Declaration_Study.ipynb

1{ 2 "cells": [ 3 { 4 "cell_type": "markdown", 5 "id": "9a59273e", 6 "metadata": {}, 7 "source": [ 8 "# Declaration and persistent state change — completed secondary analysis\n", 9 "\n", 10 "**Version 1.0 | 17 September 2026 | Prepared for Joseph Anady**\n", 11 "\n", 12 "This notebook reproduces numerical findings from verified **real-data derived arrays**, selected actual-checkpoint metrics, and complete sensitivity outcomes. Original tensor processing occurred in a source-connected session. The full raw tensors are not bundled here and are **not** loaded by this notebook. Use `code/checkpoint_analysis.py --download` for the full pinned-source computation.\n", 13 "\n", 14 "Scope: **one training run**, 167 saved adapter checkpoints, 792 gradient observations, and 79 evaluation losses. No new model inference, training, declaration measurement, or consciousness classification is performed." 15 ] 16 }, 17 { 18 "cell_type": "code", 19 "execution_count": null, 20 "id": "abffee15", 21 "metadata": {}, 22 "outputs": [], 23 "source": [ 24 "from pathlib import Path\n", 25 "import sys, json, subprocess, runpy\n", 26 "import numpy as np\n", 27 "import pandas as pd\n", 28 "ROOT = Path.cwd()\n", 29 "assert (ROOT / \"provenance/source_verification.json\").is_file(), \"Run from the package root.\"\n", 30 "sys.path.insert(0, str(ROOT / \"code\"))\n", 31 "print(\"Python\", sys.version.split()[0])\n", 32 "print(\"NumPy\", np.__version__)\n" 33 ] 34 }, 35 { 36 "cell_type": "markdown", 37 "id": "ae423ee9", 38 "metadata": {}, 39 "source": [ 40 "## 1. Verify transferred artifacts against source-session receipts" 41 ] 42 }, 43 { 44 "cell_type": "code", 45 "execution_count": null, 46 "id": "7d73f0f1", 47 "metadata": {}, 48 "outputs": [], 49 "source": [ 50 "result = subprocess.run([sys.executable, \"code/verify_package.py\"], cwd=ROOT,\n", 51 " text=True, capture_output=True, check=True)\n", 52 "print(result.stdout)\n", 53 "provenance = json.loads((ROOT / \"provenance/source_verification.json\").read_text())\n", 54 "print(json.dumps(provenance, indent=2))" 55 ] 56 }, 57 { 58 "cell_type": "markdown", 59 "id": "cb2277db", 60 "metadata": {}, 61 "source": [ 62 "## 2. Recompute full-log summaries\n", 63 "\n", 64 "The two-line descriptive split deliberately retains the original 300-step pilot domain. No inference treats sequential observations as independent trials." 65 ] 66 }, 67 { 68 "cell_type": "code", 69 "execution_count": null, 70 "id": "a2c215cf", 71 "metadata": {}, 72 "outputs": [], 73 "source": [ 74 "from analyze_tables import main\n", 75 "summary = main()\n", 76 "print(json.dumps(summary, indent=2))" 77 ] 78 }, 79 { 80 "cell_type": "markdown", 81 "id": "dfd8b477", 82 "metadata": {}, 83 "source": [ 84 "## 3. Inspect actual-checkpoint summaries and all sensitivity conditions\n", 85 "\n", 86 "The selected table contains 16 observed checkpoints. It must not be mistaken for all 167 snapshots. All 18 lag/eligibility settings are retained below. The source-style amplitude filter is factor-based and is not invariant under arbitrary reciprocal factor rescaling, even though the effective-update corner is invariant." 87 ] 88 }, 89 { 90 "cell_type": "code", 91 "execution_count": null, 92 "id": "ce8484ee", 93 "metadata": {}, 94 "outputs": [], 95 "source": [ 96 "selected = pd.read_csv(ROOT / \"data/selected_checkpoint_metrics.csv\", float_precision=\"round_trip\")\n", 97 "sensitivity = pd.read_csv(ROOT / \"data/corner_sensitivity.csv\", float_precision=\"round_trip\")\n", 98 "print(selected.to_string(index=False))\n", 99 "print(\"\\nComplete sensitivity table:\")\n", 100 "print(sensitivity.to_string(index=False))" 101 ] 102 }, 103 { 104 "cell_type": "code", 105 "execution_count": null, 106 "id": "be5729ba", 107 "metadata": {}, 108 "outputs": [], 109 "source": [ 110 "comparisons = json.loads((ROOT / \"data/checkpoint_comparisons.json\").read_text())\n", 111 "validation = json.loads((ROOT / \"data/tensor_validation.json\").read_text())\n", 112 "print(json.dumps(comparisons, indent=2))\n", 113 "print(json.dumps(validation, indent=2))" 114 ] 115 }, 116 { 117 "cell_type": "markdown", 118 "id": "350f3f1c", 119 "metadata": {}, 120 "source": [ 121 "## 4. Computational tests\n", 122 "\n", 123 "These check data identity, numerical properties, and mathematical implementations. Synthetic matrix tests validate formulas; they are not simulated consciousness experiments." 124 ] 125 }, 126 { 127 "cell_type": "code", 128 "execution_count": null, 129 "id": "c95bb845", 130 "metadata": {}, 131 "outputs": [], 132 "source": [ 133 "test = subprocess.run([sys.executable, \"-m\", \"unittest\", \"discover\", \"-s\", \"code\", \"-p\", \"test_*.py\", \"-v\"],\n", 134 " cwd=ROOT, text=True, capture_output=True)\n", 135 "print(test.stdout + test.stderr)\n", 136 "assert test.returncode == 0\n", 137 "(ROOT / \"results/tests.txt\").write_text(test.stdout + test.stderr)" 138 ] 139 }, 140 { 141 "cell_type": "markdown", 142 "id": "b0fc184f", 143 "metadata": {}, 144 "source": [ 145 "## 5. Figures\n", 146 "\n", 147 "Each plot is separate. Full-series versus selected-checkpoint scope is identified in titles and manuscript captions. Reference lines identify calculations, not declarations." 148 ] 149 }, 150 { 151 "cell_type": "code", 152 "execution_count": null, 153 "id": "ee3ab451", 154 "metadata": {}, 155 "outputs": [], 156 "source": [ 157 "from make_figures import figures\n", 158 "figures(show=True)" 159 ] 160 }, 161 { 162 "cell_type": "markdown", 163 "id": "1646f33f", 164 "metadata": {}, 165 "source": [ 166 "## Interpretation boundary\n", 167 "\n", 168 "The first nonzero update (step 2), gradient maximum (169), descriptive prefix split (after 181), and filtered local-bend maximum (190) are **different quantities**. Without the filter, the shorter local-bend windows select 725. This makes measurement specification central to any proposed onset claim.\n", 169 "\n", 170 "The acquired artifacts contain no declaration variable. This notebook therefore cannot estimate declaration-before-change ordering, within-chat persistence, an onset distribution across independent runs, or subjective consciousness. The manuscript treats those as distinct hypotheses requiring additional observations.\n", 171 "\n", 172 "## Full source-connected reproduction\n", 173 "\n", 174 "```bash\n", 175 "python code/checkpoint_analysis.py --download --raw-dir raw_sources --output-dir recomputed\n", 176 "python code/verify_recomputed.py recomputed raw_sources\n", 177 "```\n", 178 "\n", 179 "Those commands fetch public pinned JSON and safetensors, compute the full geometry, and verify outputs. They do not load the full base model, train a model, or make paid inference requests." 180 ] 181 } 182 ], 183 "metadata": { 184 "kernelspec": { 185 "display_name": "Python 3", 186 "language": "python", 187 "name": "python3" 188 }, 189 "language_info": { 190 "name": "python", 191 "version": "3.13.5" 192 } 193 }, 194 "nbformat": 4, 195 "nbformat_minor": 5 196} 197