European Forecast Hub” is a co-authorship. “Senior Research Fellow” is not my job
• Both models had read my email address and said so

• How I am LLM
• epinowcast community forum
• EpiNow2
• The prompts that wrote these decks