“Prediction in the Time of the Coronavirus” — @CedricChin
Why this is in the vault
Cedric’s forecasting series — primarily concerned with noise reduction over bias correction (the Kahneman/Sibony/Sunstein argument). Relevant to how RDCO frames quality in agent deployments: most failures are noise (inconsistent execution across runs), not bias (wrong target).
The core argument
A retrospective on Cedric’s own early-2020 COVID predictions and what he got right/wrong. The piece is more about prediction methodology than the specific calls — what kind of evidence updated him fastest, where his priors blinded him, what process changes he made afterward.
Mapping against Ray Data Co
Direct mapping to agent-eval methodology: when we run multiple deployments of the same agent, the variance across runs IS the metric. Most LLM failure modes are high-noise low-bias — the agent is roughly right on average but wildly inconsistent. This frames Sanity Check posts on eval design and our client-facing ‘agent reliability audit’ offer.
Related
- 2026-04-15-commoncog-process-behaviour-charts
- 2026-04-15-commoncog-two-types-of-data-analysis
- 2026-04-15-commoncog-no-truth-in-business-only-knowledge
Source: Prediction in the Time of the Coronavirus by Cedric Chin (Commoncog). 2760 words. Filed 2026-04-19 as part of Start-Here + Business-Expertise-Triad backfill cohort.