Shoshin: a beginner's mind updates on the record

Origins: the hypotheses, and what they were called → ← COVID-19: a sourced timeline ← What the US government said

Someone with no stake in the answer sits down with the documents on the other two pages, in order, and updates one belief as each one arrives. They start at exactly fifty-fifty, because they have not decided anything yet. That is the whole of their method, and you can take it away from them at any point by changing the numbers.

This animation is not tied to a fixed dataset. It reads the other two pages live. Every time an event is added, corrected or removed over there, this reasoning changes here, automatically, with no edit to this page. A new event enters the walk immediately with a weight of exactly 1.00 — no evidential effect — until somebody records a judgment about it and says why. The version of the data being reasoned over is stamped underneath.

All of it is open source: the sources, the criteria, the argument, the process and the prompts. The weights below are the only opinion on this page, they are written down as numbers rather than as adjectives, and you can move every one of them.

The process and the prompts This page's source The repository

Shoshin is the beginner's mind: approaching something without the expertise that tells you in advance what you are going to find. Here it means a prior of 0.50 and no other commitments. The figure is deliberately unassuming — she is not an authority, she has no credentials, and nothing on this page depends on trusting her.

The arithmetic is one line. Belief is carried as odds; each document multiplies the odds by a weight; the bar shows the result as a probability. A weight above 1 makes the hypothesis more likely, below 1 less likely, exactly 1 changes nothing. Nothing else happens.

The walk moves in both directions, and it is supposed to. The heaviest single weight on this page is not against the authorities — it is the CDC reversing a popular guidance change within days of receiving contrary data, at obvious political cost, which is exactly what the hypothesis predicts and is weighted accordingly. A model that only ever ratchets one way is not a model.

The weights are ours and they are contestable, which is why they are exposed as sliders and listed with reasons in the ledger below. Set them all to 1.00 and the belief never moves from 0.50 — that is the shape of the claim that nothing here is evidence of anything.

Beginner's mind.
No prior commitments.

P(H) where H = "the authorities in this record were following the evidence available to them at the time"

50.0%

0% — following something else following the evidence — 100%

Independence discount 0.60 1.00 treats every document as an independent observation, which they are not. Lower values shrink every weight toward 1.00 together.

Loading the data from the other two pages…

The same number, drawn against the calendar

The running column below, as a line. It is flat between documents because belief only moves when one arrives, and steps at each one. Blue is above the starting 0.50, red below. The chart stays pinned while you scroll the ledger, so you can watch a weight move the line as you drag it. Click a point to send the reader there.

Running P(H) 50.0%

The ledger: every document, its weight, and why

This is the whole model. Move a slider and the walk re-runs against your number instead of ours. Rows marked no judgment recorded are events that exist on the timelines but that nobody has assessed yet; they sit at 1.00 and change nothing.

Date	Document	Weight	Why that weight (ours)	Running P(H)

The strongest objections to this whole exercise

These are not strawmen. This page is the most contestable thing in the repository, and it should be.

The weights are the argument, and we chose them. A Bayesian frame looks like arithmetic, and arithmetic looks like objectivity, but every conclusion here is carried by numbers we assigned by judgment. The maths is window dressing on an opinion unless you move the sliders yourself — which is why they are sliders.
The events are not independent, and the maths assumes they are. Multiplying weights is only valid for independent observations. The California closures are four views of one decision; the vaccine authorizations share regulators and evidence. Treating them as separate updates overstates how much the belief should move, in whichever direction it moves. This objection is real enough that it is a control rather than a paragraph — the independence discount above shrinks every weight toward 1.00 together, and it ships set to 0.60 rather than 1.00 because we think the naive walk is wrong. Set it to 1.00 to see what we are discounting. Even 0.60 is a guess: the honest treatment would group correlated documents explicitly, and that is a change worth sending us.
The hypothesis is too coarse to be true or false. "The authorities were following the evidence" bundles dozens of different bodies, in different jurisdictions, over three years, under wildly different information. There is no single fact of the matter for the number to converge on.
Selection still dominates. This page inherits the other two pages' selection of events, and a Bayesian walk over a selected sample tells you about the selection at least as much as about the world.
A number implies a precision nobody has. Reporting a posterior to one decimal place is a rhetorical move. The honest output of this exercise is a direction and a rough size, and the page should probably say "somewhat more likely than not" rather than a percentage.

How to argue with this page

Move the sliders — that is the argument, and it costs nothing. If you want the change to stick for everyone else, change the numbers in the file and send it back; the weights live in a plain table at the top of this page's source, next to the reasons. Adding a reason for one of the unjudged rows is the most useful contribution here, and disagreeing with a weight we set is the second.

The weights, in the source Send us a change Take your own copy Or just tell us what's wrong