Probability of capture

L plays Bayesian detective.

Three suspects. Each evidence card you play multiplies their odds by a likelihood ratio. Watch every P(Kira) column climb — or fall — toward L's 95% arrest line.

L's working theory

Kira's pattern points to Japan, the Kanto region. Multiple suspects fit. We need distinguishing evidence.

Story arc

Belief trajectory · P(Kira)0 plays

Light Misa Matsuda 95% line

Active suspect prior5%

Starting probability before any evidence. Resets on suspect switch.

Evidence cards · likelihood ratios

Each card updates all three suspects by its own per-suspect LR. Badges: ↑ Light · ↗ Misa · ↘ Matsuda (control).

Add your own evidence

LR (Light)

Investigation log

[start] Three suspects loaded. Awaiting evidence.

How to use

Click a suspect card to make them active (the prior slider and verdict follow). Click evidence cards to apply each card's likelihood ratio to all three suspects at once — every column moves in parallel.

Pick an arc (Pre-FBI / Yotsuba / Final) to lock or unlock evidence by canon timing. 🎲 Random priors investigates from a fresh state. Add your own evidence in the custom box below.

The Bayesian math

posterior_odds = prior_odds × LR per piece of evidence. For numerical stability we sum log-LRs in log-odds space instead of multiplying.

An LR of 10 means the evidence is 10× more likely under "this suspect is Kira." LR < 1 is exonerating evidence — it pulls the column down.

Why per-suspect LRs matter

The same evidence updates different suspects by different amounts. "FBI agent dies after tailing X" is strong for Light (he was tailed by Raye Penber) but doesn't apply to Misa or a random officer — they were never tailed.

That's the real detective trick. A flat "this is suspicious" doesn't move probability — only evidence that distinguishes one hypothesis from another does.

L's arrest threshold (95%)

L explicitly says he's "5% sure" early on and "95% sure" later. The dashed line near the top of each column is that 95% mark. When a P(Kira) column crosses it, the suspect card lights red — L acts.

Try to find a sequence of evidence that pushes only Light past 95% while keeping Misa and the control below.

About the LR values

The likelihood ratios attached to each evidence card are illustrative, not empirically measured. There is no published forensic dataset assigning LRs to fictional Death Note events. The values here are hand-picked to produce a defensible, story-faithful posterior trajectory under L's stated 5% → 95% confidence arc — and to teach the shape of Bayesian belief revision, not its forensic ground truth.

Order-of-magnitude calibration was anchored to the canonical guidance in Aitken & Taroni (2004) and Royal Statistical Society (2010): "weak" evidence ≈ LR 2–10, "moderate" ≈ 10–100, "strong" ≈ 100–1000. We never use LRs above 1000 because in real casework that level effectively means a unique match. Treat the numbers as defaults to perturb, not as truth.

Citations

Bayes (1763) — "An essay towards solving a problem in the doctrine of chances," Phil. Trans. R. Soc.
Aitken & Taroni (2004) — Statistics and the Evaluation of Evidence for Forensic Scientists, 2nd ed., Wiley. LR verbal-scale conventions.
Royal Statistical Society (2010) — "Practitioner Guide No. 2: Assessing the probative value of evidence."
Kahneman & Tversky (1973) — "On the psychology of prediction," Psychol. Rev. 80, 237. Base-rate neglect.
Robertson, Vignaux & Berger (2016) — Interpreting Evidence, 2nd ed., Wiley. Bayesian forensic framework.

Built by Joshua Chua · Fable rebuild · vanilla JS + Canvas