Where the frame comes from
The path the manifest's entry names, joined onto the folder predictions are mounted at. Nothing holds the joined path inside that folder, so a manifest naming a path that climbs out of it is read from wherever it leads.
A path leading to no file warns, and the turn ends.
Which rows are candidates
The whole frame is read in, and every row whose reason code is anything but the usable one is dropped.
A frame with nothing left after that ends the turn without a word.
Which candidate wins
The one with the greatest moment, whatever order the frame held its rows in. So the frame is treated as a set rather than a sequence, and a row appended out of order is still found.
The moment it must beat
Strictly above the last moment recorded for that strategy and series. A repeat ends the turn without a word.
What the prediction is held to
It is cut to a whole number, and the result weighed against the two allowed values.
So a figure standing between them is taken as the whole number below it rather than refused, and one that will not cut to a whole number at all raises rather than warning.
Related
Source evidence
---
row: N71-01
baseline: 49aa659
created: 2026-11-08 21:14 UTC
modified: 2026-11-08 21:14 UTC
evidence:
- claim: "The frame is read from the manifest entry's path joined onto the folder predictions are mounted at, with nothing holding the joined path inside that folder"
source: "`nexus/strategy/predict_loop.py` 465, joining what 382 read"
- claim: "A path leading to no file warns and the turn ends"
source: "`nexus/strategy/predict_loop.py` 466-471"
- claim: "The whole frame is read in, and rows whose reason code is anything but the usable one are dropped"
source: "`nexus/strategy/predict_loop.py` 473, then 474"
- claim: "A frame with nothing left after that dropping ends the turn without a word"
source: "`nexus/strategy/predict_loop.py` 476-477"
- claim: "Of what survives, the row with the greatest moment is the one taken, whatever order the frame held them in"
source: "`nexus/strategy/predict_loop.py` 479"
- claim: "That moment must stand strictly above the last one recorded for the same strategy and series, and a repeat ends the turn without a word"
source: "`nexus/strategy/predict_loop.py` 394-397"
- claim: "The prediction is cut to a whole number and then weighed against the two values, so one standing between them is taken as the whole number below it"
source: "`nexus/strategy/predict_loop.py` 399, then 400-405"
- claim: "A prediction that will not cut to a whole number raises rather than warning"
source: "`nexus/strategy/predict_loop.py` 399, standing in no attempt"
---