dimitarpg13/semsimula-ladder-live-owt-d384-l4-none

Text Generation • Updated 1 day ago • 657

Note ARM 'no-exchange', Fock mechanism on, L=4, live gradients. 50.10 settled, 1.006x the matched GPT-2: parity at equal tokens, within eval noise, with 2.3x its parameters (76.8M vs 33.7M) and half its depth; one seed, untuned baseline. −30.2% against the same model with detached sources (71.75). First clean forecastability test (CG3): passes. SCAF CLEAN at seven audits, Tiers A and B zero.

dimitarpg13/semsimula-ladder-live-owt-d384-l2-attention-potential-rfqk

Text Generation • Updated 1 day ago • 755

Note ARM 'attention_potential' with hardened routing (QK-normalised, clamped per-head logit scale, own 0.3 clip), Fock on, L=2, everything live. 52.90 settled, 1.062x the matched GPT-2. 0.4% better than the same model without the field (53.12): within eval noise, so the hardened field adds nothing measurable. 2.4% better than the unhardened version (54.21). The most refinement-ready Fock model: +115% at 1.5x the steps, against +1,274-1,342% for its L=2 siblings (the conservative SR2: +50%). One seed.

dimitarpg13/semsimula-ladder-live-owt-d384-l2-none

Text Generation • Updated 1 day ago • 615

Note ARM 'no-exchange', Fock mechanism on, L=2, live gradients. 53.12 settled, 1.066x the matched GPT-2. −8.0% against the L=2 conservative-only live model (57.76): what the register mechanism adds when trained. The L=4 live model is 5.7% better: depth helps at matched mechanism and convention, at 1.9x the step cost. −20.7% against the same model with detached sources (66.98). Vφ inert; refinement fails 9x worse than conservative-only. One seed.

dimitarpg13/semsimula-ladder-live-owt-d384-l2-attention-potential

Text Generation • Updated 1 day ago • 749

Note ARM 'attention_potential' (conservative exchange field), Fock on, L=2, everything live. 54.21 settled, 1.088x the matched GPT-2. 2.1% WORSE than the same model without the field (53.12): as trained, the field adds nothing. It is load-bearing (switching it off costs +56%) but substitutes for other work. Its routing weights grew 5-6x unchecked (no QK-norm); the hardened version (QK-norm, next item) settles at 52.90. −11.3% against its live-field-only probe (61.11). One seed.

dimitarpg13/semsimula-ladder-live-owt-d384-l2-attention-potential-rfqk-norc

Text Generation • Updated 1 day ago • 348

Note XRB-F: the conservative-only model (F3.1) plus G3′'s hardened exchange field, with no register read. 54.27 settled, 1.090x the matched GPT-2. The field alone lowers F3.1's perplexity by 6.0%, three quarters of what the register path gives (8.0%); beside the registers it adds 0.4%. Alone it carries loss everywhere, as the register path does alone. Refines like F3.1 (+134% at 1.5x the steps). An exploratory instrument; one seed.

dimitarpg13/semsimula-ladder-live-owt-d384-l2-none-norc-pm64-pmx16-sr2

Text Generation • Updated 1 day ago • 655

Note PMX: the first model with Poisson-mode registers, the wells integrated exactly on the SR2 base. 54.63 settled: the best conservative model on the ladder, 6.3% better than SR2 and 2.8% behind the Fock model. Every force a gradient; momentum lock gone (Gate 1 +91%). Not refinement-ready under its pre-registered criterion: the per-layer flow reaches a limit 0.39 nats from the trained step. Single seed, 0.5B tokens.

dimitarpg13/semsimula-ladder-live-owt-d384-l2-none-norc

Text Generation • Updated 1 day ago • 609

Note ARM 'conservative-only', live gradients. Vθ + Vφ + ξ with no Fock mechanism and no exchange field: every force is a gradient. 57.76 settled, 1.160x the matched GPT-2, the best L=2 model without the register path. Same forward pass as its Gen 2 twin (87.93); only the gradient convention differs. Causality: SCAF CLEAN at seven audits and an independent future-perturbation check exactly 0. Single seed.

dimitarpg13/semsimula-ladder-live-owt-d384-l2-none-norc-sr2

Text Generation • Updated 1 day ago • 595

Note SR2: the conservative-only model with its stiff modes integrated as exact damped oscillators. 58.32 settled, level with F3.1 (+1.0%). Its layer step is a coarse sample of a convergent per-layer flow (pre-registered, replicated on fresh data): run each layer in k substeps with substeps_per_layer = k and perplexity converges, within 0.10 to 0.13 nats of k = 1. Depth becomes an inference-time knob. Single seed.

dimitarpg13/semsimula-ladder-owt-d384-l2-gpt2-matched

Text Generation • Updated 1 day ago • 803

Note SHARED REFERENCE for Gen 2 and Gen 3. Matched GPT-2 baseline, 49.81 settled: matched tokenizer, corpus, batches, width and token budget; depth NOT matched (L=8 against the ladder's L=2). Retrained on the Fock ladder's exact schedule (W2, peak 1.2e-3): 48.37; on that schedule it trails the Fock models until the learning-rate decay.