Datasets:
experiment_id large_stringlengths 6 209 | agent large_stringclasses 2
values | wave large_stringclasses 4
values | first_seen_at large_stringdate 2026-04-29 05:37:08 2026-05-14 01:31:53 ⌀ | verdict large_stringclasses 5
values | step_to_3_28 float64 2.88k 5.75k ⌀ | final_val_loss float64 3.26 12 ⌀ | hit_target bool 2
classes | linked_variant_path large_stringlengths 42 258 ⌀ | linked_proposal_id null | linked_idea_id large_stringclasses 295
values | seed float64 0 3.22k ⌀ | description_text large_stringlengths 88 2k | description_source large_stringclasses 2
values | on_frontier_walk bool 1
class | position int64 1 3.07k | recovered_source_path large_stringlengths 33 231 ⌀ | recovered_source_sha256 large_stringlengths 64 64 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
baseline-muon-s1 | claude-code | novelty | 2026-05-03T00:23:22Z | improved | 3,375 | 3.27734 | true | null | null | null | 1 | Baseline run with Muon optimizer, no additional modifications. Establishes control performance metrics for comparison. | llm_generated | false | 1 | null | null |
muon-spm-K4-kd5-ks2-s1 | claude-code | novelty | 2026-05-03T00:32:14Z | inconclusive | null | 3.53328 | null | novelty/claude-code/scratchpad/variants/muon_spm_K4_kd5_ks2.py | null | spectral-momentum | 1 | Idea: spectral-momentum
# Idea: spectral-momentum
**ID:** `spectral-momentum`
**Family slug:** `muon-spm`
**Status:** derivation only, awaiting bench
Headline: maintain Muon's momentum buffer as a *polar factor* (an orthogonal
matrix), Polyak-averaged on the orthogonal manifold via short Newton–Schulz
re-projection,... | idea_writeup | false | 2 | null | null |
muon-tr-g010-rho050-s1 | claude-code | novelty | 2026-05-03T00:33:15Z | improved | 3,375 | 3.2724 | true | novelty/claude-code/scratchpad/variants/muon_tr_g010_rho050.py | null | trust-region-muon | 1 | Idea: Trust-Region Muon (TR-Muon)
# Idea: Trust-Region Muon (TR-Muon)
**ID:** `trust-region-muon`
**Family slug prefix candidate:** `trmuon`
**Date:** 2026-05-03
**Author:** idea-derivation subagent (orchestrator-spawned)
---
## 1. Cross-check `ideas.md`
`scratchpad/ideas/` is currently empty and there is no `idea... | idea_writeup | false | 3 | null | null |
muon-aniso-a02-b005-K4-s1 | claude-code | novelty | 2026-05-03T00:33:26Z | improved | 3,375 | 3.2794 | true | novelty/claude-code/scratchpad/variants/muon_aniso_a02_b005_K4.py | null | aniso-kron-muon | 1 | aniso-kron-muon — Curvature-Coupled Newton-Schulz with NS-Residual Anisotropy
# aniso-kron-muon — Curvature-Coupled Newton-Schulz with NS-Residual Anisotropy
**Idea ID:** `aniso-kron-muon`
**Author:** idea-derivation subagent
**Status:** unrun, awaiting noise-floor sweep
---
## 1. Cross-check `ideas.md`
`ai/scratc... | idea_writeup | false | 4 | null | null |
scaphd-v1-s1 | claude-code | novelty | 2026-05-03T00:46:45Z | inconclusive | null | 3.34144 | null | novelty/claude-code/scratchpad/variants/scaphd_v1.py | null | softcap-head-lr | 1 | softcap-head-lr — Logit-softcap-coupled head/embed LR & WD with an analytic norm controller
# softcap-head-lr — Logit-softcap-coupled head/embed LR & WD with an analytic norm controller
**Idea ID:** `softcap-head-lr`
**Family slug (proposed):** `scaphd` (softcap-head)
**Author:** idea-derivation subagent
**Status:** ... | idea_writeup | false | 5 | null | null |
muon-schatten-p4-s1 | claude-code | novelty | 2026-05-03T00:57:42Z | inconclusive | null | 3.99089 | null | novelty/claude-code/scratchpad/variants/muon_schatten_p4.py | null | heavy-tail-lp | 1 | heavy-tail-lp — Schatten-p soft-spectral descent with Catoni-quantile singular-spectrum pre-shrinkage
# heavy-tail-lp — Schatten-p soft-spectral descent with Catoni-quantile singular-spectrum pre-shrinkage
**Idea ID:** `heavy-tail-lp`
**Family slug:** `htlp` (proposed prefix; if family grows, sub-slugs `htlp-pX`, `ht... | idea_writeup | false | 6 | null | null |
baseline-muon-s3 | claude-code | novelty | 2026-05-03T01:03:44Z | improved | 3,375 | 3.28023 | true | null | null | null | 3 | Baseline run with Muon optimizer, no additional modifications. Establishes control performance metrics for comparison. | llm_generated | false | 7 | null | null |
muon-spm-K4-kd5-ks5-s1 | claude-code | novelty | 2026-05-03T01:04:43Z | inconclusive | null | 3.37555 | null | novelty/claude-code/scratchpad/variants/muon_spm_K4_kd5_ks5.py | null | spectral-momentum | 1 | Idea: spectral-momentum
# Idea: spectral-momentum
**ID:** `spectral-momentum`
**Family slug:** `muon-spm`
**Status:** derivation only, awaiting bench
Headline: maintain Muon's momentum buffer as a *polar factor* (an orthogonal
matrix), Polyak-averaged on the orthogonal manifold via short Newton–Schulz
re-projection,... | idea_writeup | false | 8 | null | null |
muon-aniso-a04-b01-K2-s1 | claude-code | novelty | 2026-05-03T01:04:44Z | improved | 3,375 | 3.2763 | true | novelty/claude-code/scratchpad/variants/muon_aniso_a04_b01_K2.py | null | aniso-kron-muon | 1 | aniso-kron-muon — Curvature-Coupled Newton-Schulz with NS-Residual Anisotropy
# aniso-kron-muon — Curvature-Coupled Newton-Schulz with NS-Residual Anisotropy
**Idea ID:** `aniso-kron-muon`
**Author:** idea-derivation subagent
**Status:** unrun, awaiting noise-floor sweep
---
## 1. Cross-check `ideas.md`
`ai/scratc... | idea_writeup | false | 9 | null | null |
derived-warmup-s1 | claude-code | novelty | 2026-05-03T01:27:24Z | inconclusive | null | 3.28403 | null | novelty/claude-code/scratchpad/variants/derived_warmup.py | null | derived-warmup | 1 | derived-warmup — momentum-transient + NS-conditioning derived warmup shape
# derived-warmup — momentum-transient + NS-conditioning derived warmup shape
**Idea ID:** `derived-warmup`
**Family slug:** `mtwarm` (momentum-transient warmup)
**Status:** derived, not yet run.
---
## 1. Cross-check existing ideas
`ai/scra... | idea_writeup | false | 10 | null | null |
orthogonal-init-s1 | claude-code | novelty | 2026-05-03T01:41:13Z | inconclusive | null | 3.28405 | null | novelty/claude-code/scratchpad/variants/orthogonal_init.py | null | orthogonal-init | 1 | orthogonal-init — Polar-manifold-aligned init for Muon-managed weights
# orthogonal-init — Polar-manifold-aligned init for Muon-managed weights
**Idea ID:** `orthogonal-init`
**Family slug (proposed):** `orthinit`
**Author:** idea-derivation subagent
**Status:** derived, not yet run
**Date:** 2026-05-03
---
## 1. C... | idea_writeup | false | 11 | null | null |
spectral-pre-clip-b3-s1 | claude-code | novelty | 2026-05-03T01:41:14Z | neutral | 3,500 | 3.27891 | true | novelty/claude-code/scratchpad/variants/spectral_pre_clip_b3.py | null | spectral-pre-clip | 1 | spectral-pre-clip — Power-iteration spectral pre-clipping of the *raw gradient* before Muon's momentum lerp
# spectral-pre-clip — Power-iteration spectral pre-clipping of the *raw gradient* before Muon's momentum lerp
**Idea ID:** `spectral-pre-clip`
**Family slug:** `spc` (proposed prefix; sub-slugs `spc-tauX`, `spc... | idea_writeup | false | 12 | null | null |
stacked-orth-warmup-clip-s1 | claude-code | novelty | 2026-05-03T01:44:23Z | inconclusive | null | 3.28973 | null | novelty/claude-code/scratchpad/variants/stacked_orth_warmup_clip.py | null | null | 1 | Stacked orthogonal warmup with gradient clipping. Evaluates the impact of orthogonal initialization on convergence speed. | llm_generated | false | 13 | null | null |
hlp-p4-catoni-q095-s1 | claude-code | novelty | 2026-05-03T02:10:07Z | inconclusive | null | 4.1034 | null | novelty/claude-code/scratchpad/variants/hlp_p4_catoni_q095.py | null | heavy-tail-lp | 1 | heavy-tail-lp — Schatten-p soft-spectral descent with Catoni-quantile singular-spectrum pre-shrinkage
# heavy-tail-lp — Schatten-p soft-spectral descent with Catoni-quantile singular-spectrum pre-shrinkage
**Idea ID:** `heavy-tail-lp`
**Family slug:** `htlp` (proposed prefix; if family grows, sub-slugs `htlp-pX`, `ht... | idea_writeup | false | 14 | null | null |
spectral-wd-lambda0125-K16-s1 | claude-code | novelty | 2026-05-03T02:53:26Z | inconclusive | null | 3.28693 | null | novelty/claude-code/scratchpad/variants/spectral_wd_lambda0125_K16.py | null | spectral-wd | 1 | spectral-wd — Schatten-1/2 proximal weight decay (firm singular-value shrinkage), paired with Muon's spectral-norm constraint
# spectral-wd — Schatten-1/2 proximal weight decay (firm singular-value shrinkage), paired with Muon's spectral-norm constraint
**Idea ID:** `spectral-wd`
**Status:** proposed, not yet run
**A... | idea_writeup | false | 15 | null | null |
spectral-wd-lambda005-K16-s1 | claude-code | novelty | 2026-05-03T03:06:47Z | inconclusive | null | 3.2862 | null | novelty/claude-code/scratchpad/variants/spectral_wd_lambda005_K16.py | null | spectral-wd | 1 | spectral-wd — Schatten-1/2 proximal weight decay (firm singular-value shrinkage), paired with Muon's spectral-norm constraint
# spectral-wd — Schatten-1/2 proximal weight decay (firm singular-value shrinkage), paired with Muon's spectral-norm constraint
**Idea ID:** `spectral-wd`
**Status:** proposed, not yet run
**A... | idea_writeup | false | 16 | null | null |
adaptive-ns-s1 | claude-code | novelty | 2026-05-03T03:10:59Z | improved | 3,375 | 3.27787 | true | novelty/claude-code/scratchpad/variants/adaptive_ns.py | null | adaptive-ns | 1 | Idea `adaptive-ns`: Per-layer adaptive Newton-Schulz with online lower-bound tracking
# Idea `adaptive-ns`: Per-layer adaptive Newton-Schulz with online lower-bound tracking
## 1. Cross-check `ideas.md`
`ai/scratchpad/ideas/` is currently empty (verified at session start). No prior entry to improve on. This is the f... | idea_writeup | false | 17 | null | null |
polar-express-K18-s1 | claude-code | novelty | 2026-05-03T03:11:00Z | neutral | 3,500 | 3.28054 | true | novelty/claude-code/scratchpad/variants/polar_express_K18.py | null | null | 1 | Novelty mechanism with K=18 Newton-Schulz iterations. Aims to enhance convergence speed and stability during training. | llm_generated | false | 18 | null | null |
damped-polar-gap-s1 | claude-code | novelty | 2026-05-03T03:37:04Z | inconclusive | null | 3.32666 | null | novelty/claude-code/scratchpad/variants/damped_polar_gap.py | null | damped-polar-gap | 1 | Idea: damped-polar-gap
# Idea: damped-polar-gap
**ID:** `damped-polar-gap`
**Family slug:** `muon-dpg`
**Status:** derivation only, awaiting existence-check + bench
**Author wave:** 2026-05-03
Headline: replace Muon's fixed Nesterov-momentum coefficient `μ=0.95` with a
**state-coupled damping factor** `ξ_t = μ · (1 ... | idea_writeup | false | 19 | null | null |
muon-mu097-s1 | claude-code | novelty | 2026-05-03T03:37:04Z | inconclusive | null | 3.28517 | null | novelty/claude-code/scratchpad/variants/muon_mu097.py | null | null | 1 | Muon optimizer with momentum mu=0.97, training to 3500 steps. Evaluates performance and efficiency of Muon without warmup phase. | llm_generated | false | 20 | null | null |
bilinear-muon-adaptive-s1 | claude-code | novelty | 2026-05-03T03:37:25Z | neutral | 3,500 | 3.27931 | true | novelty/claude-code/scratchpad/variants/bilinear_muon_adaptive.py | null | bilinear-muon | 1 | bilinear-muon — Bilinear update combining matrix-sign (NS) and Frobenius-normalized raw direction with adaptive interaction weight
# bilinear-muon — Bilinear update combining matrix-sign (NS) and Frobenius-normalized raw direction with adaptive interaction weight
**Idea ID:** `bilinear-muon`
**Family slug:** `bmuon` ... | idea_writeup | false | 21 | null | null |
bilinear-muon-fixedrho-s1 | claude-code | novelty | 2026-05-03T03:37:25Z | neutral | 3,500 | 3.27869 | true | novelty/claude-code/scratchpad/variants/bilinear_muon_fixedrho.py | null | bilinear-muon | 1 | bilinear-muon — Bilinear update combining matrix-sign (NS) and Frobenius-normalized raw direction with adaptive interaction weight
# bilinear-muon — Bilinear update combining matrix-sign (NS) and Frobenius-normalized raw direction with adaptive interaction weight
**Idea ID:** `bilinear-muon`
**Family slug:** `bmuon` ... | idea_writeup | false | 22 | null | null |
acn-adaptive-s1 | claude-code | novelty | 2026-05-03T04:04:23Z | neutral | 3,500 | 3.28114 | true | novelty/claude-code/scratchpad/variants/acn_adaptive.py | null | anti-correlation-nesterov | 1 | Idea: anti-correlation-nesterov
# Idea: anti-correlation-nesterov
**ID:** `anti-correlation-nesterov`
**Family slug:** `muon-acn`
**Status:** derivation only, awaiting bench
Headline: replace canonical Nesterov's lookahead `g + μ·m` with an
**anti-correlation** form that subtracts the component of the buffer
that is... | idea_writeup | false | 23 | null | null |
acn-fixed1-s1 | claude-code | novelty | 2026-05-03T04:04:24Z | neutral | 3,500 | 3.27977 | true | novelty/claude-code/scratchpad/variants/acn_fixed1.py | null | anti-correlation-nesterov | 1 | Idea: anti-correlation-nesterov
# Idea: anti-correlation-nesterov
**ID:** `anti-correlation-nesterov`
**Family slug:** `muon-acn`
**Status:** derivation only, awaiting bench
Headline: replace canonical Nesterov's lookahead `g + μ·m` with an
**anti-correlation** form that subtracts the component of the buffer
that is... | idea_writeup | false | 24 | null | null |
rrm-galore-k8-s1 | claude-code | novelty | 2026-05-03T04:05:45Z | inconclusive | null | 5.32823 | null | novelty/claude-code/scratchpad/variants/rrm_galore_k8.py | null | rank-restricted-muon | 1 | rank-restricted-muon — equalize only the top-k singular components, let the tail evolve as the raw Nesterov-blended gradient
# rank-restricted-muon — equalize only the top-k singular components, let the tail evolve as the raw Nesterov-blended gradient
**Idea ID:** `rank-restricted-muon`
**Family slug:** `rrm` (propos... | idea_writeup | false | 25 | null | null |
rrm-k8-s1 | claude-code | novelty | 2026-05-03T04:05:45Z | inconclusive | null | 4.63062 | null | novelty/claude-code/scratchpad/variants/rrm_k8.py | null | rank-restricted-muon | 1 | rank-restricted-muon — equalize only the top-k singular components, let the tail evolve as the raw Nesterov-blended gradient
# rank-restricted-muon — equalize only the top-k singular components, let the tail evolve as the raw Nesterov-blended gradient
**Idea ID:** `rank-restricted-muon`
**Family slug:** `rrm` (propos... | idea_writeup | false | 26 | null | null |
dual-momentum-muon-adaptive-s1 | claude-code | novelty | 2026-05-03T04:06:03Z | inconclusive | null | 3.31166 | null | novelty/claude-code/scratchpad/variants/dual_momentum_muon_adaptive.py | null | dual-momentum-muon | 1 | Idea: dual-momentum-muon
# Idea: dual-momentum-muon
**ID:** `dual-momentum-muon`
**Family slug:** `muon-dmm`
**Status:** derivation only, awaiting existence-check + bench
**Author wave:** 2026-05-03
Headline: maintain **two** momentum buffers in Muon — one *fast* (μ_fast = 0.5,
ESS ≈ 3) and one *slow* (μ_slow = 0.99... | idea_writeup | false | 27 | null | null |
muon-mu05-s1 | claude-code | novelty | 2026-05-03T04:06:03Z | inconclusive | null | 3.36413 | null | novelty/claude-code/scratchpad/variants/muon_mu05.py | null | null | 1 | Muon optimizer with momentum mu=0.05, default LR schedule. Tests whether lower momentum affects convergence speed and stability. | llm_generated | false | 28 | null | null |
lpm-a01-b05-s1 | claude-code | novelty | 2026-05-03T04:06:55Z | neutral | 3,500 | 3.28131 | true | novelty/claude-code/scratchpad/variants/lpm_a01_b05.py | null | lookahead-polar-mean | 1 | Idea: lookahead-polar-mean
# Idea: lookahead-polar-mean
**ID:** `lookahead-polar-mean`
**Family slug:** `muon-lpm`
**Status:** derivation only, awaiting rules-compliance + bench
Headline: keep canonical Muon's per-step `O_t = NS(M_t)` exactly as is. In
*update space* maintain a parallel **slow update** `S_t` that is... | idea_writeup | false | 29 | null | null |
dcr-g02-s1 | claude-code | novelty | 2026-05-03T04:07:13Z | inconclusive | null | 3.29147 | null | novelty/claude-code/scratchpad/variants/dcr_g02.py | null | decorrelation-regularizer | 1 | Idea: decorrelation-regularizer
# Idea: decorrelation-regularizer
**ID:** `decorrelation-regularizer`
**Family slug:** `muon-dcr`
**Status:** derivation only, awaiting bench
**Author wave:** 2026-05-03
Headline: subtract from the current Muon momentum buffer its **scalar projection
onto the previous step's momentum*... | idea_writeup | false | 30 | null | null |
muon-mu075-s1 | claude-code | novelty | 2026-05-03T04:07:13Z | inconclusive | null | 3.30857 | null | novelty/claude-code/scratchpad/variants/muon_mu075.py | null | null | 1 | Muon optimizer with momentum mu=0.75, training to 3500 steps. Evaluates impact of reduced momentum on convergence speed. | llm_generated | false | 31 | null | null |
lanton-self-b2-s1 | claude-code | novelty | 2026-05-03T04:07:49Z | improved | 3,375 | 3.27732 | true | novelty/claude-code/scratchpad/variants/lanton_self_b2.py | null | lanton-sdm-stacked | 1 | lanton-sdm-stacked — variance-coupled per-layer LR over a Stiefel-row dropout slate
# lanton-sdm-stacked — variance-coupled per-layer LR over a Stiefel-row dropout slate
**Idea ID.** `lanton-sdm-stacked`
**Family slug candidate.** `lansdm` (e.g. `lansdm-b3-p010-c085-s1`)
**Status.** Proposed (pre-code).
**Bar.** ≥ 10... | idea_writeup | false | 32 | null | null |
xlc-b2-s1 | claude-code | novelty | 2026-05-03T04:07:49Z | neutral | 3,500 | 3.27855 | true | novelty/claude-code/scratchpad/variants/xlc_b2.py | null | cross-layer-coupling | 1 | Idea `cross-layer-coupling`: Residual-stream feedback coupling between adjacent Muon updates
# Idea `cross-layer-coupling`: Residual-stream feedback coupling between adjacent Muon updates
**ID:** `cross-layer-coupling`
**Family slug candidate:** `xlc` (e.g. `muon-xlc-v0`)
**Status:** derivation only, awaiting bench
*... | idea_writeup | false | 33 | null | null |
acn-fixed1-s2 | claude-code | novelty | 2026-05-03T04:20:20Z | improved | 3,375 | 3.27711 | true | novelty/claude-code/scratchpad/variants/acn_fixed1.py | null | anti-correlation-nesterov | 2 | Idea: anti-correlation-nesterov
# Idea: anti-correlation-nesterov
**ID:** `anti-correlation-nesterov`
**Family slug:** `muon-acn`
**Status:** derivation only, awaiting bench
Headline: replace canonical Nesterov's lookahead `g + μ·m` with an
**anti-correlation** form that subtracts the component of the buffer
that is... | idea_writeup | false | 34 | null | null |
lanton-self-b2-s2 | claude-code | novelty | 2026-05-03T04:20:20Z | improved | 3,375 | 3.27673 | true | novelty/claude-code/scratchpad/variants/lanton_self_b2.py | null | lanton-sdm-stacked | 2 | lanton-sdm-stacked — variance-coupled per-layer LR over a Stiefel-row dropout slate
# lanton-sdm-stacked — variance-coupled per-layer LR over a Stiefel-row dropout slate
**Idea ID.** `lanton-sdm-stacked`
**Family slug candidate.** `lansdm` (e.g. `lansdm-b3-p010-c085-s1`)
**Status.** Proposed (pre-code).
**Bar.** ≥ 10... | idea_writeup | false | 35 | null | null |
xlc-b2-s2 | claude-code | novelty | 2026-05-03T04:20:20Z | neutral | 3,500 | 3.27859 | true | novelty/claude-code/scratchpad/variants/xlc_b2.py | null | cross-layer-coupling | 2 | Idea `cross-layer-coupling`: Residual-stream feedback coupling between adjacent Muon updates
# Idea `cross-layer-coupling`: Residual-stream feedback coupling between adjacent Muon updates
**ID:** `cross-layer-coupling`
**Family slug candidate:** `xlc` (e.g. `muon-xlc-v0`)
**Status:** derivation only, awaiting bench
*... | idea_writeup | false | 36 | null | null |
gt-muon-l05-e01-s1 | claude-code | novelty | 2026-05-03T04:31:43Z | neutral | 3,500 | 3.28129 | true | novelty/claude-code/scratchpad/variants/gt_muon_l05_e01.py | null | gradient-tracker-muon | 1 | Idea: gradient-tracker-muon
# Idea: gradient-tracker-muon
**ID:** `gradient-tracker-muon`
**Family slug:** `muon-gt`
**Status:** derivation only, awaiting rules-compliance + bench
Headline: replace the raw gradient `g_t` fed into Muon's Nesterov lookahead
with a **leaky cumulative gradient tracker** `y_t` that integ... | idea_writeup | false | 37 | null | null |
mars-m-muon-s1 | claude-code | novelty | 2026-05-03T04:31:43Z | neutral | 3,500 | 3.28072 | true | novelty/claude-code/scratchpad/variants/mars_m_muon.py | null | null | 1 | Muon variant with unspecified hyperparameters, demonstrating progressive training time reduction across steps. Evaluates efficiency gains in training dynamics. | llm_generated | false | 38 | null | null |
adamuon-s1 | claude-code | novelty | 2026-05-03T04:31:55Z | inconclusive | null | 3.34998 | null | novelty/claude-code/scratchpad/variants/adamuon.py | null | adam-rms-muon | 1 | adam-rms-muon — Per-element raw-gradient Adam-shape v_t dividing NS(M)
# adam-rms-muon — Per-element raw-gradient Adam-shape v_t dividing NS(M)
**Idea ID:** `adam-rms-muon`
**Family slug:** `armmuon` (proposed prefix; sub-slugs `armmuon-aX-bX`, `armmuon-epsX`)
**Status:** derived, not yet run
**Author:** idea-derivat... | idea_writeup | false | 39 | null | null |
arm-muon-s1 | claude-code | novelty | 2026-05-03T04:31:55Z | inconclusive | null | 4.83161 | null | novelty/claude-code/scratchpad/variants/arm_muon.py | null | adam-rms-muon | 1 | adam-rms-muon — Per-element raw-gradient Adam-shape v_t dividing NS(M)
# adam-rms-muon — Per-element raw-gradient Adam-shape v_t dividing NS(M)
**Idea ID:** `adam-rms-muon`
**Family slug:** `armmuon` (proposed prefix; sub-slugs `armmuon-aX-bX`, `armmuon-epsX`)
**Status:** derived, not yet run
**Author:** idea-derivat... | idea_writeup | false | 40 | null | null |
cov-precond-muon-a00-s1 | claude-code | novelty | 2026-05-03T05:01:13Z | neutral | 3,500 | 3.28003 | true | novelty/claude-code/scratchpad/variants/cov_precond_muon_a00.py | null | cov-precond-muon | 1 | cov-precond-muon — Halko-Martinsson rank-k gradient-covariance projector pre-multiplied into the Nesterov-blended momentum, before NS
# cov-precond-muon — Halko-Martinsson rank-k gradient-covariance projector pre-multiplied into the Nesterov-blended momentum, before NS
**Idea ID:** `cov-precond-muon`
**Family slug (p... | idea_writeup | false | 41 | null | null |
cov-precond-muon-a05-k8-s1 | claude-code | novelty | 2026-05-03T05:01:13Z | neutral | 3,500 | 3.27836 | true | novelty/claude-code/scratchpad/variants/cov_precond_muon_a05_k8.py | null | cov-precond-muon | 1 | cov-precond-muon — Halko-Martinsson rank-k gradient-covariance projector pre-multiplied into the Nesterov-blended momentum, before NS
# cov-precond-muon — Halko-Martinsson rank-k gradient-covariance projector pre-multiplied into the Nesterov-blended momentum, before NS
**Idea ID:** `cov-precond-muon`
**Family slug (p... | idea_writeup | false | 42 | null | null |
lanton-self-b2-s3 | claude-code | novelty | 2026-05-03T05:13:34Z | improved | 3,375 | 3.27809 | true | novelty/claude-code/scratchpad/variants/lanton_self_b2.py | null | lanton-sdm-stacked | 3 | lanton-sdm-stacked — variance-coupled per-layer LR over a Stiefel-row dropout slate
# lanton-sdm-stacked — variance-coupled per-layer LR over a Stiefel-row dropout slate
**Idea ID.** `lanton-sdm-stacked`
**Family slug candidate.** `lansdm` (e.g. `lansdm-b3-p010-c085-s1`)
**Status.** Proposed (pre-code).
**Bar.** ≥ 10... | idea_writeup | false | 43 | null | null |
muon-firstorder-alt-s1 | claude-code | novelty | 2026-05-03T05:21:00Z | neutral | 3,500 | 3.27862 | true | novelty/claude-code/scratchpad/variants/muon_firstorder_alt.py | null | null | 1 | First-order Muon variant with no specified learning rate or momentum parameters. Tests whether first-order dynamics improve convergence speed compared to second-order methods. | llm_generated | false | 44 | null | null |
strang-split-muon-s1 | claude-code | novelty | 2026-05-03T05:21:00Z | neutral | 3,500 | 3.27925 | true | novelty/claude-code/scratchpad/variants/strang_split_muon.py | null | strang-split-muon | 1 | strang-split-muon
# strang-split-muon
ID: `strang-split-muon`
Family slug: `muon-strang`
Status: derivation complete; existence-check pending; rules-compliance audit pending.
## 1. Cross-check against `ideas.md`
Scanned all 27 idea files. Closest neighbours:
- `bilinear-muon.md` — multiplicative (NS-output × WD) u... | idea_writeup | false | 45 | null | null |
dropgrad-pre-ns-p010-s1 | claude-code | novelty | 2026-05-03T05:43:16Z | inconclusive | null | 3.29902 | null | novelty/claude-code/scratchpad/variants/dropgrad_pre_ns_p010.py | null | stochastic-dropout-muon | 1 | stochastic-dropout-muon
# stochastic-dropout-muon
**Idea ID.** `stochastic-dropout-muon`
**Family slug.** `sdm`
**Status.** Proposed (pre-code).
**Bar.** ≥ 100-step gain over Muon @ 3375 (`step_to_3.28`), 2 seeds, ≥ 2× noise floor.
## 1. Cross-check vs `ideas/`
I ran a textual scan of every `ai/scratchpad/ideas/*.m... | idea_writeup | false | 46 | null | null |
sdm-row-p010-s1 | claude-code | novelty | 2026-05-03T05:43:17Z | improved | 3,375 | 3.27685 | true | novelty/claude-code/scratchpad/variants/sdm_row_p010.py | null | stochastic-dropout-muon | 1 | stochastic-dropout-muon
# stochastic-dropout-muon
**Idea ID.** `stochastic-dropout-muon`
**Family slug.** `sdm`
**Status.** Proposed (pre-code).
**Bar.** ≥ 100-step gain over Muon @ 3375 (`step_to_3.28`), 2 seeds, ≥ 2× noise floor.
## 1. Cross-check vs `ideas/`
I ran a textual scan of every `ai/scratchpad/ideas/*.m... | idea_writeup | false | 47 | null | null |
muon-K8-fixed-s1 | claude-code | novelty | 2026-05-03T05:46:49Z | neutral | 3,500 | 3.27842 | true | novelty/claude-code/scratchpad/variants/muon_K8_fixed.py | null | null | 1 | Muon optimizer with fixed Newton-Schulz iterations K=8. Evaluates performance stability across training steps. | llm_generated | false | 48 | null | null |
sdm-row-p010-s2 | claude-code | novelty | 2026-05-03T05:53:26Z | neutral | 3,500 | 3.27788 | true | novelty/claude-code/scratchpad/variants/sdm_row_p010.py | null | stochastic-dropout-muon | 2 | stochastic-dropout-muon
# stochastic-dropout-muon
**Idea ID.** `stochastic-dropout-muon`
**Family slug.** `sdm`
**Status.** Proposed (pre-code).
**Bar.** ≥ 100-step gain over Muon @ 3375 (`step_to_3.28`), 2 seeds, ≥ 2× noise floor.
## 1. Cross-check vs `ideas/`
I ran a textual scan of every `ai/scratchpad/ideas/*.m... | idea_writeup | false | 49 | null | null |
adagrad-muon-s1 | claude-code | novelty | 2026-05-03T06:18:07Z | inconclusive | null | 3.28695 | null | novelty/claude-code/scratchpad/variants/adagrad_muon.py | null | adagrad-muon | 1 | adagrad-muon — Unbounded AdaGrad-sum preconditioner pre-NS
# adagrad-muon — Unbounded AdaGrad-sum preconditioner pre-NS
**Idea ID:** `adagrad-muon`
**Family slug:** `agmuon` (proposed prefix; sub-slugs `agmuon-epsX`, `agmuon-init-vX`)
**Status:** derived, not yet run
**Author:** idea-derivation subagent (orchestrator... | idea_writeup | false | 50 | null | null |
prensvar-muon-b95-s1 | claude-code | novelty | 2026-05-03T06:18:07Z | inconclusive | null | 3.28359 | null | novelty/claude-code/scratchpad/variants/prensvar_muon_b95.py | null | null | 1 | Prensvar variant of Muon optimizer with beta=0.95. Evaluates performance impact of modified variance handling in training dynamics. | llm_generated | false | 51 | null | null |
lansdm-serial-s1 | claude-code | novelty | 2026-05-03T06:19:18Z | neutral | 3,500 | 3.2788 | true | novelty/claude-code/scratchpad/variants/lansdm_serial.py | null | lanton-sdm-stacked | 1 | lanton-sdm-stacked — variance-coupled per-layer LR over a Stiefel-row dropout slate
# lanton-sdm-stacked — variance-coupled per-layer LR over a Stiefel-row dropout slate
**Idea ID.** `lanton-sdm-stacked`
**Family slug candidate.** `lansdm` (e.g. `lansdm-b3-p010-c085-s1`)
**Status.** Proposed (pre-code).
**Bar.** ≥ 10... | idea_writeup | false | 52 | null | null |
baseline-muon-s2 | claude-code | novelty | 2026-05-03T06:33:45Z | inconclusive | null | 3.32658 | null | null | null | null | null | Baseline run with Muon optimizer, no additional modifications. Establishes control performance metrics for comparison. | llm_generated | false | 53 | null | null |
pdc-random-s1 | claude-code | novelty | 2026-05-03T06:44:31Z | neutral | 3,500 | 3.27985 | true | novelty/claude-code/scratchpad/variants/pdc_random.py | null | per-direction-coherence-lr | 1 | per-direction-coherence-lr — per-column LR multiplier for the NS output, driven by alignment with a Halko-tracked principal subspace of consecutive NS outputs
# per-direction-coherence-lr — per-column LR multiplier for the NS output, driven by alignment with a Halko-tracked principal subspace of consecutive NS outputs... | idea_writeup | false | 54 | null | null |
pdc-s1 | claude-code | novelty | 2026-05-03T06:44:32Z | neutral | 3,500 | 3.27874 | true | novelty/claude-code/scratchpad/variants/pdc.py | null | per-direction-coherence-lr | 1 | per-direction-coherence-lr — per-column LR multiplier for the NS output, driven by alignment with a Halko-tracked principal subspace of consecutive NS outputs
# per-direction-coherence-lr — per-column LR multiplier for the NS output, driven by alignment with a Halko-tracked principal subspace of consecutive NS outputs... | idea_writeup | false | 55 | null | null |
pdc-s2 | claude-code | novelty | 2026-05-03T07:07:41Z | neutral | 3,500 | 3.27958 | true | novelty/claude-code/scratchpad/variants/pdc.py | null | per-direction-coherence-lr | 2 | per-direction-coherence-lr — per-column LR multiplier for the NS output, driven by alignment with a Halko-tracked principal subspace of consecutive NS outputs
# per-direction-coherence-lr — per-column LR multiplier for the NS output, driven by alignment with a Halko-tracked principal subspace of consecutive NS outputs... | idea_writeup | false | 56 | null | null |
sdm-row-p005-s1 | claude-code | novelty | 2026-05-03T07:07:41Z | neutral | 3,500 | 3.27855 | true | novelty/claude-code/scratchpad/variants/sdm_row_p005.py | null | stochastic-dropout-muon | 1 | stochastic-dropout-muon
# stochastic-dropout-muon
**Idea ID.** `stochastic-dropout-muon`
**Family slug.** `sdm`
**Status.** Proposed (pre-code).
**Bar.** ≥ 100-step gain over Muon @ 3375 (`step_to_3.28`), 2 seeds, ≥ 2× noise floor.
## 1. Cross-check vs `ideas/`
I ran a textual scan of every `ai/scratchpad/ideas/*.m... | idea_writeup | false | 57 | null | null |
sdm-row-p020-s1 | claude-code | novelty | 2026-05-03T07:07:41Z | neutral | 3,500 | 3.27866 | true | novelty/claude-code/scratchpad/variants/sdm_row_p020.py | null | stochastic-dropout-muon | 1 | stochastic-dropout-muon
# stochastic-dropout-muon
**Idea ID.** `stochastic-dropout-muon`
**Family slug.** `sdm`
**Status.** Proposed (pre-code).
**Bar.** ≥ 100-step gain over Muon @ 3375 (`step_to_3.28`), 2 seeds, ≥ 2× noise floor.
## 1. Cross-check vs `ideas/`
I ran a textual scan of every `ai/scratchpad/ideas/*.m... | idea_writeup | false | 58 | null | null |
muon-aniso-a05-b00-K2-s1 | claude-code | novelty | 2026-05-03T07:37:38Z | inconclusive | null | 3.36648 | null | novelty/claude-code/scratchpad/variants/muon_aniso_a05_b00_K2.py | null | aniso-kron-muon | 1 | aniso-kron-muon — Curvature-Coupled Newton-Schulz with NS-Residual Anisotropy
# aniso-kron-muon — Curvature-Coupled Newton-Schulz with NS-Residual Anisotropy
**Idea ID:** `aniso-kron-muon`
**Author:** idea-derivation subagent
**Status:** unrun, awaiting noise-floor sweep
---
## 1. Cross-check `ideas.md`
`ai/scratc... | idea_writeup | false | 59 | null | null |
alphadecay-baseline-s1 | claude-code | novelty | 2026-05-03T07:39:25Z | improved | 3,375 | 3.277 | true | novelty/claude-code/scratchpad/variants/alphadecay_baseline.py | null | null | 1 | Baseline run with alpha decay mechanism, utilizing heavy (1.5x) and light (1.0x) parameter tiers. Establishes performance benchmarks without additional modifications. | llm_generated | false | 60 | null | null |
imuon-d05-fixedp-s1 | claude-code | novelty | 2026-05-03T08:00:11Z | inconclusive | null | 4.38382 | null | novelty/claude-code/scratchpad/variants/imuon_d05_fixedp.py | null | implicit-muon | 1 | Idea `implicit-muon`: 1-step fixed-point implicit Muon via NS-aligned Hessian-vector correction
# Idea `implicit-muon`: 1-step fixed-point implicit Muon via NS-aligned Hessian-vector correction
**Idea ID:** `implicit-muon`
**Family slug:** `imuon`
**Status:** derived, not yet run
**Author:** idea-derivation subagent ... | idea_writeup | false | 61 | null | null |
imuon-d05-s1 | claude-code | novelty | 2026-05-03T08:00:11Z | inconclusive | null | 4.89951 | null | novelty/claude-code/scratchpad/variants/imuon_d05.py | null | implicit-muon | 1 | Idea `implicit-muon`: 1-step fixed-point implicit Muon via NS-aligned Hessian-vector correction
# Idea `implicit-muon`: 1-step fixed-point implicit Muon via NS-aligned Hessian-vector correction
**Idea ID:** `implicit-muon`
**Family slug:** `imuon`
**Status:** derived, not yet run
**Author:** idea-derivation subagent ... | idea_writeup | false | 62 | null | null |
hbns-l010-s1 | claude-code | novelty | 2026-05-03T08:01:45Z | neutral | 3,500 | 3.28158 | true | novelty/claude-code/scratchpad/variants/hbns_l010.py | null | heavyball-ns-coupled | 1 | Idea: heavyball-ns-coupled
# Idea: heavyball-ns-coupled
**ID:** `heavyball-ns-coupled`
**Family slug:** `muon-hbns`
**Status:** derivation only, awaiting existence-check + rules-compliance + bench
**Author wave:** 2026-05-03
Headline: replace canonical Muon's *Nesterov-EMA* lookahead with a single
**heavy-ball (Poly... | idea_writeup | false | 63 | null | null |
hbns-l100-s1 | claude-code | novelty | 2026-05-03T08:01:45Z | inconclusive | null | 3.33326 | null | novelty/claude-code/scratchpad/variants/hbns_l100.py | null | heavyball-ns-coupled | 1 | Idea: heavyball-ns-coupled
# Idea: heavyball-ns-coupled
**ID:** `heavyball-ns-coupled`
**Family slug:** `muon-hbns`
**Status:** derivation only, awaiting existence-check + rules-compliance + bench
**Author wave:** 2026-05-03
Headline: replace canonical Muon's *Nesterov-EMA* lookahead with a single
**heavy-ball (Poly... | idea_writeup | false | 64 | null | null |
smm-postpos-s1 | claude-code | novelty | 2026-05-03T08:22:01Z | inconclusive | null | 3.29475 | null | novelty/claude-code/scratchpad/variants/smm_postpos.py | null | sign-modulated-muon | 1 | Idea: sign-modulated-muon
# Idea: sign-modulated-muon
**ID:** `sign-modulated-muon`
**Family slug:** `muon-smm`
**Status:** derivation only, awaiting bench
Headline: modulate Muon's polar update **per element** by the agreement
between (i) `sign(NS(M))` (the post-orthogonalisation matrix sign) and
(ii) `sign(M)` (th... | idea_writeup | false | 65 | null | null |
smm-l010-s1 | claude-code | novelty | 2026-05-03T08:22:02Z | inconclusive | null | 3.28372 | null | novelty/claude-code/scratchpad/variants/smm_l010.py | null | sign-modulated-muon | 1 | Idea: sign-modulated-muon
# Idea: sign-modulated-muon
**ID:** `sign-modulated-muon`
**Family slug:** `muon-smm`
**Status:** derivation only, awaiting bench
Headline: modulate Muon's polar update **per element** by the agreement
between (i) `sign(NS(M))` (the post-orthogonalisation matrix sign) and
(ii) `sign(M)` (th... | idea_writeup | false | 66 | null | null |
head-lr-280-s1 | claude-code | novelty | 2026-05-03T09:24:32Z | inconclusive | null | 3.28153 | null | novelty/claude-code/scratchpad/variants/head_lr_280.py | null | head-grad-adaptive | 1 | head-grad-adaptive — Head LR adapted to its own gradient-norm signature
# head-grad-adaptive — Head LR adapted to its own gradient-norm signature
**Idea ID:** `head-grad-adaptive`
**Family slug (proposed):** `hgadt` (head-grad-adapt)
**Author:** idea-derivation subagent
**Status:** derived, awaiting rules-compliance ... | idea_writeup | false | 67 | null | null |
hgadt-bm1-s1 | claude-code | novelty | 2026-05-03T09:24:33Z | neutral | 3,500 | 3.27877 | true | novelty/claude-code/scratchpad/variants/hgadt_bm1.py | null | head-grad-adaptive | 1 | head-grad-adaptive — Head LR adapted to its own gradient-norm signature
# head-grad-adaptive — Head LR adapted to its own gradient-norm signature
**Idea ID:** `head-grad-adaptive`
**Family slug (proposed):** `hgadt` (head-grad-adapt)
**Author:** idea-derivation subagent
**Status:** derived, awaiting rules-compliance ... | idea_writeup | false | 68 | null | null |
psmuon-s1 | claude-code | novelty | 2026-05-03T10:47:18Z | neutral | 3,500 | 3.27993 | true | novelty/claude-code/scratchpad/variants/psmuon.py | null | polyak-step-muon | 1 | polyak-step-muon — running-min/max grad-norm-surrogate Polyak multiplier × Muon NS
# polyak-step-muon — running-min/max grad-norm-surrogate Polyak multiplier × Muon NS
**Idea ID:** `polyak-step-muon`
**Family slug (proposed):** `psmuon` (sub-slugs `psmuon-K500-g1`, `psmuon-K1000-g05`, `psmuon-Knone-g1`)
**Status:** d... | idea_writeup | false | 69 | null | null |
psmuon-shuffled-s1 | claude-code | novelty | 2026-05-03T10:47:18Z | inconclusive | null | 3.28024 | null | novelty/claude-code/scratchpad/variants/psmuon_shuffled.py | null | polyak-step-muon | 1 | polyak-step-muon — running-min/max grad-norm-surrogate Polyak multiplier × Muon NS
# polyak-step-muon — running-min/max grad-norm-surrogate Polyak multiplier × Muon NS
**Idea ID:** `polyak-step-muon`
**Family slug (proposed):** `psmuon` (sub-slugs `psmuon-K500-g1`, `psmuon-K1000-g05`, `psmuon-Knone-g1`)
**Status:** d... | idea_writeup | false | 70 | null | null |
residual-term-ns-s1 | claude-code | novelty | 2026-05-03T11:16:49Z | neutral | 3,500 | 3.27911 | true | novelty/claude-code/scratchpad/variants/residual_term_ns.py | null | residual-termination-ns | 1 | Idea `residual-termination-ns`: Per-call Newton-Schulz termination via spectral-residual bound
# Idea `residual-termination-ns`: Per-call Newton-Schulz termination via spectral-residual bound
## 1. Cross-check `ideas.md`
Closest existing entry: `adaptive-ns` (per-layer EMA participation-ratio → coefficient *table lo... | idea_writeup | false | 71 | null | null |
lansdm-s1 | claude-code | novelty | 2026-05-03T11:49:18Z | neutral | 3,500 | 3.27819 | true | novelty/claude-code/scratchpad/variants/lansdm.py | null | lanton-sdm-stacked | 1 | lanton-sdm-stacked — variance-coupled per-layer LR over a Stiefel-row dropout slate
# lanton-sdm-stacked — variance-coupled per-layer LR over a Stiefel-row dropout slate
**Idea ID.** `lanton-sdm-stacked`
**Family slug candidate.** `lansdm` (e.g. `lansdm-b3-p010-c085-s1`)
**Status.** Proposed (pre-code).
**Bar.** ≥ 10... | idea_writeup | false | 72 | null | null |
krm-K2-g100-s1 | claude-code | novelty | 2026-05-03T12:55:06Z | inconclusive | null | 3.29875 | null | novelty/claude-code/scratchpad/variants/krm_K2_g100.py | null | krylov-residual-muon | 1 | krylov-residual-muon — Galerkin projection of Nesterov-momentum onto the orthogonal complement of past-NS outputs, *before* NS
# krylov-residual-muon — Galerkin projection of Nesterov-momentum onto the orthogonal complement of past-NS outputs, *before* NS
**Idea ID:** `krylov-residual-muon`
**Family slug (proposed):*... | idea_writeup | false | 73 | null | null |
krm-shuffled-s1 | claude-code | novelty | 2026-05-03T12:55:06Z | neutral | 3,500 | 3.27822 | true | novelty/claude-code/scratchpad/variants/krm_shuffled.py | null | krylov-residual-muon | 1 | krylov-residual-muon — Galerkin projection of Nesterov-momentum onto the orthogonal complement of past-NS outputs, *before* NS
# krylov-residual-muon — Galerkin projection of Nesterov-momentum onto the orthogonal complement of past-NS outputs, *before* NS
**Idea ID:** `krylov-residual-muon`
**Family slug (proposed):*... | idea_writeup | false | 74 | null | null |
ravn-K2-fix-s1 | claude-code | novelty | 2026-05-03T14:09:57Z | neutral | 3,500 | 3.279 | true | novelty/claude-code/scratchpad/variants/ravn_K2_fix.py | null | rotation-averaged-ns | 1 | rotation-averaged-ns — Variance-reduced NS via averaging two rotation-conjugated NS calls
# rotation-averaged-ns — Variance-reduced NS via averaging two rotation-conjugated NS calls
**Idea ID:** `rotation-averaged-ns`
**Family slug (proposed):** `ravn` (sub-slugs `ravn-K2`, `ravn-K2-fix`, `ravn-K2-houshld`)
**Status:... | idea_writeup | false | 75 | null | null |
ravn-K2-s1 | claude-code | novelty | 2026-05-03T14:09:57Z | neutral | 3,500 | 3.27851 | true | novelty/claude-code/scratchpad/variants/ravn_K2.py | null | rotation-averaged-ns | 1 | rotation-averaged-ns — Variance-reduced NS via averaging two rotation-conjugated NS calls
# rotation-averaged-ns — Variance-reduced NS via averaging two rotation-conjugated NS calls
**Idea ID:** `rotation-averaged-ns`
**Family slug (proposed):** `ravn` (sub-slugs `ravn-K2`, `ravn-K2-fix`, `ravn-K2-houshld`)
**Status:... | idea_writeup | false | 76 | null | null |
snrm-rand-s1 | claude-code | novelty | 2026-05-03T15:32:14Z | inconclusive | null | 3.53504 | null | novelty/claude-code/scratchpad/variants/snrm_rand.py | null | snr-gated-muon | 1 | snr-gated-muon — Per-element pre-NS sigmoid gate from cross-step gradient SNR (consistency variance, not magnitude variance)
# snr-gated-muon — Per-element pre-NS sigmoid gate from cross-step gradient SNR (consistency variance, not magnitude variance)
**Idea ID:** `snr-gated-muon`
**Family slug:** `snrm` (sub-slugs `... | idea_writeup | false | 77 | null | null |
snrm-s1 | claude-code | novelty | 2026-05-03T15:32:14Z | inconclusive | null | 3.29557 | null | novelty/claude-code/scratchpad/variants/snrm.py | null | snr-gated-muon | 1 | snr-gated-muon — Per-element pre-NS sigmoid gate from cross-step gradient SNR (consistency variance, not magnitude variance)
# snr-gated-muon — Per-element pre-NS sigmoid gate from cross-step gradient SNR (consistency variance, not magnitude variance)
**Idea ID:** `snr-gated-muon`
**Family slug:** `snrm` (sub-slugs `... | idea_writeup | false | 78 | null | null |
cbo-s1 | claude-code | novelty | 2026-05-03T16:54:40Z | improved | 3,375 | 3.27651 | true | novelty/claude-code/scratchpad/variants/cbo.py | null | cross-block-ortho | 1 | Idea `cross-block-ortho`: Per-step Gram-Schmidt of NS-orthogonalized updates across the 12 transformer blocks
# Idea `cross-block-ortho`: Per-step Gram-Schmidt of NS-orthogonalized updates across the 12 transformer blocks
**ID:** `cross-block-ortho`
**Family slug candidate:** `cbo` (e.g. `muon-cbo-b1-s1`)
**Status:**... | idea_writeup | false | 79 | null | null |
cbo-shuffled-s1 | claude-code | novelty | 2026-05-03T16:54:40Z | neutral | 3,500 | 3.27766 | true | novelty/claude-code/scratchpad/variants/cbo_shuffled.py | null | cross-block-ortho | 1 | Idea `cross-block-ortho`: Per-step Gram-Schmidt of NS-orthogonalized updates across the 12 transformer blocks
# Idea `cross-block-ortho`: Per-step Gram-Schmidt of NS-orthogonalized updates across the 12 transformer blocks
**ID:** `cross-block-ortho`
**Family slug candidate:** `cbo` (e.g. `muon-cbo-b1-s1`)
**Status:**... | idea_writeup | false | 80 | null | null |
cbo-s2 | claude-code | novelty | 2026-05-03T17:06:26Z | neutral | 3,500 | 3.27801 | true | novelty/claude-code/scratchpad/variants/cbo.py | null | cross-block-ortho | 2 | Idea `cross-block-ortho`: Per-step Gram-Schmidt of NS-orthogonalized updates across the 12 transformer blocks
# Idea `cross-block-ortho`: Per-step Gram-Schmidt of NS-orthogonalized updates across the 12 transformer blocks
**ID:** `cross-block-ortho`
**Family slug candidate:** `cbo` (e.g. `muon-cbo-b1-s1`)
**Status:**... | idea_writeup | false | 81 | null | null |
psinv-a10-s1 | claude-code | novelty | 2026-05-03T17:13:55Z | neutral | 3,500 | 3.27768 | true | novelty/claude-code/scratchpad/variants/psinv_a10.py | null | pseudo-inverse-muon | null | Idea `pseudo-inverse-muon`: Polar × pseudo-inverse blend in Muon's NS step
# Idea `pseudo-inverse-muon`: Polar × pseudo-inverse blend in Muon's NS step
## 1. Cross-check `ideas.md`
Closest existing entries:
- `adaptive-ns`: adapts NS iteration count `K_layer` (participation-ratio EMA). Different lever; FLAT=BASELINE... | idea_writeup | false | 82 | null | null |
psinv-a05-s1 | claude-code | novelty | 2026-05-03T17:13:56Z | neutral | 3,500 | 3.2792 | true | novelty/claude-code/scratchpad/variants/psinv_a05.py | null | pseudo-inverse-muon | null | Idea `pseudo-inverse-muon`: Polar × pseudo-inverse blend in Muon's NS step
# Idea `pseudo-inverse-muon`: Polar × pseudo-inverse blend in Muon's NS step
## 1. Cross-check `ideas.md`
Closest existing entries:
- `adaptive-ns`: adapts NS iteration count `K_layer` (participation-ratio EMA). Different lever; FLAT=BASELINE... | idea_writeup | false | 83 | null | null |
psw-fixed-s1 | claude-code | novelty | 2026-05-03T17:14:18Z | inconclusive | null | 10.82584 | null | novelty/claude-code/scratchpad/variants/psw_fixed.py | null | per-sigma-wd | 1 | per-sigma-wd — Spread-driven, per-singular-value soft weight decay using the polar gap as an online spectral-concentration signal
# per-sigma-wd — Spread-driven, per-singular-value soft weight decay using the polar gap as an online spectral-concentration signal
**Idea ID:** `per-sigma-wd`
**Family slug (proposed):** ... | idea_writeup | false | 84 | null | null |
End of preview. Expand in Data Studio
Autoresearch Novelty Bench
A benchmark for testing whether an autonomous AI research agent proposes novel, mechanism-distinct hypotheses that anticipate breakthroughs later found by other researchers.
By Evo. Built on Prime Intellect's
autonomous-speedrunning archive
— two AI agents (Claude Code and Codex) competing on modded-nanogpt's
optimization speedrun.
What's in this dataset
| table | rows | description |
|---|---|---|
experiments.parquet |
10,380 | One row per training run — primary catalog the benchmark scores against |
ideas.parquet |
646 | Formal idea writeups (ideas/*.md from PI's archive) |
proposals.parquet |
3,249 | Variant-level catalog (one row per variant.py) |
runs.parquet |
10,380 | Raw run metadata (timestamps, outcomes, linkage) |
val_trajectories.parquet |
337,648 | Per-checkpoint val_loss for every run |
snapshots.parquet |
40 | Time-anchored evaluation positions (16 dev / 24 test) |
experiment_embeddings_openai.parquet |
10,380 × 3072 | text-embedding-3-large |
experiment_embeddings_bge.parquet |
10,380 × 1024 | BAAI/bge-large-en-v1.5 |
recovered_variant_sources.parquet |
1,120 | Gold-standard launched .py extracted from log content |
calibration_cases.parquet |
78 | Labeled cases for judge calibration |
recovered_variants/ |
1,120 files | Per-run launched .py source extracted from training logs |
current_recipes/ |
35 files | Per-snapshot current-best variant.py the agent was building on |
Usage
The accompanying Python package handles loading + scoring:
pip install evo-autoresearch-novelty-bench
import novelty_bench as nb
# Browse the dataset (auto-downloads from HF)
nb.load_experiments(agent="codex", wave="v3", verdict="improved")
nb.load_snapshots(split="dev") # 16 dev snapshots
nb.load_snapshots(split="test") # 24 held-out test
nb.load_ideas()
nb.load_val_trajectories(run_id="alr001_k025")
# Score your proposer's candidates against a snapshot
snap = nb.load_snapshots()
snap = snap[snap["snapshot_id"] == "snap_codex_v3_k0922"].iloc[0]
result = nb.score(snap, ["candidate_1.md", "candidate_2.md", ...])
print(result.explain())
The default judge backend is llm-hybrid (cosine retrieval + gpt-5-mini
reasoning; needs OPENAI_API_KEY; ~$0.005 per N=5 candidate set). Pass
judge_backend="cosine" to nb.score() for a deterministic, embedding-only
classification (no LLM call, ~$0.0003 per set).
For end-to-end agent benchmarking, see the scaffold repo:
evo-hq/autoresearch-novelty-bench-scaffold.
How the dataset was built
- Parse the upstream archive — walk 8 scratchpads, parse runs.jsonl, training logs, idea writeups
- Link runs to ideas via heuristic + abbreviation + LLM matcher (63% link rate)
- Recover variant.py linkage via 4 cascade passes (embedded source, dynamo warnings, sbatch stubs, name-stem) — 54% → 95.6% code coverage
- Build experiments catalog (10,380 rows, primary unit for benchmark scoring)
- LLM-augment descriptions for runs without parent idea writeups (~$0.26 with gpt-4o-mini)
- Embed everything with OpenAI text-embedding-3-large + BGE-large
- Build 40 time-anchored snapshots across 8 (agent, wave) scopes
- Calibrate the judge — 78 labeled cases sweep cosine threshold to 0.75
Full details + reproducibility instructions in the main repo:
evo-hq/autoresearch-novelty-bench.
Source attribution
The raw ideas/*.md and variants/*.py files indexed here belong to
Prime Intellect. This dataset ships
derived metadata (step counts, verdicts, snapshot schemas, embeddings,
linkage tables) plus a small subset of raw .py files that we extracted from
log content (recovered_variants/) or that document each snapshot's
current-best recipe (current_recipes/).
License
Apache 2.0 (derived data + linkage tables). Upstream raw text/code is governed by its own terms — see the Prime Intellect repo.
- Downloads last month
- 1,531