Updated 2026-06-18, Euler node eu-g7-007 (RTX PRO 6000 Blackwell, 24 h alloc). Autonomous overnight session: Phase 4 — pluck GRPO ("pluck go"), plus two report upgrades (matched-subset reporting, full-detail rollout traces). This tab and the Regression-PPL tab are kept live as results land.

Now running

Experiment	Status	Detail
E1 — pluck v1 GRPO (survival, fermieval, ames, 300 steps, LoRA)	all 3 done & evaluated	parse → 1.00 everywhere; trained pluck ties-or-beats trained MoG on all three (table below).
E3 — cure-exemplar probe (survival, `pluck_cure` prompt)	done	best survival model overall (−1.853, beats MoG −1.99 & v1 pluck −2.10) — but the gain is on uncensored rows and point masses are mis-placed; structure reachable, not correctly used.
E4 — variance control (v1 pluck survival re-run, same config)	done	v1 reruns at −2.10…−2.22 (variance ≈0.1 nat); cure −1.85 beats it beyond variance → the cure gain is real (in the mean/tail; median unchanged).

E5 — cure variance check (2nd pluck_cure survival run) done — reproduced both cure runs (−1.85, −1.92) beat both v1 runs (−2.10, −2.22); gap real (~0.27 nat). E6 — bigger model: 8B (Qwen3-8B pluck, survival) done capacity test answer: NO structure — branch≡0.000 through all 300 steps, like 4B (ncomp~1.8 gaussian mixtures). Best raw-ls −2.06 ≈ 4B. Held-out eval pending GPU. E6 — bigger model: 14B (Qwen3-14B pluck, survival) done & evaluated fits at util 0.45 (~78 GB, no OOM). branch≡0, ncomp collapsed to ~1.0; held-out −1.976 (best pluck, ≈ a 4B MoG).

Model-size scaling on survival (held-out, best ckpt)

model	ls-mean ↑	median ↑	CRPS ↓	cov90	structure
4B pluck	−2.099	−1.793	1.251	0.81	branch≡0, ncomp~1.8
8B pluck	−2.135	−1.823	1.306	0.77	branch≡0, ncomp~1.8
14B pluck	−1.976	−1.767	1.060	0.86	branch≡0, ncomp~1.0

Capacity hypothesis rejected across 4B/8B/14B: a bigger model does not unlock pluck's structure (branch≡0 everywhere; 14B uses the least). Scaling 3.5× buys ~0.12 nat (14B pluck ≈ 4B MoG) — diminishing calibration gains, no structural emergence. The limit is the reward, not model size. GPU now free.

Phase 4 campaign conclusion (E1–E4, E5 confirming). Trained pluck ties-or-beats trained MoG everywhere as an exact-inference engine; the cure exemplar gives a real extra gain but the model never deploys structure correctly — realizing PPL structure needs a reward that pays for it. Full write-up in the Regression-PPL tab and 5-regression-ppl/phase4_log.md.

Headline — trained pluck vs trained MoG (matched, full held-out 200)

dataset · series	ls-mean ↑	median ↑	abs-err ↓	CRPS ↓	cov 90%
survival · trained MoG	−1.99	−1.79	2.61	1.71	0.83
survival · trained pluck	−2.10	−1.79	1.79	1.25	0.81
fermieval · trained MoG	−7.11	−1.62	4.20	3.52	0.78
fermieval · trained pluck	−4.93	−1.61	3.26	2.80	0.80
ames · trained MoG	−0.567	−0.440	0.314	0.229	0.955
ames · trained pluck	−0.555	−0.421	0.304	0.220	0.950

Survival: trade-off (tied median log-score, MoG +0.1 nat mean, pluck −31% abs-err / −27% CRPS). fermieval: pluck wins outright on the full holdout (its single wide gaussian is adaptive for heavy tails). In both, pluck never uses point-mass / factor structure — it collapses to the gaussian subset MoG covers, so the program format helps as an inference engine, not via richer structure. Full tables + the censored-row structure probe in the Regression-PPL tab.

Done this session

Item	Status	Outcome
Matched-subset reporting (project-wide standard)	done	`matched_report.py`; Phase 3 table rebuilt on matched subsets + parse rates
Full-detail expandable rollout traces	done	each example: prompt → reasoning → gold → completion; good + both failure modes
Pluck GRPO trainer built + self-tested	done	`reward_pluck.py` (caches inference, ~1 ms/rollout), `train_pluck.py`, `pluck_session.sh`, `eval_pluck.sh`
E2 — pluck_v2 improved-prompt baseline	negative — dropped	v2 prompt HURT zero-shot (survival parse 0.70→0.33; log-space framing confused the model). Cheap baseline killed the v2 GRPO branch, saving ~1.8 h GPU.

Findings so far

GRPO lifts pluck parse rate on its own (survival 0.70 base → 1.00 by step ~26), just like MoG — no format curriculum needed.
Trained pluck ≈ a MoG expressed as a program. Reading rollouts (not just the metric): 28/32 programs at step 50 use flip/if, but they branch between two gaussians → a gaussian mixture (ncomp≈1.8). The model does not use point-mass "cure" leaves (the one structure a MoG can't express). So the open question sharpens to: does a program-mixture beat an XML-mixture when both are trained?
Prompt complexity ≠ better (E2): the base model needs fewer rules, not more; GRPO, not prompting, is what fixes the variadic-+. and division failures.

Queued (to keep the GPU busy)

Merge + held-out 200-cell eval for each pluck dataset as it finishes (eval_pluck.sh).
Trained-vs-trained MoG vs pluck on matched subsets — the headline Phase 4 result.
Structure probe: on censored survival rows a point-mass cure leaf is exactly right and a gaussian MoG cannot express it — does trained pluck use it, or leave the free lunch? (Split survival eval by censored vs not.)
If time: one 600-step pluck run on fermieval (the multiplicative / log-space-sum case where pluck's (+.) composition could most plausibly beat MoG).

openestimate dropped as a 4th training dataset — only 256 examples, 56 left after the 200-cell holdout. Campaign log: 5-regression-ppl/phase4_log.md.