| Age | Commit message (Expand) | Author |
|---|
| 2026-07-13 | Stage-2 data pipeline: FineWeb-Edu sample-10BT -> 32k BPE -> bins; --data fla... | Yuren Hao |
| 2026-07-12 | bf16 mixed-precision (--amp): gate PASSED (cos 0.9682 vs fp32 0.9687), traine... | Yuren Hao |
| 2026-07-12 | E-tier wave-1 ledger (RESULT 14 + HW map rows) + cost model v2 recalibrated o... | Yuren Hao |
| 2026-07-12 | RESULT 13: bsign neutral at 4k (1.7382 vs 1.7314); 42M gap-chasing closed at ... | Yuren Hao |
| 2026-07-11 | RESULT 11: stage1b sealed (gap 0.050); beta ceiling unfound (3e-3 -> cos 0.99... | Yuren Hao |
| 2026-07-10 | RESULT 10: epoch endpoints (BP 1.2509/1.2750 vs EP 1.4802, +0.22 erosion cost... | Yuren Hao |
| 2026-07-10 | Sustained dose-response sealed (3e-4/5e-4/1e-3 -> 0.975/0.983/0.990); ship --... | Yuren Hao |
| 2026-07-10 | RESULT 6: wall-2 = EP-specific marginal under-convergence at the contractivit... | Yuren Hao |
| 2026-07-10 | OLMo2-standard architecture (user directive): norm-after-sublayer RMSNorm blo... | Yuren Hao |
| 2026-07-10 | AUDIT: retract confounded Muon verdict; tone down parity claims (n=3, best-of... | Yuren Hao |
| 2026-07-10 | Stage-1 epoch blowup @12100: sig story REFUTED (only +8%, cos fine till after... | Yuren Hao |
| 2026-07-10 | QK-norm: replace nn.MHA with explicit SDPA attn + --qk_norm (OLMo2-style, ana... | Yuren Hao |
| 2026-07-10 | Add --cosine (warmup->cosine decay to 0.1x lr) to both cascade trainers for f... | Yuren Hao |
| 2026-07-10 | K refuted as L12 lever (cos K-invariant K3==K8); BP 3-seed sealed 1.919; mech... | Yuren Hao |
| 2026-07-09 | BP-free formal audit (5.6-sol): EP update verified local (108 bitwise zero-in... | Yuren Hao |
| 2026-07-09 | A0.4 precision gate (TF32 harmless cos 0.9946==fp32; pure-bf16 cos 0.9427) + ... | Yuren Hao |
| 2026-07-09 | casc_eq_train: exact mode (dtop_every=1) now DEFAULT — bisection showed the... | Yuren Hao |
| 2026-07-09 | casc_eq_train v7: two algorithmic dedups — d_top refreshed every other roun... | Yuren Hao |
| 2026-07-09 | casc_eq_train v6: graph-reuse fb (rebuild graphs feed next round's vjps), --c... | Yuren Hao |
| 2026-07-09 | cascade root cause found: tied readout + default N(0,1) embedding init = path... | Yuren Hao |
| 2026-07-09 | casc_eq_train v4: quality-governed estimator — periodic measured cos(EP,BP)... | Yuren Hao |
| 2026-07-09 | casc_eq_train v3: adaptive beta (tok-sigma stiffness compensation), adaptive-... | Yuren Hao |
| 2026-07-09 | cascade equilibrium solver solved: fb (forward-backward message passing) K=3 ... | Yuren Hao |
| 2026-07-09 | casc_eq_train: equilibrium-mode cascade-EP trainer (GS-reverse solver K sweep... | Yuren Hao |
| 2026-07-09 | cascade: retire zil per user directive (EP route = equilibrium two-phase); pr... | Yuren Hao |
| 2026-07-09 | cascade: zil interleaved scheme = exact gradients at all depths (cos 1.0000 L... | Yuren Hao |
| 2026-07-09 | cascade ablation program: plan doc (5 claims, tiers 0-4), probe v2 (jacobi/gs... | Yuren Hao |
| 2026-07-09 | cascade_probe: layered-energy EP on L distinct standard blocks — free equil... | Yuren Hao |
| 2026-07-09 | lt_ep_train: --gov_keepjr now keeps the jr floor through reg_delay too (fastf... | Yuren Hao |
| 2026-07-09 | backfill_wandb: replay historical run logs into W&B (ept-c512 / ept-tol / ept... | Yuren Hao |
| 2026-07-09 | lt_ep_train: --rr_floor (resreg floor during free/delay phase) for govfloor arms | Yuren Hao |
| 2026-07-09 | lt_ep_train: optional W&B mirroring (--wandb/--wandb_run), failure-proof logg... | Yuren Hao |
| 2026-07-07 | SPECTRAL GOVERNOR (--governor): the from-scratch recipe candidate, dip-map-ca... | Yuren Hao |
| 2026-07-06 | dp_ep smoke PASSED: 2-GPU lockstep clean, ~107% linear scaling (allreduce neg... | Yuren Hao |
| 2026-07-06 | Tier-3 gates: Anderson math-yes/impl-no (parked for v2); bf16polish UNSAFE ne... | Yuren Hao |
| 2026-07-06 | aggregate speed bench: speed tier hf+sd 1.47x; accuracy tier hf+sd+t80+avg (0... | Yuren Hao |
| 2026-07-06 | --holoavg shipped (gate 0.913->0.936 @t2sel160, batch2 +0.051) + farm wave-2 ... | Yuren Hao |
| 2026-07-06 | bp_lm twin control + env forensics probes + holo_a_track_avg (semi-convergenc... | Yuren Hao |
| 2026-07-05 | estimator forensics closed: cos ceiling = adjoint TRUNCATION + SEMI-CONVERGEN... | Yuren Hao |
| 2026-07-05 | r-sweep: cos(EP,BPTT) FLAT across r=0.02..0.4 on both operators — nudge-SNR... | Yuren Hao |
| 2026-07-05 | --sdpa: fused flash attention in the no_grad relax loop — 1.45x free phase,... | Yuren Hao |
| 2026-07-05 | holofast: exact halved-jvp AEP track — 1.55x nudged phase, gradient-gate pa... | Yuren Hao |
| 2026-07-04 | queue v3: 5th arm abl_delay_hr2 (redx-faithful hr0.2 front) + every-100 saves... | Yuren Hao |
| 2026-07-04 | magic-s2000 study: reg_delay/noadaptc flags, 4-arm queue v2, redx trajectory ... | Yuren Hao |
| 2026-07-03 | floss-ept: graded finite-horizon LE penalty (--floss) + three-arm from-scratc... | Yuren Hao |
| 2026-07-03 | omega/norm-family refuted as stability signal; fingerprint story retracted; e... | Yuren Hao |
| 2026-07-03 | Initial commit: ept — backprop-free equilibrium transformer (EP) | Yuren Hao |
