agent commited on
Commit
cef9a1f
·
1 Parent(s): 7735a06

RESEARCH_LOG: prod_v1 complete - the flat-surface finding (absolute-scale neutrality across volume/objective/train-depth/infer-depth on final), the decoupling decode, and the eliminated-lever table

Browse files
Files changed (1) hide show
  1. docs/RESEARCH_LOG.md +48 -0
docs/RESEARCH_LOG.md CHANGED
@@ -3168,3 +3168,51 @@ at push-every 10.
3168
  rebuild kills jobs (never push mid-run; exfil-first always); artifact
3169
  predicates vs state predicates for launches; cp1252 everywhere needs
3170
  PYTHONIOENCODING=utf-8. Full incident ledger in the day-file diary.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3168
  rebuild kills jobs (never push mid-run; exfil-first always); artifact
3169
  predicates vs state predicates for launches; cp1252 everywhere needs
3170
  PYTHONIOENCODING=utf-8. Full incident ledger in the day-file diary.
3171
+
3172
+ ### prod_v1 COMPLETE: the flat surface finding (2026-11-19 morning)
3173
+
3174
+ **The run**: fastwide + 1200 iters (~77k games) at the sweep-verdict recipe
3175
+ (steps-4 + MC), 2.64h. Internal excellence throughout: eval-vs-init peak
3176
+ 0.963 (ties project record, D3-lean), v_mae 0.728 -> 0.468 best, self-play
3177
+ decisive rate ROSE monotonically (0.674 -> 0.743), capture-seeking
3178
+ non-greedy policy character (23/30 argmax captures; 10/23 agree with
3179
+ material-greedy).
3180
+
3181
+ **The central finding - the absolute scale did not move**:
3182
+ - prod_v1@1200 vs SF@1320 pooled n=32: 0.047 (3 draws) vs
3183
+ fast_wide@600's 0.042 (n=12) - statistically indistinguishable
3184
+ (~705 vs ~775 point estimates, overlapping wide CIs)
3185
+ - 2x volume was absolute-scale-NEUTRAL at this recipe/architecture
3186
+
3187
+ **The decoupling decoded**: the draw-heavy final cells (D13/D14) correlate
3188
+ with entropy floor lows (2.01-2.09); self-play decisive rate still rose -
3189
+ the internal improvements are real but live in a closed self-play
3190
+ universe that cannot see what SF exploits.
3191
+
3192
+ **The depth lever died on the final model**: gen600's +0.117 depth-extrapolation
3193
+ gain (steps 4->20 vs material) measured -0.008 on prod_v1@1200 (both CIs
3194
+ overlap) - the phase-3 entropy sharpening consumed it. AND: prod_v1@steps-4
3195
+ beats material 51/60 with zero losses (0.925) - the material reference has
3196
+ saturated too; internal references are nearly exhausted as discriminators.
3197
+
3198
+ **The eliminated-lever table** (all MEASURED null on the absolute scale):
3199
+ volume (600->1200: flat), TD(lambda) (swept: loses + late-sag),
3200
+ training depth (swept: no gain at 40 iters), inference depth on the final
3201
+ model (dead by sharpening).
3202
+
3203
+ **The binding constraint** (measured, not inferred): the representational gap
3204
+ - free-capture 0.66 (a third of outright gifts declined), the rank-4/5
3205
+ attention dead zone, self-referential optimization blind to external
3206
+ exploitation.
3207
+
3208
+ **The Letter**: C (final cell CI 0.695 < 0.83 bar) - and the diagnosis shows
3209
+ the letter's original question (resume?) is answered by the flat-surface
3210
+ finding: resume is the wrong lever. The real options are the s8-jitter
3211
+ restoration probe (cheap, staged), the structural attack on the
3212
+ representation gap (high ceiling), or ceiling acceptance.
3213
+
3214
+ **Ledger of the night's incidents** (all resolved, all logged): 5 loss-spikes
3215
+ (all single-batch artifacts, 5-for-5), the hub rate-limit self-saturation
3216
+ (exfil covered everything), the chain wedges/duplicates (manual endgame
3217
+ takeover, letter computed by pre-committed rule), the rebuild stall on the
3218
+ transformers dep (still pending at morning; probe + experiments await it).