dimitarpg13/semsimula-fock-parflm-anisogaussian-vtheta-owt-d384-verlet-instability
Text Generation • Updated • 57
Semantic Simulation — Gen 2: CfC/BAOAB, detached sources
Gen 2. CfC/BAOAB integrator; Vφ/ξ sources detached. SUPERSEDED by Gen 3 (live gradients); kept as its paired control.
Note NOT A LADDER ARM. PROLOGUE, Verlet-era. The trained stiffness pushed omega*dt past Verlet's stability bound of 2; the E/P spikes here are why the programme moved to CfC+BAOAB, whose A-substep is the exact harmonic flow and has no such bound. Every model below uses that integrator.
dimitarpg13/semsimula-ladder-owt-d384-l2-gpt2-matched
Text Generation • Updated • 801Note ARM 'gpt2-matched' — matched GPT-2 baseline. The reference. 49.81 settled. Matched tokenizer, corpus, batches, width and token budget; depth is NOT matched (L=8 against the ladder's L=2).
dimitarpg13/semsimula-ladder-owt-d384-l2-attention
Text Generation • Updated • 696Note ARM 'attention' — Fock-PARFLM, exchange field on. 63.51 settled, 1.275x the matched GPT-2. Removes the transformer; keeps the xi-routed non-conservative exchange field and the Fock register mechanism. Its gap to the no-exchange arm (+5.2%) was measured with the V_phi and xi sources detached in both arms, and clips on 29.2% of steps against 0.0%, so it is not a price of the exchange field. Gen 2: superseded by its live-gradient retrain.
dimitarpg13/semsimula-ladder-owt-d384-l2-attention-potential
Text Generation • Updated • 699Note ARM 'attention_potential' — exchange field as a potential. 80.90 settled. The conservative twin of the attention arm, parameter-matched exactly, differing only in whether the exchange field is a force or a potential gradient. Its +27.4% gap is WITHDRAWN as a price of conservativity: its sources were detached, so it sent no gradient into any hidden state. With the identical forward force and live gradients it settled at 61.11, below the attention arm's 63.51. Gen 2: superseded.
dimitarpg13/semsimula-ladder-owt-d384-l2-none
Text Generation • Updated • 687Note ARM 'no-exchange' (named 'none' in the repo and notebook) — Fock-PARFLM, no exchange field. 66.98 settled, 1.345x the matched GPT-2. Fock-PARFLM with the register mechanism on. Its layer step is NOT a geodesic: R(geo) = 1.09, the reverse channel ~90% of the deflection. Trained with the V_phi and xi sources detached. Gen 2: superseded by its live-gradient retrain.
dimitarpg13/semsimula-ladder-owt-d384-l2-none-norc
Text Generation • Updated • 698Note ARM 'conservative-only' (repo: none-norc) — the Fock mechanism off. 87.93 settled. Removes the Fock mechanism, so every force is the gradient of a scalar potential; its layer step is the damped V_theta geodesic followed by LayerNorm (R = 0.0003). Its +31.3% gap to the no-exchange arm is WITHDRAWN as the price of the Fock mechanism: its only inter-token channels, V_phi and xi, had their sources detached. Its live-gradient retrain is far better. Gen 2: superseded.
dimitarpg13/semsimula-ladder-owt-d384-l2-attention-potential-rglive
Text Generation • Updated • 453Note NOT A LADDER ARM. The gradient-path probe that found the starvation. The same forward force as the attention_potential arm, bit-identical, with only the exchange field's loss gradient made live (V_phi and xi still detached, so Gen 2). 80.90 -> 61.11 settled, below the attention arm's 63.51, with 0.8% clip-hit against 29.2%; SCAF CLEAN at all seven audits. Its Gen 3 counterpart is the everything-live run.