{
 "figure": "Figure 1",
 "caption": "Figure 1. The calibrated concept profile in six steps, left to right. (1) A generator with a declared distribution draws the batteries: set A (the construct in its role), set B (its token out of role) and set C (ordinary code). (2) The model at one checkpoint is run on all three, giving raw activations at each prompt's read site. (3) Per neuron, A minus B in units of that neuron's own spread on C at the same kind of position -- the yardstick -- gives the calibrated profile. (4) Steps 1 to 3 are repeated on a second battery drawn from the same distribution, and the spread between the two profiles is the noise unit every later figure is read in. (5) Nine checks on the profiles give the licence: two must pass, seven are reported. (6) The same battery and the same process at every checkpoint make readings comparable across training, with a stated resolution.",
 "files": [
  "<data>/validation/paper1/FIG1_method.png",
  "<data>/validation/paper1/FIG1_method.svg"
 ],
 "what": "a schematic of the method; nothing in it is measured",
 "steps": [
  "prompt generator",
  "raw readings",
  "the profile",
  "the noise unit",
  "the licence",
  "readings across training"
 ],
 "layout": "one row of six, arrows consecutive 1->2 ... 5->6, and a dashed return 6 -> 2 labelled 'at each checkpoint'",
 "running_example": "for",
 "figsize_inches": [
  7.4,
  1.71
 ],
 "axes_units": [
  100,
  23.1
 ],
 "isotropic": true,
 "cropped": "bbox_inches=tight: no whitespace above or below",
 "provenance": {
  "inputs": [
   "none: a diagram, not a reading"
  ],
  "reference": null,
  "battery_prompt_hash": "570b42db13f8fe5772b8d1a9ad7d55d6fa1279dae5948ad94e8363789bcbc1ad",
  "code_commit": "78d9e01",
  "design": "9-Paper-1/2610-18Oct_PAPER-1-Design.md sha256 47f6a664\u2026"
 }
}