Skip to content

Commit 7dd2dfb

Browse files
alexarjeclaude
andcommitted
The design records its own wrong prediction: mixtures are burstier than their parts
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GFjgT8igZuQu4EPMvSqK6u
1 parent b47d779 commit 7dd2dfb

1 file changed

Lines changed: 18 additions & 0 deletions

File tree

plans/2026-08-30-fragment-reassociation-design.md

Lines changed: 18 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -68,3 +68,21 @@ Returns per-mover trajectories in the single-person extractor's contract, plus
6868
of scope entirely.
6969
- Any claim about which mover is which *person*: the output says "mover 1" and
7070
"mover 2" per segment, and mapping movers to Ole and Lisa stays a human act.
71+
72+
## Measured on the corpus, same day
73+
74+
Run on the co-located day (1,630 fragments, 159 minutes): with three mover slots
75+
(the third absorbs researchers entering frame), 374 breaks --- 351 "no plausible
76+
mover", 23 ambiguous --- yielding 375 segments, the longest 6.6 minutes. Only one
77+
segment over two minutes has BOTH movers covering most of its span (2.6 minutes at
78+
minute 111): chains are real but thin, and the v2 lever the design named ---
79+
appearance ReID, with its re-decode --- is what more per-dancer coverage costs.
80+
81+
And the downstream check was measured and found to be wrongly designed. The
82+
expectation that the ensemble sits between the two movers held for Space only
83+
(0.12 between 0.07 and 0.15) and failed for Time, Weight and Flow, where the
84+
ensemble sits ABOVE both movers --- correctly: burst concentration and spectral
85+
roughness of a MIXTURE of two movers exceed each component's, because interleaved
86+
burst patterns are burstier than either alone. These indices are not convex in
87+
their inputs, so "between" was never the right prediction. The check to keep is
88+
Space's; the failed three are a lesson about the indices, not about the chains.

0 commit comments

Comments
 (0)