{
 "measurement": "Whether a skill's cost depends on how many skills the session already has",
 "claude_code_version": "2.1.231",
 "why": "The 2.1.231 sweep measured 40 project skills at +1,570 tokens under --setting-sources project, against a figure of +80 published by this site from 2.1.223. Before that could be reported as a version change, one confound had to be excluded. The same sweep found the marginal cost of a skill collapsing from 39.0 to 3.9 tokens somewhere between 200 and 400 skills. This machine loads 29 user-scope skills and 36 plugin skills, none of which load under the isolation flag. (An earlier version of this file said 242, from a raw find that counted marketplace clones and cache copies. The corrected count is 65, and it does not support the hypothesis, which is recorded here rather than quietly dropped.) The original measurement's ~58,300-token baseline says it ran without that flag. So the published figure may not be a different VERSION at all, but the same behaviour measured from a starting point already past the collapse.",
 "verdict": "CONFIRMED. Same version, same fixture, same round: 40 skills cost 1,558 tokens on a clean project and 152 on this machine with its own configuration loaded, a 10.2x difference with nothing changed but whether that configuration was present. The published figure was not measuring a different version. It was measuring a configured machine, which sits in the cheap regime. Why a configured machine sits there is NOT established: this one loads only 65 skills, below the count at which the clean sweep sees the collapse.",
 "method": {
  "shape": "N skill directories under .claude/skills/<name>/SKILL.md, prose descriptions cut to 90 characters, 200-byte bodies. Prompt: 'Reply with only: OK'.",
  "estimator": "MEDIAN of the rounds. Declared once and used for every figure here.",
  "paired": "Each arm gets its own floor in its own round. The two arms have completely different floors, so no delta is ever taken against the other arm's floor.",
  "arms": {
   "isolated": "--setting-sources project. No user-scope or plugin skills load.",
   "unisolated": "No isolation flag. Whatever this machine has installed loads too."
  },
  "controls": {
   "flag": "The isolation flag is asserted PRESENT on the isolated arm and asserted ABSENT on the unisolated arm, so the treatment is proved on both sides rather than only one.",
   "arrival": "Every reply had to be exactly OK.",
   "fixture": "Skills counted back off disk per run.",
   "same_round": "Both arms run inside one round before the next begins."
  }
 },
 "results": {
  "isolated": {
   "cells": {
    "40": {
     "delta": 1558.0,
     "per_skill": 38.95,
     "rounds": 4,
     "min": 1360,
     "max": 1558,
     "spread": 198
    },
    "150": {
     "delta": 5848.0,
     "per_skill": 38.99,
     "rounds": 4,
     "min": 5848,
     "max": 5848,
     "spread": 0
    },
    "1000": {
     "delta": 11699.0,
     "per_skill": 11.7,
     "rounds": 4,
     "min": 11699,
     "max": 11699,
     "spread": 0
    }
   },
   "floors": [
    22528,
    22528,
    22528,
    22528
   ],
   "floor_drift": 0
  },
  "unisolated": {
   "cells": {
    "40": {
     "delta": 152.0,
     "per_skill": 3.8,
     "rounds": 4,
     "min": 152,
     "max": 302,
     "spread": 150
    },
    "150": {
     "delta": 579.0,
     "per_skill": 3.86,
     "rounds": 4,
     "min": 579,
     "max": 729,
     "spread": 150
    },
    "1000": {
     "delta": 3875.0,
     "per_skill": 3.88,
     "rounds": 4,
     "min": 3875,
     "max": 3875,
     "spread": 0
    }
   },
   "floors": [
    60041,
    60191,
    60191,
    60191
   ],
   "floor_drift": 150
  }
 },
 "what_this_means_for_the_published_figure": "A skill's cost is not a property of the skill. It is a property of how many skills the session already has. The site's published 1.5 tokens per skill was measured on a configured machine, where it is right. The 39 tokens measured here is right on a clean project. Both are correct measurements of different situations, which is why one did not reproduce the other.",
 "known_limits": {
  "the_unisolated_arm_changes_more_than_skill_count": "Dropping the isolation flag also loads a user-level CLAUDE.md, plugins, MCP servers, commands, agents and hooks, not only skills. So this arm shows what a configured machine actually costs; it does not identify which part of that configuration is responsible. The clean sweep establishes only that on a clean project the marginal cost falls with the size of the listing when project skills are the sole variable. It cannot establish the cause here, because it never varies pre-existing user or plugin skills, and this machine loads too few of them to be past the collapse.",
  "not_a_mechanism": "No claim is made about what changes between 200 and 400 skills, or about what the listing does internally. Only the cost is measured.",
  "an_intermittent_150_token_component": "The unisolated arm carries a 150-token component that fires on some calls and not others. In round 3 it was absent from the floor and from the 1000-skill cell but present in the 40 and 150 cells, shifting those two deltas up by exactly 150. Absolute context for every unisolated treatment cell is identical across all four rounds, so the drift is in which calls carry that component, not in the treatment. The median across four rounds is unaffected, which is why the estimator is a median.",
  "the_isolation_flag_does_not_explain_the_last_factor_of_two": "The published figure is 80 tokens for 40 skills and 1,461 for 1,000. This arm measures 152 and 3,875 for the same counts in the same regime. The flag accounts for the order of magnitude; a residual factor of roughly 1.9 to 2.6 is not accounted for and could be version, fixture detail, a different plugin load on the day, or the earlier measurement reading a different usage field. No claim is made about which.",
  "one_machine_one_model": "Opus on Claude Code 2.1.231. The unisolated arm is specific to this machine's installed configuration.",
  "nothing_from_the_machine_config_is_recorded": "The unisolated arm runs against a real user configuration. Only token counts and the reply string are recorded from it.",
  "the_cause_of_the_configured_arm_being_cheap_is_NOT_established": "The configured arm measures 3.8 tokens per skill, which is the collapsed rate. The obvious explanation, that the machine already has enough skills to be past the collapse, is REFUTED by counting them: only 65 load, and the clean sweep puts the marginal cost at 39 tokens per skill at that count. So dropping the isolation flag moves the price by 10.2x for a reason this corpus does not identify. Dropping it also loads a user CLAUDE.md, plugin commands, plugin agents, MCP servers and hooks, any of which could matter. What is measured is that a configured machine is in the cheap regime; what is not measured is which part of the configuration puts it there.",
  "the_two_routes_do_not_agree_at_this_machine_s_own_count": "The configured marginal cost, 3.8 to 3.9 tokens per skill, matches the collapsed regime the clean sweep reaches past 400 skills. It does NOT match the clean sweep's rate at this machine's actual loaded count of 65, which is 39. The two routes agree with each other only at the far end of the curve, which is why no count-to-price mapping is claimed here."
 },
 "runs": [
  {
   "arm": "floor_isolated",
   "isolated": true,
   "skills": 0,
   "round": 0,
   "status": "ok",
   "context": 22528,
   "delta": 0
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 40,
   "round": 0,
   "status": "ok",
   "skills_on_disk": 40,
   "context": 24086,
   "delta": 1558,
   "per_skill": 38.95
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 150,
   "round": 0,
   "status": "ok",
   "skills_on_disk": 150,
   "context": 28376,
   "delta": 5848,
   "per_skill": 38.99
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 1000,
   "round": 0,
   "status": "ok",
   "skills_on_disk": 1000,
   "context": 34227,
   "delta": 11699,
   "per_skill": 11.7
  },
  {
   "arm": "floor_unisolated",
   "isolated": false,
   "skills": 0,
   "round": 0,
   "status": "ok",
   "context": 60191,
   "delta": 0
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 40,
   "round": 0,
   "status": "ok",
   "skills_on_disk": 40,
   "context": 60343,
   "delta": 152,
   "per_skill": 3.8
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 150,
   "round": 0,
   "status": "ok",
   "skills_on_disk": 150,
   "context": 60770,
   "delta": 579,
   "per_skill": 3.86
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 1000,
   "round": 0,
   "status": "ok",
   "skills_on_disk": 1000,
   "context": 64066,
   "delta": 3875,
   "per_skill": 3.88
  },
  {
   "arm": "floor_isolated",
   "isolated": true,
   "skills": 0,
   "round": 1,
   "status": "ok",
   "context": 22528,
   "delta": 0
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 40,
   "round": 1,
   "status": "ok",
   "skills_on_disk": 40,
   "context": 24086,
   "delta": 1558,
   "per_skill": 38.95
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 150,
   "round": 1,
   "status": "ok",
   "skills_on_disk": 150,
   "context": 28376,
   "delta": 5848,
   "per_skill": 38.99
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 1000,
   "round": 1,
   "status": "ok",
   "skills_on_disk": 1000,
   "context": 34227,
   "delta": 11699,
   "per_skill": 11.7
  },
  {
   "arm": "floor_unisolated",
   "isolated": false,
   "skills": 0,
   "round": 1,
   "status": "ok",
   "context": 60191,
   "delta": 0
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 40,
   "round": 1,
   "status": "ok",
   "skills_on_disk": 40,
   "context": 60343,
   "delta": 152,
   "per_skill": 3.8
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 150,
   "round": 1,
   "status": "ok",
   "skills_on_disk": 150,
   "context": 60770,
   "delta": 579,
   "per_skill": 3.86
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 1000,
   "round": 1,
   "status": "ok",
   "skills_on_disk": 1000,
   "context": 64066,
   "delta": 3875,
   "per_skill": 3.88
  },
  {
   "arm": "floor_isolated",
   "isolated": true,
   "skills": 0,
   "round": 2,
   "status": "ok",
   "context": 22528,
   "delta": 0
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 40,
   "round": 2,
   "status": "ok",
   "skills_on_disk": 40,
   "context": 23888,
   "delta": 1360,
   "per_skill": 34.0
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 150,
   "round": 2,
   "status": "ok",
   "skills_on_disk": 150,
   "context": 28376,
   "delta": 5848,
   "per_skill": 38.99
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 1000,
   "round": 2,
   "status": "ok",
   "skills_on_disk": 1000,
   "context": 34227,
   "delta": 11699,
   "per_skill": 11.7
  },
  {
   "arm": "floor_unisolated",
   "isolated": false,
   "skills": 0,
   "round": 2,
   "status": "ok",
   "context": 60191,
   "delta": 0
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 40,
   "round": 2,
   "status": "ok",
   "skills_on_disk": 40,
   "context": 60343,
   "delta": 152,
   "per_skill": 3.8
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 150,
   "round": 2,
   "status": "ok",
   "skills_on_disk": 150,
   "context": 60770,
   "delta": 579,
   "per_skill": 3.86
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 1000,
   "round": 2,
   "status": "ok",
   "skills_on_disk": 1000,
   "context": 64066,
   "delta": 3875,
   "per_skill": 3.88
  },
  {
   "arm": "floor_isolated",
   "isolated": true,
   "skills": 0,
   "round": 3,
   "status": "ok",
   "context": 22528,
   "delta": 0
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 40,
   "round": 3,
   "status": "ok",
   "skills_on_disk": 40,
   "context": 24086,
   "delta": 1558,
   "per_skill": 38.95
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 150,
   "round": 3,
   "status": "ok",
   "skills_on_disk": 150,
   "context": 28376,
   "delta": 5848,
   "per_skill": 38.99
  },
  {
   "arm": "isolated",
   "isolated": true,
   "skills": 1000,
   "round": 3,
   "status": "ok",
   "skills_on_disk": 1000,
   "context": 34227,
   "delta": 11699,
   "per_skill": 11.7
  },
  {
   "arm": "floor_unisolated",
   "isolated": false,
   "skills": 0,
   "round": 3,
   "status": "ok",
   "context": 60041,
   "delta": 0
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 40,
   "round": 3,
   "status": "ok",
   "skills_on_disk": 40,
   "context": 60343,
   "delta": 302,
   "per_skill": 7.55
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 150,
   "round": 3,
   "status": "ok",
   "skills_on_disk": 150,
   "context": 60770,
   "delta": 729,
   "per_skill": 4.86
  },
  {
   "arm": "unisolated",
   "isolated": false,
   "skills": 1000,
   "round": 3,
   "status": "ok",
   "skills_on_disk": 1000,
   "context": 63916,
   "delta": 3875,
   "per_skill": 3.88
  }
 ],
 "machine_skill_count": {
  "skills_that_actually_load": 65,
  "user_scope": 29,
  "installed_plugins": 36,
  "skill_md_files_on_disk_anywhere": 213,
  "how_counted": "Plugin skills counted as skills/*/SKILL.md and skills/*/*/SKILL.md under each installPath in ~/.claude/plugins/installed_plugins.json, plus ~/.claude/skills/*/SKILL.md. A raw `find ~/.claude/plugins -name SKILL.md` returns 213, but that counts marketplace clones (22 and 6), a temp git checkout (14) and cache copies of versions that are not the installed one. An earlier version of this corpus published 242 from the raw count and was wrong.",
  "why_it_matters": "65 is BELOW the 200-to-400 bracket where the clean sweep sees the marginal cost collapse. So the pre-existing SKILL COUNT does not explain why the configured arm is already in the cheap regime. Something else about a loaded configuration does, and this corpus does not identify what."
 }
}