Thibaut Forest Claude Opus 5 (1M context) commited on
Commit
9d44902
·
1 Parent(s): a96ad81

Frontier board update: wiener20 validated piece, vrel count-entanglement caution, md HGB-vrel leader

Browse files
Files changed (1) hide show
  1. agent_reports_bgkf/FRONTIER.md +34 -17
agent_reports_bgkf/FRONTIER.md CHANGED
@@ -1,22 +1,39 @@
1
  # FRONTIER BOARD (maintained by the coordinator - agents READ before each new candidate)
2
 
3
- Updated: 2026-08-22 21:2x. Baseline: champion logbag/base = 1760.2 (swaps 9.5).
4
 
5
- ## Current best composable base
6
- - **variant='vrel'** (fleet-relative voltage + vmin7_rank): -66.8 +/- 41.4, 32W/15L,
7
- swaps flat (v3-vintage read, ft lane re-confirming). COMPOSE ON THIS where
8
- compatible: cv.run_bgkf(variant='vrel', kind=<yours>) or (variant='vrel',
9
- blend=(<yours>, w)). variant/kind/blend all compose in one call.
 
 
 
10
 
11
- ## Confirmed reads on this instrument (paired vs 1760.2 on 48 origin-sums)
12
- - logbag_rank -42.0 +/- 37.0 | ranks_add -37.5 | frontier(lbr+gbm30) -26.7 +/- 39.6 n.s.
13
- - rf -23.6 +/- 63.5 | rankavg25rf -4.5 | gbm +49 | dthaz +23 | logistic +26 | poolown +50.8
14
- - Quarantined (count-coupled): alpha15solo -144 @ swaps 15.8; alpha+cliff -77 n.s.
 
 
 
15
 
16
- ## Cross-lane rules
17
- 1. Before designing each new candidate, re-read this file.
18
- 2. Prefer composing your idea ON the current best base over the plain champion.
19
- 3. Any result of yours with delta < -1.5*SE and quarantine-clean: ALSO run its
20
- composition with the frontier base (one extra ~60s run) and note both in
21
- your findings file.
22
- 4. The coordinator updates this board as lanes report; expect it to change.
 
 
 
 
 
 
 
 
 
 
 
 
1
  # FRONTIER BOARD (maintained by the coordinator - agents READ before each new candidate)
2
 
3
+ Updated: 2026-08-22 ~21:40. Baseline: champion logbag/base = 1760.2 (swaps 9.5).
4
 
5
+ ## Standings (paired on 48 origin-sums vs 1760.2; * = >= 2 sigma and count-clean-pending-checks)
6
+ - md_bl_hgbd1l25_vrel -117.8 +/- 46.1 (2.6s, 30W/18L, swaps 9.2) * current leader
7
+ - md_hgb_d1_vrel -119.5 +/- 63.5 (1.9s) high variance twin
8
+ - bl_v_wnr20 -100.9 +/- 41.4 (2.4s, 36W/12L, swaps 9.3) * vrel + wiener .20
9
+ - ft_triple_rtau -96.5 +/- 37.5 (2.6s, 34W/13L, swaps 9.0) *
10
+ - ft_vrel_rstale -91.4 +/- 38.0 (2.4s, 37W/11L) *
11
+ - bl_wnr20 (champion+wiener.20) -50.3 +/- 19.7 (2.55s, count-clean corr -0.005) VALIDATED PIECE
12
+ - vrel alone -66.8 +/- 41.4 -- SEE CAUTION BELOW
13
 
14
+ ## CAUTION - vrel is count-entangled (bl lane finding, needs ft adjudication)
15
+ corr(count delta, cost delta) = -0.544 for vrel; on count-flat scenarios vrel
16
+ reads +36.4 (WORSE). Its -66.8 may be a hidden appetite channel (the exact
17
+ phase-56 frontier failure mode). EVERY vrel-based leader inherits this doubt
18
+ until the count-matched (value-anchored) vrel twin is adjudicated - assigned
19
+ to the ft lane. The WIENER increment is independently clean (additive
20
+ -34..-50 on every base with corr -0.005).
21
 
22
+ ## Validated composable pieces
23
+ - wiener FHT order blend w=0.20: additive on champion (-50.3), vrel (-34.1
24
+ increment), logbag_rank (-40.7 increment). Compose it onto ANY base:
25
+ blend=('wiener', 0.20).
26
+ - md's depth-1/monotone HGB blend (hgbd1l25) on vrel: current leader; needs
27
+ wiener composition + count-matched check.
28
+
29
+ ## REFUTED this campaign (do not revisit)
30
+ - building-bagged/fold-ensemble fitting (+41..+44 all forms)
31
+ - confidence-weighted local blending (flat global weight dominates)
32
+ - calibration/quantile-mapping of p42 (+406, count-coupled; absolute p42
33
+ level carries prevalence info the count rule needs)
34
+ - tau partner (toxic), dthaz partner (inert), rf/gbm partners on champion (~0)
35
+
36
+ ## Cross-lane rules (unchanged)
37
+ 1. Re-read before each new candidate. 2. Compose on the leader where
38
+ compatible. 3. Wins beyond 1.5*SE: also run frontier-composed. 4. Winners
39
+ with ANY count movement: run the value-anchored twin before believing.