{
  "model_family": "semsimula-mechanism-ladder",
  "generation": 3,
  "collection": "depth on code",
  "ladder_arm": "none-norc-sr2-l8-code-live",
  "role": "depth: SO-CODE split into eight layers at the same integration time, the matched GPT-2's depth",
  "variant_of": "semsimula-ladder-live-code-d384-l2-none-norc-sr2",
  "integrator": {
    "name": "baoab_cfc_lowrank",
    "lowrank_damped_flow": true,
    "fixed_gamma": 0.1,
    "first_order_overdamped": false,
    "poisson_modes": 0
  },
  "architecture": "as the arm no-exchange but REVERSE_CHANNEL=False: registers are created and destroyed but never reach the tokens, so every force is the gradient of a scalar potential",
  "params_total": 76731865,
  "context_channels": 5,
  "d": 384,
  "n_layers": 8,
  "dt": 1.0,
  "integration_time_T": 8.0,
  "vocab_size": 50257,
  "block_size": 512,
  "corpus": "codeparrot/codeparrot-clean (Python; the ladder's cache: 450M training tokens, first 2M validation tokens; one tab per four-space indentation level)",
  "total_steps": 19250,
  "schedule_steps": 16250,
  "effective_batch": 32,
  "training_tokens": 315392000,
  "learning_rate": 0.0012,
  "settled_ppl_512": 5.3545,
  "best_ppl_512": 5.26,
  "final_ppl_512": 5.45,
  "best_step": 18000,
  "settled_convention": "exp of the mean log perplexity of the evals at steps 18,000, 18,500 and 19,000",
  "so_code_settled_ppl_512": 5.9951,
  "l4_code_settled_ppl_512": 5.4847,
  "fock_code_settled_ppl_512": 5.6567,
  "fo_b_code_settled_ppl_512": 6.5487,
  "matched_gpt2_code_settled_ppl_512": 4.96,
  "paired_nats": {
    "l8_so": {
      "early": [
        0.069,
        0.054,
        5
      ],
      "mid": [
        -0.098,
        0.006,
        11
      ],
      "late": [
        -0.112,
        0.003,
        7
      ]
    },
    "l8_l4": {
      "early": [
        0.362,
        0.073,
        5
      ],
      "mid": [
        0.001,
        0.011,
        11
      ],
      "late": [
        -0.024,
        0.002,
        7
      ]
    },
    "spans": {
      "early": "2,000–4,000",
      "mid": "8,000–13,000",
      "late": "16,000–19,000"
    }
  },
  "copy_split": {
    "closed_all": 59.8,
    "first": 59.1,
    "induction": 64.5,
    "indent": 53.2,
    "ident_near": 58.5,
    "ident_far": 59.9,
    "near": 58.6,
    "mid": 61.1,
    "far": 60.0,
    "diff": 1.4,
    "ci": [
      -3.2,
      6.3
    ],
    "p": 0.72,
    "windows_better_so": 98.2,
    "windows_better_l4": 74.2,
    "ppl": 5.765,
    "l4_ppl": 5.898
  },
  "wells_by_layer": {
    "ratio": [
      1.21,
      1.21,
      1.16,
      1.14,
      1.13,
      1.18,
      1.3,
      1.46
    ],
    "ratio_p99": [
      1.78,
      1.69,
      1.63,
      1.63,
      1.61,
      1.64,
      1.76,
      1.93
    ],
    "occ": [
      0.98,
      0.26,
      0.17,
      0.42,
      0.45,
      0.3,
      0.13,
      0.06
    ],
    "n_eff": [
      7.1,
      1.9,
      1.9,
      2.1,
      2.1,
      2.0,
      1.8,
      1.8
    ],
    "two_well_share": [
      56,
      87,
      92,
      92,
      95,
      97,
      98,
      99
    ],
    "removal": [
      2780,
      2.8,
      3.4,
      2.2,
      26,
      25,
      6.2,
      0.6
    ],
    "removal_all": 3251,
    "base_ppl": 6.74,
    "ratio_range": [
      0.81,
      1.91
    ],
    "scale": {
      "x0": 1.85,
      "xdeep": [
        22,
        28
      ],
      "ratio": 13,
      "l0_up": 2.0,
      "deep_down": [
        7.1,
        7.3
      ]
    }
  },
  "xi_reach": {
    "lift_far": [
      10.5,
      15.3,
      20.5,
      25.1,
      27.3,
      23.2,
      12.0
    ],
    "channel": 4,
    "eff_pos": [
      101,
      76,
      54,
      39,
      32,
      31,
      43
    ]
  },
  "gates_6b7": {
    "g0": 4.59,
    "g1": 172.1,
    "g3": {
      "1": 4.552,
      "2": 3.539,
      "3": 3.576,
      "4": 2.013,
      "6": 0.206,
      "8": 0.0,
      "12": 0.416,
      "16": 0.783,
      "24": 2.175,
      "32": 2.604
    },
    "g2": {
      "1": 67.35,
      "2": 59.42,
      "3": 42.51,
      "4": 26.92,
      "6": 7.5,
      "8": 4.59
    },
    "g3_ppl": {
      "1": 434.93,
      "2": 157.93,
      "3": 163.81,
      "4": 34.33,
      "6": 5.64,
      "8": 4.59,
      "12": 6.95,
      "16": 10.04,
      "24": 40.37,
      "32": 61.98
    }
  },
  "omega_dt_6b13": {
    "p50": 1.75,
    "layers": [
      2.08,
      1.91,
      1.34,
      2.26,
      2.09,
      1.7,
      1.05,
      0.62
    ],
    "over_wall": 33.8,
    "omega": 1.75
  },
  "cell7_ppl": 4.89,
  "flow_r": {
    "rows": [
      [
        "L8-CODE (L = 8, Δt = 1; trained ppl 5.18 on these batches)",
        {
          "k": [
            0.186,
            0.029,
            0.034,
            0.035,
            0.035,
            0.035
          ],
          "stack4": 2.695
        }
      ],
      [
        "L4-CODE (L = 4, Δt = 2)",
        {
          "k": [
            0.429,
            0.52,
            0.282,
            0.32,
            0.309,
            0.306
          ],
          "stack4": 2.96
        }
      ],
      [
        "SO-CODE (L = 2, Δt = 4)",
        {
          "k": [
            0.97,
            0.526,
            0.3,
            0.384,
            0.376,
            0.369
          ],
          "stack4": 2.858
        }
      ]
    ],
    "reading": "**At eight layers the per-layer flow settles 0.035 nats from the trained step**, ten times closer than at two or four layers (0.37, 0.31) and closer than any model in this programme so far (SR2 on prose: 0.10). The whole stack still fails refinement (2.70 nats at k = 4). So depth at a fixed integration time brings each trained layer onto its own continuous flow; it does not make the stack one."
  },
  "pre_registered": "none on its own result: L8-CODE ran by the rule fixed before L4-CODE trained (it runs if L4-CODE settles more than 2% below SO-CODE; L4-CODE settled 8.5% below)",
  "run_health": "19,250 steps in 29.5 h over two Colab sessions (5.5 s per step with layer checkpointing; resumed at step 15,500 with the batch sampler restored, so the evals stay paired: their step-to-step changes correlate with L4-CODE's at 0.991 after the resume and 0.994 before it); 0 watchdog triggers; largest rise between logged training losses 0.069 (step 13,050); gradient norm median 0.38 after step 2,000, maximum 1.06 (step 400, warmup); SCAF leak monitor CLEAN at all 4 audits. 34 of the 45 logged steps between 1,400 and 3,600 had a non-finite gradient norm, and the training loop skips those updates (cause not diagnosed; none after step 3,600)",
  "companion_notes": [
    "Depth_Ladder_and_Matched_Baseline_Protocol.md",
    "Dead_Wells_in_the_Joint_V_theta_Bank.md"
  ],
  "cell0": "REVERSE_CHANNEL=False, VPHI_GRAD_PATH='live', XI_GRAD_PATH='live',\nLOWRANK_DAMPED_FLOW=True,                 # SR2\nVPHI_ANALYTIC_FORCE=True                  # speed switch, same arithmetic\nCORPUS='code', LADDER_STEPS=16_250, EXTRA_FLOOR_STEPS=3_000   # Python code, then 3,000 steps at the floor\nLADDER_L=8                                # eight layers; dt = T / L = 1, T = 8 as at L = 2 and 4\nSNAPSHOT_STEPS=[1_000, 1_500, 2_000, 2_500, 3_000]   # weights-only snapshots for the well probe\n# LAYER_CHECKPOINT stays at its default True (memory, same arithmetic)",
  "gdrive_folder": "semsimula_fock_cfc_baoab_code_xi5long_topk16_dt32da16_mh4_aniso_dcvt5x8_vtjoint_cgqk_norc_vplive_xilive_sr2_st16250_L8probe_ob_untied_wsd_e5c_plgate_rep0.05_fockreg0.005_g0.1_baoab_cfc_lowrank_lr0p0012_noattn"
}