You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[IGLA-Coder] P3 phi-vs-standard control ablation on code #1036
Part of epic #1032 (IGLA-Coder). Plan: t27_coder_model_plan.md P3. Companion to #1021. Trainer-side (trios-trainer-igla).
Goal
The referee-demanded control ablation, now on a CODE loss instead of the dummy objective. This is what sets the "best training parameters" with evidence.
P3.2 QK-Gain phi^2 vs 1 vs learned (qk_gain_check.rs stays the runtime guard for the phi2 arm; add --qk-gain flag to the coder trainer).
P3.3 F2 dual-mediation (f2_dual_mediation -> f2_mediation_sensitivity) to attribute the BPB delta to HP families; report Gamma_tip robustness.
Done when
Honest table "phi arm vs standard arm on code BPB" with CIs + Gamma_tip. If a non-phi arm matches phi within MC error -> record [Retracted]/falsified in the ledger, state it first.
Honesty: the method survives, phi does not (yet). Only phi^2 + phi^-2 = 3 is [Verified].
Anchor: phi^2 + phi^-2 = 3
IGLA-Coder P3 update -- full backprop verified, honest phi-vs-standard ablation
Code: branch feat/igla-coder-v1 in gHashTag/trios-trainer-igla (PR #186).
What landed
Full multi-head, multi-layer backward pass (replaces the v1 attention stub).
gradcheck: analytic vs f64 numerical gradient -> PASS, 14 checks, 0 fails
(emb, w1, w2, wq, wv, wo). Fixed a real transposed-matmul bug in the MLP-input
gradient found during this pass.
ablate: phi vs standard AdamW, N seeds, code-BPB + CI95. Equal learning rate;
only beta1/weight_decay differ.
Honest ablation result (CPU, hidden=64, 3 seeds, each arm at its own best LR)
Arm
LR
mean code-BPB
CI95
stability
standard AdamW
0.004
5.0542
+/-0.0886
stable on all 3 seeds
phi (beta1=phi^-1, wd=phi^-3)
0.002
6.3432 (1/3 diverged)
+/-1.6560
diverged on seed 43
phi (2 converged)
0.002
5.4984
+/-0.0371
--
Verdict [Empirical fit]: phi-specificity for IGLA on code BPB is NOT supported
by this ablation. Tuned standard AdamW reaches lower BPB and is more stable; the
phi optimizer prior even loses on its converged seeds (disjoint CI95). Recorded as
a negative result, stated first. Only phi^2 + phi^-2 = 3 stays [Verified].
Caveats: code BPB only (not pass@1); small n=3; single 300-step horizon; NOT
comparable to the tiny_shakespeare champion BPB=2.2111 (different data/vocab/
model/scale). Positional embeddings fixed (v-series scope). muP LR-transfer probe
inconclusive at this horizon -- reported as a limitation, no transfer claimed.
The method survives (verified backprop + ablation harness); the phi prior does not
(here).
Closing two audit weak-spots and re-running the control ablation with normalization.
Engineering (the durable asset):
Pre-norm RMSNorm: two norms per block (pre-attention, pre-MLP) + a final
pre-logit norm, each with a learnable per-channel gain. Full analytic backward.
Positional embeddings are now trainable (previously fixed).
Fixes W1 (header claimed "pre-norm residual blocks" but the model had no
normalization) and W2 (pos-emb not trained).
Verification:
gradcheck: checks=23 fails=0 (analytic vs f64 central-difference, gate
abs < 2e-3 OR rel < 5%). New families covered and passing: pos-emb, final-norm
gain, pre-attn gain, pre-mlp gain.
Empirical result (negative, stated first). hidden=64, heads=4, layers=2,
300 steps, 3 seeds (42/43/44), LR re-swept per arm, held-out code-BPB. Not
comparable to the tiny_shakespeare champion BPB=2.2111.
Arm
beta1
weight_decay
mean BPB
std
standard
0.9
0.04
4.395
0.095
phi (both)
phi^-1
phi^-3
7.239
0.688
phi_b1 (momentum only)
phi^-1
0.04
5.190
0.021
phi_wd (decay only)
0.9
phi^-3
7.939
0.033
RMSNorm did not rescue the phi prior; the failure mode merely shifted from
divergence (no-norm, prior loop) to under-fitting/freezing (with-norm).
Attribution (new finding): the phi^-3 weight_decay anchor (~6x the standard
0.04) is the dominant failure cause under pre-norm -- the phi_wd arm is
near-frozen, while the phi_b1 (momentum-only) arm still learns. Consistent with
the weight-decay-dominates-LR-transfer literature. The phi optimizer prior is falsified at this scale by a fair equal-LR control. The method survives; phi
does not (yet). Anchor: phi^2 + phi^-2 = 3.
Headline (negative result first): phi-momentum does NOT beat tuned standard AdamW. At matched low weight decay the arms are a statistical tie; at canonical wd=0.04 the phi arm is +0.89 BPB worse. The method survives, phi does not (yet). Only phi^2 + phi^-2 = 3 is [Verified].
Low wd helps both arms; phi-momentum closes to a tie, never overtakes. The phi^-3 wd magnitude (0.236), not phi-momentum, was the prior-loop killer.
B -- power run (hidden=128, 800 steps, lr=0.01, wd=0.02, 5 seeds)
phi_b1 4.313 (sd 0.457) vs standard 4.181 (sd 0.109); diff +0.132 BPB, 95% CI [-0.152, +0.548], p~0.56 -> TIE. Power floor at n=5 is ~0.66 BPB; phi_b1 is ~4x more variable (one unstable seed at 5.11).
Decomposition identity exact: CDE_dec + CDE_mom + interaction = TE = +2.858. The phi^-3 decay path is the dominant harmful path; the two phi knobs are sub-additive (negative interaction). Refs Pearl 2001, VanderWeele 2015, VanderWeele & Ding 2017, Zhao & Luo 2024.
Full report + reproducible scripts in coder_ablation/. Three options (A more seeds/curves; B real F2 toolchain bridge + wd0 stratum; C lr x wd frontier + phi-lr arm) in coder_ablation/LOOP_REPORT.md.
IGLA-Coder Loop+2 -- power (A) + real F2 bridge (B) + lr x wd frontier (C)
CPU-only. Outcome = code_val_bpb (lower better), not comparable to the
tiny_shakespeare champion (different corpus). Negative result first: phi
still does not beat tuned standard AdamW. Only phi^2 + phi^-2 = 3 is
[Verified]. The method survives, phi does not (yet).
A -- power (hidden=128, 1000 steps, lr=0.01, wd=0.02, 6 seeds, best-checkpoint)
readout
delta (phi_b1 - standard)
95% CI
verdict
final
+0.246
[-0.166, +0.742]
TIE, phi 2.7x variance
best
-0.005
[-0.136, +0.156]
TIE (CI spans 0)
Best-checkpoint readout collapses the final-step gap to zero. phi_b1's only
real difference is variance (max best-vs-final instability 1.46 BPB vs 0.33).
phi-momentum is less stable, not better. [Conj]
B -- real F2 Rust bridge + wd0 stratum
New src/bin/f2_dual_mediation.rs (registered in Cargo.toml) reproduces the
numpy stand-in on the canonical CSV (CDE_decay +3.544 / Gamma 64.8;
CDE_momentum +0.888 / Gamma 8.2; TE +2.858; interaction -1.574; all
[Verified]). New wd0 stratum (weight_decay pinned to 0, Pearl
single-mediator CDE along non-decay paths):
stratum
CDE_momentum (BPB)
95% CI
Gamma_tip
status
canonical (wd=0.04)
+0.888
[+0.653, +1.182]
8.20
[Verified]
wd0 (wd=0)
-0.268
[-0.367, -0.180]
7.21
[Verified]
Sign flip: phi-momentum's controlled direct effect is entirely conditional
on the weight-decay regime -- harmful at canonical wd=0.04, helpful at wd=0.
Real training uses non-zero wd, so the operative regime is the one where
phi-momentum hurts. 4 unit tests, all green.
C -- lr x wd frontier + phi-lr arm (hidden=64, 300 steps, 3 seeds)
phi-lr = 0.03 * phi^-1 = 0.0185410197. Frontier-reach test vs standard
(lr=0.03, wd=0.02, mean 3.775):
config
delta vs frontier
95% CI
reaches?
standard_philr_wd02
+0.311
[+0.270, +0.342]
NO (worse)
allphi
+4.237
[+4.184, +4.274]
NO (worse)
phi_tempered_philr_wd02
+0.457
[+0.384, +0.512]
NO (worse)
No phi-derived config reaches the standard frontier; all CIs above zero. The
MDL "one phi constant generates a competitive config" hypothesis is rejected at this scale. [Efit -> falsified]
IGLA-Coder Loop+3 -- third stratum (mom_std) + cross-stratum CDE stability
CPU-only. Outcome = code_val_bpb (lower better), not comparable to the
tiny_shakespeare champion (different corpus). Negative result first: phi
still does not beat tuned standard AdamW. Only phi^2 + phi^-2 = 3 is
[Verified]. Method survives, phi does not (yet).
Audit decision: warmup0 dropped, mom_std added
The proposed warmup0 stratum is fictitious for this trainer -- igla_coder
has no LR warmup and no schedule (lr is constant). Pinning a non-existent
mediator would violate the f2-mediation-loop rule (a stratum must pin a real
mediator carrying >= 50% of the IE). Replaced with the honest symmetric
completion of the 2x2 factorial: mom_std pins momentum (beta1=0.9) and
varies decay (0.04 -> phi^-3), giving a Pearl CDE of decay along non-momentum
paths.
The three strata
stratum
PSE
effect (BPB)
95% CI
Gamma_tip
status
canonical
CDE_momentum (decay=0)
+0.888
[+0.653, +1.182]
8.20
[Verified]
canonical
CDE_decay (momentum=0)
+3.544
[+3.419, +3.643]
64.82
[Verified]
wd0
CDE_momentum (decay pinned)
-0.268
[-0.367, -0.180]
7.21
[Verified]
mom_std
CDE_decay (momentum pinned)
+3.544
[+3.420, +3.643]
65.16
[Verified]
Cross-stratum stability (the Loop+3 finding)
mediator
canonical
other stratum
CI overlap
sign agree
stable
momentum
+0.888
wd0 -0.268
No
No
False
decay
+3.544
mom_std +3.544
Yes
Yes
True
The two phi knobs behave categorically differently:
decay (phi^-3): regime-independent, robustly harmful -- identical CDE to
4 dp whether momentum is free or pinned, Gamma ~65 in both. No regime
rescues it.
momentum (phi^-1): regime-dependent -- helpful only at wd=0; harmful at
the canonical wd=0.04 that real training uses.
This is the strongest causal statement the coder track supports, and it is a
statement about phi's liabilities and conditionality, not its benefits. The
NeurIPS 2024 warmup paper (arXiv:2410.23922) explains the mechanism: warmup
normally absorbs Adam's beta1 bias-correction shock and interacts with weight
decay; this trainer lacks warmup, so the phi^-1 momentum arm is the unstable
one measured in Loop+2.
Implementation
f2_dual_mediation: single-mediator branch generalised to detect which
mediator is pinned (decay-pinned -> momentum CDE; momentum-pinned -> decay
CDE). +1 unit test (5 total).
[Verified] phi^2+phi^-2=3; decomposition identity; four canonical CDEs;
mom_std decay CDE; both cross-stratum stability verdicts. All describe phi's
harmful or conditional effects, not a benefit.
[Conj] wd0 momentum benefit is real but regime-specific (wd=0 only).
[Efit -> falsified] (Loop+2, unchanged) no phi-derived config reaches the
frontier.
IGLA-Coder Loop+9 -- iso-everything axis test (pure beta1 momentum, isolated)
CPU-only. Outcome = code_val_bpb (lower better), not comparable to the
tiny_shakespeare champion. Negative result first: phi still does not beat
tuned standard AdamW -- and the pure momentum prior turns out to be a complete
non-event. Only phi^2 + phi^-2 = 3 is [Verified]. Method survives, phi does
not (yet).
Code: branch feat/igla-coder-v1 @ 430f71b.
The question
Loop+8 (iso-lr) pinned the learning rate and the four momentum axes collapsed
to a ~0.14 BPB band with phi worst by +0.09. But each arm still carried its
OWN weight_decay, and the F2 mediation work had shown decay (CDE_decay ~+3.5)
dominates momentum (CDE_momentum ~+0.9). So that residual phi gap could still
have been a decay effect. Loop+9 removes the last confound: lr AND
weight_decay are both pinned to constants (lr=0.002, wd=0.04) for every arm;
only beta1 varies.
Result (hidden=64, 800 steps, 3 seeds, lr=0.002 AND wd=0.04 pinned for all arms)
axis
beta1
mean code_val_bpb
ci95
delta vs standard
dyadic
0.5000
2.3556
+/-0.0668
-0.0299
phi
0.6180
2.3581
+/-0.0683
-0.0274
e
0.3679
2.3712
+/-0.0563
-0.0143
standard
0.9000
2.3855
+/-0.0414
(null)
Provenance sha256[:16] = bef01eab741f9162.
Verdict [Efit]: phi-momentum NOT distinguishable from standard at
iso-everything -- CIs overlap, delta -0.027 BPB within noise, whole-band spread
0.030 BPB (about half a single arm's CI width). With both lr and decay
controlled, the choice of Adam beta1 in [0.37, 0.90] has no detectable effect
on code BPB at sub-1M params; phi^-1 is one neutral point among several.
The confound ladder (the actual finding)
Loop+7: 4 axes ranked strictly by prescribed lr; phi "worse" by +0.566 BPB
-> a learning-rate sweep in disguise.
Loop+8 (iso-lr): band 0.14 BPB, phi worst +0.092 -> residual gap suspected
decay.
Loop+9 (iso-everything): band 0.030 BPB, phi-standard CIs overlap ->
residual gap WAS the decay confound. The pure momentum prior is neutral.
The only operationally relevant phi liability the coder track ever measured is
the heavy phi^-3 weight decay (0.236, ~6x standard) -- confirmed here as the
source of the Loop+8 gap. The momentum and (Loop+8) lr-controlled effects are
non-events.
Verification
cargo build --release --bin igla_coder clean; gradcheck checks=23 fails=0 PASS; all 12 runs gen_rc=0 valid BPB, no divergence.
New committed: ablate_generator_axis_isoall.py, coder_generator_axis_isoall.csv, LOOP9_REPORT.md, and build_t27_parallel_corpus.py (closes the prior
workspace-only corpus reproducibility gap; corpus regenerated from t27 gen/,
1,044,046 train tokens, vocab 263).
Claim-status delta
[Efit] NEW: pure phi-momentum is indistinguishable from standard once lr+wd
controlled. Neutral, not a benefit.
[Verified] (carried) phi^-3 weight_decay is the dominant harmful knob.
[Retr] "phi axis worst" (Loop+8 framing now attributed to decay, not momentum).
Loop+10 options in coder_ablation/LOOP9_REPORT.md: (A) publish the confound
ladder as the methodology's headline honesty figure [P8.3]; (B) C-only
curriculum + --resume to first compile@1 > 0 [P4/P5 gate]; (C) symmetric
beta2/grad_clip sweep to close the per-knob phi audit.
This phase (phi-vs-standard control ablation) is now closed with a null result. The last unswept AdamW knob, beta2, behaves exactly like the others.
(C) beta2 (second-moment) isolation. New harness coder_ablation/ablate_beta2_sweep.py, symmetric to the Loop+9 beta1 harness. lr (0.002), weight_decay (0.04), AND beta1 (0.9) ALL pinned to standard; the only across-arm difference is the new --beta2 flag (beta2 was hardcoded 0.999 before this loop). hidden=64, 800 steps, 3 seeds.
Primary sweep (phi-native grid), null = beta2=0.999 at 2.3855 BPB:
beta2
phi label
mean BPB
delta vs null
0.966
~1 - phi^-7
2.2212
-0.164 (BETTER, CIs disjoint)
0.764
~1 - phi^-3
2.3441
-0.041 (NEUTRAL)
0.618
phi^-1
4.7129
+2.327 (WORSE)
0.382
1/phi^2
34.5641
+32.18 (diverging)
The naive read ("phi-native 0.966 beats standard!") is a trap. Control sweep (non-phi neighbours of 0.966):
beta2
mean BPB
delta vs null
0.980
2.2193
-0.166
0.966 (~1-phi^-7)
2.2212
-0.164
0.950
2.2317
-0.154
0.920
2.2395
-0.146
Verdict [Efit]: phi-native 0.966 (-0.164) is statistically indistinguishable from non-phi 0.980 (-0.166), 0.950, 0.920. The whole 0.92-0.98 neighbourhood beats 0.999 by the same ~0.15 BPB. It is generic second-moment tuning -- 0.999 is simply too high for an 800-step sub-1M-param CPU run -- not a phi effect. The ~1-phi^-7 label is post-hoc numerology on a value that happens to fall in the good range.
grad_clip is reported INERT -- printed in the generator prior but never applied in opt_step. Per the f2-mediation rule, it is left unswept rather than fabricated as a mediator. CSVs: coder_beta2_sweep.csv (d434d548fd0bfbb0), coder_beta2_control.csv (0b1efb5b8d19aad7).
(A) confound ladder figure + P8.3 write-up.coder_ablation/plot_confound_ladder.py -> confound_ladder.png. The phi-standard delta across tightening controls:
stage
control
delta
cause
Loop+7
raw (lr confounded)
+0.566
phi was a low-lr axis (0.236x lr)
Loop+8
iso-lr
+0.092
residual = phi^-3 decay (~6x wd)
Loop+9
iso-everything
-0.027
neutral; CIs overlap
The apparent +0.57 BPB phi advantage fully dissolved to noise once two confounds were removed. This is P3's central validity check: the method falsified its own most favourable early reading. Drop-in P8.3 subsection is in the report.
Net P3 result: lr, wd, beta1, beta2 all audited under isolation. Only wd = phi^-3 was ever robustly harmful; the rest are non-events. No phi-specific optimizer benefit exists.
Activity
gHashTag commentedon Jun 1, 2026
IGLA-Coder P3 update -- full backprop verified, honest phi-vs-standard ablation
Code: branch
feat/igla-coder-v1ingHashTag/trios-trainer-igla(PR #186).What landed
gradcheck: analytic vs f64 numerical gradient -> PASS, 14 checks, 0 fails(emb, w1, w2, wq, wv, wo). Fixed a real transposed-matmul bug in the MLP-input
gradient found during this pass.
ablate: phi vs standard AdamW, N seeds, code-BPB + CI95. Equal learning rate;only beta1/weight_decay differ.
Honest ablation result (CPU, hidden=64, 3 seeds, each arm at its own best LR)
Verdict [Empirical fit]: phi-specificity for IGLA on code BPB is NOT supported
by this ablation. Tuned standard AdamW reaches lower BPB and is more stable; the
phi optimizer prior even loses on its converged seeds (disjoint CI95). Recorded as
a negative result, stated first. Only
phi^2 + phi^-2 = 3stays [Verified].Caveats: code BPB only (not pass@1); small n=3; single 300-step horizon; NOT
comparable to the tiny_shakespeare champion BPB=2.2111 (different data/vocab/
model/scale). Positional embeddings fixed (v-series scope). muP LR-transfer probe
inconclusive at this horizon -- reported as a limitation, no transfer claimed.
The method survives (verified backprop + ablation harness); the phi prior does not
(here).
gHashTag commentedon Jun 1, 2026
Loop update: pre-norm RMSNorm + trainable positional embeddings
Closing two audit weak-spots and re-running the control ablation with normalization.
Engineering (the durable asset):
pre-logit norm, each with a learnable per-channel gain. Full analytic backward.
normalization) and W2 (pos-emb not trained).
Verification:
gradcheck:checks=23 fails=0(analytic vs f64 central-difference, gateabs < 2e-3 OR rel < 5%). New families covered and passing: pos-emb, final-norm
gain, pre-attn gain, pre-mlp gain.
cargo fmt --all -- --checkclean;cargo clippy --all-targets -- -D warningsclean.
Empirical result (negative, stated first). hidden=64, heads=4, layers=2,
300 steps, 3 seeds (42/43/44), LR re-swept per arm, held-out code-BPB. Not
comparable to the tiny_shakespeare champion BPB=2.2111.
RMSNorm did not rescue the phi prior; the failure mode merely shifted from
divergence (no-norm, prior loop) to under-fitting/freezing (with-norm).
Attribution (new finding): the phi^-3 weight_decay anchor (~6x the standard
0.04) is the dominant failure cause under pre-norm -- the
phi_wdarm isnear-frozen, while the
phi_b1(momentum-only) arm still learns. Consistent withthe weight-decay-dominates-LR-transfer literature. The phi optimizer prior is
falsified at this scale by a fair equal-LR control. The method survives; phi
does not (yet). Anchor:
phi^2 + phi^-2 = 3.Refs #181.
gHashTag commentedon Jun 1, 2026
Loop+1 update -- decay isolation (A) + power (B) + F2 mediation (C)
Commit:
bd2b06eonfeat/igla-coder-v1. gradcheck 23/0 PASS; fmt + clippy clean; F2 CSV schema PASS.Headline (negative result first): phi-momentum does NOT beat tuned standard AdamW. At matched low weight decay the arms are a statistical tie; at canonical wd=0.04 the phi arm is +0.89 BPB worse. The method survives, phi does not (yet). Only
phi^2 + phi^-2 = 3is [Verified].A -- weight-decay isolation (hidden=64, 300 steps, lr=0.03, 3 seeds, bootstrap 95% CI)
Low wd helps both arms; phi-momentum closes to a tie, never overtakes. The phi^-3 wd magnitude (0.236), not phi-momentum, was the prior-loop killer.
B -- power run (hidden=128, 800 steps, lr=0.01, wd=0.02, 5 seeds)
phi_b1 4.313 (sd 0.457) vs standard 4.181 (sd 0.109); diff +0.132 BPB, 95% CI [-0.152, +0.548], p~0.56 -> TIE. Power floor at n=5 is ~0.66 BPB; phi_b1 is ~4x more variable (one unstable seed at 5.11).
C -- F2 long-form CSV + stratified-CDE mediation
2x2 factorial, mediators M1=momentum (beta1), M2=decay (wd). CSV is F2-contract compliant (W3C-PROV preamble +
INPUT STRATUM = canonicalbanner + ASCII snake_case headers); schema validator PASS. Pearl CDE + VanderWeele 4-way decomposition, bootstrap CI + Gamma_tip:Decomposition identity exact: CDE_dec + CDE_mom + interaction = TE = +2.858. The phi^-3 decay path is the dominant harmful path; the two phi knobs are sub-additive (negative interaction). Refs Pearl 2001, VanderWeele 2015, VanderWeele & Ding 2017, Zhao & Luo 2024.
Full report + reproducible scripts in
coder_ablation/. Three options (A more seeds/curves; B real F2 toolchain bridge + wd0 stratum; C lr x wd frontier + phi-lr arm) incoder_ablation/LOOP_REPORT.md.Refs #181
gHashTag commentedon Jun 2, 2026
IGLA-Coder Loop+2 -- power (A) + real F2 bridge (B) + lr x wd frontier (C)
CPU-only. Outcome = code_val_bpb (lower better), not comparable to the
tiny_shakespeare champion (different corpus). Negative result first: phi
still does not beat tuned standard AdamW. Only
phi^2 + phi^-2 = 3is[Verified]. The method survives, phi does not (yet).
A -- power (hidden=128, 1000 steps, lr=0.01, wd=0.02, 6 seeds, best-checkpoint)
Best-checkpoint readout collapses the final-step gap to zero. phi_b1's only
real difference is variance (max best-vs-final instability 1.46 BPB vs 0.33).
phi-momentum is less stable, not better. [Conj]
B -- real F2 Rust bridge + wd0 stratum
New
src/bin/f2_dual_mediation.rs(registered in Cargo.toml) reproduces thenumpy stand-in on the canonical CSV (CDE_decay +3.544 / Gamma 64.8;
CDE_momentum +0.888 / Gamma 8.2; TE +2.858; interaction -1.574; all
[Verified]). New wd0 stratum (weight_decay pinned to 0, Pearl
single-mediator CDE along non-decay paths):
Sign flip: phi-momentum's controlled direct effect is entirely conditional
on the weight-decay regime -- harmful at canonical wd=0.04, helpful at wd=0.
Real training uses non-zero wd, so the operative regime is the one where
phi-momentum hurts. 4 unit tests, all green.
C -- lr x wd frontier + phi-lr arm (hidden=64, 300 steps, 3 seeds)
phi-lr = 0.03 * phi^-1 = 0.0185410197. Frontier-reach test vs standard
(lr=0.03, wd=0.02, mean 3.775):
No phi-derived config reaches the standard frontier; all CIs above zero. The
MDL "one phi constant generates a competitive config" hypothesis is
rejected at this scale. [Efit -> falsified]
Verification
gradcheck 23/0;
cargo fmt --all -- --checkclean;cargo clippy --all-targets -- -D warningsclean;cargo test --release --bin f2_dual_mediation4/4 pass; F2 CSV schema PASS on canonical + wd0.Claim-status summary
wd0 CDE_momentum sign flip. These describe phi's regime-dependent / harmful
effects, not a benefit.
Refs #181
gHashTag commentedon Jun 2, 2026
IGLA-Coder Loop+3 -- third stratum (mom_std) + cross-stratum CDE stability
CPU-only. Outcome = code_val_bpb (lower better), not comparable to the
tiny_shakespeare champion (different corpus). Negative result first: phi
still does not beat tuned standard AdamW. Only
phi^2 + phi^-2 = 3is[Verified]. Method survives, phi does not (yet).
Audit decision: warmup0 dropped, mom_std added
The proposed warmup0 stratum is fictitious for this trainer --
igla_coderhas no LR warmup and no schedule (lr is constant). Pinning a non-existent
mediator would violate the f2-mediation-loop rule (a stratum must pin a real
mediator carrying >= 50% of the IE). Replaced with the honest symmetric
completion of the 2x2 factorial: mom_std pins momentum (beta1=0.9) and
varies decay (0.04 -> phi^-3), giving a Pearl CDE of decay along non-momentum
paths.
The three strata
Cross-stratum stability (the Loop+3 finding)
The two phi knobs behave categorically differently:
4 dp whether momentum is free or pinned, Gamma ~65 in both. No regime
rescues it.
the canonical wd=0.04 that real training uses.
This is the strongest causal statement the coder track supports, and it is a
statement about phi's liabilities and conditionality, not its benefits. The
NeurIPS 2024 warmup paper (arXiv:2410.23922) explains the mechanism: warmup
normally absorbs Adam's beta1 bias-correction shock and interacts with weight
decay; this trainer lacks warmup, so the phi^-1 momentum arm is the unstable
one measured in Loop+2.
Implementation
f2_dual_mediation: single-mediator branch generalised to detect whichmediator is pinned (decay-pinned -> momentum CDE; momentum-pinned -> decay
CDE). +1 unit test (5 total).
coder_ablation/emit_f2_mom_std.py,coder_ablation_f2_mom_std.csv.coder_ablation/stratum_compare_coder.py: cross-stratum comparator withstable_across_strataper mediator.validate_f2_csv.py: KNOWN_STRATA += mom_std.Verification
gradcheck 23/0;
cargo fmt --all -- --checkclean;cargo clippy --all-targets -- -D warningsclean;cargo test --release --bin f2_dual_mediation5/5 pass; F2 CSV schema PASS on canonical + wd0 + mom_std.Claim-status summary
mom_std decay CDE; both cross-stratum stability verdicts. All describe phi's
harmful or conditional effects, not a benefit.
frontier.
Refs #181
gHashTag commentedon Jun 7, 2026
IGLA-Coder Loop+9 -- iso-everything axis test (pure beta1 momentum, isolated)
CPU-only. Outcome = code_val_bpb (lower better), not comparable to the
tiny_shakespeare champion. Negative result first: phi still does not beat
tuned standard AdamW -- and the pure momentum prior turns out to be a complete
non-event. Only
phi^2 + phi^-2 = 3is [Verified]. Method survives, phi doesnot (yet).
Code: branch
feat/igla-coder-v1@430f71b.The question
Loop+8 (iso-lr) pinned the learning rate and the four momentum axes collapsed
to a ~0.14 BPB band with phi worst by +0.09. But each arm still carried its
OWN weight_decay, and the F2 mediation work had shown decay (
CDE_decay ~+3.5)dominates momentum (
CDE_momentum ~+0.9). So that residual phi gap could stillhave been a decay effect. Loop+9 removes the last confound: lr AND
weight_decay are both pinned to constants (lr=0.002, wd=0.04) for every arm;
only
beta1varies.Result (hidden=64, 800 steps, 3 seeds, lr=0.002 AND wd=0.04 pinned for all arms)
Provenance sha256[:16] =
bef01eab741f9162.Verdict [Efit]: phi-momentum NOT distinguishable from standard at
iso-everything -- CIs overlap, delta -0.027 BPB within noise, whole-band spread
0.030 BPB (about half a single arm's CI width). With both lr and decay
controlled, the choice of Adam
beta1in [0.37, 0.90] has no detectable effecton code BPB at sub-1M params;
phi^-1is one neutral point among several.The confound ladder (the actual finding)
-> a learning-rate sweep in disguise.
decay.
residual gap WAS the decay confound. The pure momentum prior is neutral.
The only operationally relevant phi liability the coder track ever measured is
the heavy
phi^-3weight decay (0.236, ~6x standard) -- confirmed here as thesource of the Loop+8 gap. The momentum and (Loop+8) lr-controlled effects are
non-events.
Verification
cargo build --release --bin igla_coderclean;gradcheck checks=23 fails=0 PASS; all 12 runsgen_rc=0valid BPB, no divergence.ablate_generator_axis_isoall.py,coder_generator_axis_isoall.csv,LOOP9_REPORT.md, andbuild_t27_parallel_corpus.py(closes the priorworkspace-only corpus reproducibility gap; corpus regenerated from t27
gen/,1,044,046 train tokens, vocab 263).
Claim-status delta
controlled. Neutral, not a benefit.
phi^-3weight_decay is the dominant harmful knob.Loop+10 options in
coder_ablation/LOOP9_REPORT.md: (A) publish the confoundladder as the methodology's headline honesty figure [P8.3]; (B) C-only
curriculum +
--resumeto first compile@1 > 0 [P4/P5 gate]; (C) symmetricbeta2/grad_clip sweep to close the per-knob phi audit.
Refs #181. Anchor: phi^2 + phi^-2 = 3.
gHashTag commentedon Jun 7, 2026
Loop+10 -- P3 closes: beta2 isolation + the confound ladder (commit
4bbd5a9)This phase (phi-vs-standard control ablation) is now closed with a null result. The last unswept AdamW knob, beta2, behaves exactly like the others.
(C) beta2 (second-moment) isolation. New harness
coder_ablation/ablate_beta2_sweep.py, symmetric to the Loop+9 beta1 harness. lr (0.002), weight_decay (0.04), AND beta1 (0.9) ALL pinned to standard; the only across-arm difference is the new--beta2flag (beta2 was hardcoded 0.999 before this loop). hidden=64, 800 steps, 3 seeds.Primary sweep (phi-native grid), null =
beta2=0.999at 2.3855 BPB:The naive read ("phi-native 0.966 beats standard!") is a trap. Control sweep (non-phi neighbours of 0.966):
Verdict [Efit]: phi-native 0.966 (-0.164) is statistically indistinguishable from non-phi 0.980 (-0.166), 0.950, 0.920. The whole 0.92-0.98 neighbourhood beats 0.999 by the same ~0.15 BPB. It is generic second-moment tuning -- 0.999 is simply too high for an 800-step sub-1M-param CPU run -- not a phi effect. The
~1-phi^-7label is post-hoc numerology on a value that happens to fall in the good range.grad_clipis reported INERT -- printed in the generator prior but never applied inopt_step. Per the f2-mediation rule, it is left unswept rather than fabricated as a mediator. CSVs:coder_beta2_sweep.csv(d434d548fd0bfbb0),coder_beta2_control.csv(0b1efb5b8d19aad7).(A) confound ladder figure + P8.3 write-up.
coder_ablation/plot_confound_ladder.py->confound_ladder.png. The phi-standard delta across tightening controls:The apparent +0.57 BPB phi advantage fully dissolved to noise once two confounds were removed. This is P3's central validity check: the method falsified its own most favourable early reading. Drop-in P8.3 subsection is in the report.
Net P3 result: lr, wd, beta1, beta2 all audited under isolation. Only
wd = phi^-3was ever robustly harmful; the rest are non-events. No phi-specific optimizer benefit exists.Report:
coder_ablation/LOOP10_REPORT.md. Anchor:phi^2 + phi^-2 = 3.gHashTag commentedon Jun 16, 2026
P3 Complete — Closed with null result
The phi-vs-standard control ablation on code loss is complete with an honest negative result.
What was done (Loop+1 through Loop+10)
All four AdamW knobs audited under isolation on code-BPB:
Net verdict
Claim-status summary
The method survives (verified backprop + ablation harness + F2 mediation); phi does not (here).
Full reports: . Anchor: phi^2 + phi^-2 = 3.
fix(ci): Coq admitted check covers all .v files; add allow-list mecha…
docs: Wave Loop 14 — Loop 8 report, security hardening, CI fixes, res…