Laborator commited on
Commit
c812bf0
·
verified ·
1 Parent(s): 16e9055

Add the half-year/year total_revealed=417 bucketing caveat and the determinism note

Browse files
Files changed (1) hide show
  1. index.html +10 -0
index.html CHANGED
@@ -316,6 +316,16 @@ input[type=range].lever-input:focus-visible::-moz-range-thumb{box-shadow:0 0 0 4
316
  the model learning to predict the future. What the lever measures is the rate at which
317
  unforgeable signal accumulates — not forecasting skill.
318
  <br/><br/>
 
 
 
 
 
 
 
 
 
 
319
  No live model runs in this Space. Every number, the curve and the table are verbatim
320
  outputs of the original experimental run, bundled into <code>data.json</code>. Full code,
321
  raw per-schedule results and methodology:
 
316
  the model learning to predict the future. What the lever measures is the rate at which
317
  unforgeable signal accumulates — not forecasting skill.
318
  <br/><br/>
319
+ One artefact worth naming, visible in the table: “half year” (180 days) and “year”
320
+ (365 days) both report <b>417</b> revealed outcomes, and half-year’s Brier
321
+ (<b>0.209</b>) comes out slightly better than “2 months” (<b>0.220</b>). This is
322
+ bucketing, not signal. Every reveal flushes the whole accumulated backlog at once, so
323
+ at long periods the count depends on where the reveal dates happen to land inside the
324
+ 17-month stream rather than on the period itself — two different periods can land the
325
+ same number of flushes over the window. Read the corridor as flat, not as a ranking:
326
+ the differences between finite schedules are within this artefact, and the only real
327
+ step in the data is the one to <i>never</i>.
328
+ <br/><br/>
329
  No live model runs in this Space. Every number, the curve and the table are verbatim
330
  outputs of the original experimental run, bundled into <code>data.json</code>. Full code,
331
  raw per-schedule results and methodology: