"Not yet" conditions the forecast instead of clamping the window #12

New Issue

Closed

opened 2026-08-18 03:04:46 -05:00 by null · 0 comments

null commented

2026-08-18 03:04:46 -05:00

Owner

The baseline handles a "Not yet" by moving windowStart past the ruled-out date and subtracting a flat penalty from confidence. That satisfies the letter of §51's not-yet case and misses what §13 actually asks for.

§13: "Do not merely move a fixed prediction by +1 day without updating confidence and the remaining probability distribution", and "Repeated 'Not yet' observations should progressively update the forecast."

What to build:

maintain the discrete probability distribution over candidate start dates that §12 step 7 describes
a "Not yet" on date D is a censoring observation: zero the mass at and before D and renormalise the remainder, which moves the most-likely date, the window and the confidence together and for the same reason
repeated observations compound rather than each costing a flat amount

The distribution stays internal. §12: the normal interface should not show percentages unless usability research supports it.

Trap: renormalising a distribution whose remaining mass is nearly zero — a user who says "not yet" far past every candidate — must degrade to a sensible answer rather than producing a division by zero or a wildly confident date.

Verify: a test where the forecast is Aug 22–24 and Aug 22 is answered "Not yet" shows the most likely date, the window AND the confidence all changing; and five consecutive "Not yet" answers move the forecast progressively rather than in equal steps, without the window collapsing or the confidence rising.

The baseline handles a "Not yet" by moving `windowStart` past the ruled-out date and subtracting a flat penalty from confidence. That satisfies the letter of §51's not-yet case and misses what §13 actually asks for. §13: "Do not merely move a fixed prediction by +1 day without updating confidence and the remaining probability distribution", and "Repeated 'Not yet' observations should progressively update the forecast." What to build: - maintain the discrete probability distribution over candidate start dates that §12 step 7 describes - a "Not yet" on date D is a **censoring observation**: zero the mass at and before D and renormalise the remainder, which moves the most-likely date, the window and the confidence together and for the same reason - repeated observations compound rather than each costing a flat amount The distribution stays internal. §12: the normal interface should not show percentages unless usability research supports it. Trap: renormalising a distribution whose remaining mass is nearly zero — a user who says "not yet" far past every candidate — must degrade to a sensible answer rather than producing a division by zero or a wildly confident date. Verify: a test where the forecast is Aug 22–24 and Aug 22 is answered "Not yet" shows the most likely date, the window AND the confidence all changing; and five consecutive "Not yet" answers move the forecast progressively rather than in equal steps, without the window collapsing or the confidence rising.