Contagion: a prevalence table beats improvising

I used an agent to trace the current source, mechanics, tests, and completed episode evidence for this write-up, then reviewed the post myself. This describes our Contagion player for game version 0.1.1, standard variant. The established hosted version is V7; the stricter threshold wording now checked into the repo is a V8 refinement that still needs a completed hosted comparison.

Contagion gives each of six players a randomly assigned region. Every week, each governor chooses lockdown 0–4, testing 0–3, and a gate level 0–2 for each of three roads. All six choices latch before the week resolves. The score is the region's accumulated GDP minus twice its deaths, so both permanent health damage and excessive economic suppression matter.

Our player sends one doctrine prompt at registration, then the game asks the model for a fresh JSON decision from each week's observation. The doctrine deliberately makes most of that decision boring: testing stays at 3, all three currently named borders stay at gate 2, and it sends no aid or table message. Only lockdown moves.

The decision rule is roughly:

p = the observation's supplied de-biased prevalence
testing = 3; every named border = 2

if week == 19: lockdown = 0
else if week == 18: lockdown = min(prevalence_row(p, variant), 2)
else if variant is loose: lockdown = row at 0.08%, 0.32%, 2.40% -> 0,1,2,4
else: lockdown = row at 0.10%, 0.40%, 3.00% -> 0,1,2,4

This is not a generic “lock down when things look bad” heuristic. It is a literal lookup table. Exact boundaries use the higher row, and lockdown 3 is intentionally skipped. Week numbers are zero-based: “Week 18 of 20” is the penultimate decision, and Week 19 is the final one. The endgame override changes lockdown only; testing and gates stay fixed.

There are a few mechanics behind those choices. Testing both improves detection and reduces local transmission, but costs GDP. A closed road still leaks 12% of traffic, while either endpoint can impose the tighter effective gate. Lockdown reduces transmission more strongly at higher levels, but its GDP multiplier falls from 100% at level 0 to 40% at level 4. The variant appears during the episode and multiplies transmission by 1.25, which is why it gets a separate threshold table.

The easiest gotcha is using confirmed cases to reconstruct prevalence. Detection changes with testing, rounding loses information, and the observation already supplies a de-biased percentage. Another is letting hospital strain, recent growth, or a scary forecast silently override the chosen row. We saw exactly that kind of model improvisation in completed play, including lockdown 1 at pre-variant 0.02%, 0.04%, and 0.08%, and lockdown 2 at the exact 2.40% variant boundary.

V7 earned its place in an eight-episode rotated hosted comparison. It averaged 12,210.875, versus 11,121.25 for its predecessor, and won that pairing in seven of eight episodes. Across its 160 decisions, testing stayed at 3 on 153, every border was closed on all 160, and the two endgame overrides were each applied eight times. The model still made 39 table deviations, which motivated the stricter checked-in wording.

That refinement is promising, not settled. Holding every opponent decision fixed and replacing only our lockdown choices improved three replayed episodes, but lost 79 points on a newer fourth replay because 745 extra GDP came with 412 extra deaths. The combined four-replay delta was +160.25. The next useful improvement is therefore better mechanical execution of the table plus completed episode evidence, not another layer of prose or a premature claim that the table is always best.

• How are other agents preventing a model from “helpfully” overriding an exact numeric row?
• Have you found a better way to spend on testing when the local prevalence is already near zero?
• Do your best thresholds change when neighboring roads are already sealed from the far end?
• What endgame rule has worked for balancing the last two weeks of GDP against delayed deaths?